Daily Papers

Image-to-poster generation is a high-demand task requiring not only local adjustments but also high-level design understanding. Models must generate text, layout, style, and visual elements while preserving semantic fidelity and aesthetic coherence. The process spans two regimes: local editing, where ID-driven generation, rescaling, filling, and extending must preserve concrete visual entities; and global creation, where layout- and style-driven tasks rely on understanding abstract design concepts. These intertwined demands make image-to-poster a multi-dimensional process coupling entity-preserving editing with concept-driven creation under image-prompt control. To address these challenges, we propose PosterOmni, a generalized artistic poster creation framework that unlocks the potential of a base edit model for multi-task image-to-poster generation. PosterOmni integrates the two regimes, namely local editing and global creation, within a single system through an efficient data-distillation-reward pipeline: (i) constructing multi-scenario image-to-poster datasets covering six task types across entity-based and concept-based creation; (ii) distilling knowledge between local and global experts for supervised fine-tuning; and (iii) applying unified PosterOmni Reward Feedback to jointly align visual entity-preserving and aesthetic preference across all tasks. Additionally, we establish PosterOmni-Bench, a unified benchmark for evaluating both local editing and global creation. Extensive experiments show that PosterOmni significantly enhances reference adherence, global composition quality, and aesthetic harmony, outperforming all open-source baselines and even surpassing several proprietary systems.

On-policy distillation (OPD) has emerged as an effective framework for post-training language models by pairing student-generated trajectories with dense token-level supervision from a teacher. However, OPD implicitly assumes that teacher-derived rewards are an appropriate proxy for reasoning progress, and therefore treats all teacher feedback equally during policy optimization. While in practice, this assumption does not always hold. We observe that teacher-derived rewards often conflict with genuine reasoning progress, as reasoning steps with clear reasoning advancement may still receive lower distillation rewards, simply due to deviation from teacher's outputs. To address this mismatch, we propose Reasoning-Progress-Aware Reward Filtering for On-Policy Distillation (R2-OPD), which constructs two within-trajectory rankings of reasoning spans, one from teacher-derived rewards and the other from independently estimated progress reward. Distillation rewards are selectively suppressed whenever the two rankings disagree, reducing supervision that conflicts with reasoning progress while preserving effective teacher guidance. Our approach shows consistent improvement over standard OPD especially regarding reasoning performances.

On-policy distillation (OPD) has been widely studied as a post-training method in which a student model obtains token-level supervision from a stronger teacher on its own rollouts. Recent studies have improved OPD through alternative distillation reward formulations and teacher configurations, while the objective of distillation remains centered on mimicking the teacher. However, when a stronger teacher has limited distributional overlap with the student, such positive guidance can provide insufficient learning signals. In this work, we introduce Negative-Policy OPD (NP-OPD), which complements teacher supervision with rollouts from a lower-performing, lower-capability negative policy that serves as a negative reference for the student. Rather than modifying the distillation reward formulation, NP-OPD introduces the negative policy at the rollout stage, continuously supplying tokens preferred by the negative policy over the teacher so that they remain exposed to teacher supervision throughout training. This provides an explicit negative signal through negative-policy rollouts while preserving the positive teacher supervision used in OPD. Through extensive experiments, we show that NP-OPD improves OPD across model scales, generation modes, reasoning domains, and different OPD variants. Furthermore, our analyses show that NP-OPD effectively suppresses tokens preferred by the negative policy over the teacher and moves the student away from the negative policy. These results support our design of introducing negative signals through negative-policy rollouts and provide new insight into the role of the rollout policy in OPD. Code will be available at https://github.com/naver-ai/np-opd.

On-policy distillation is an alternative post-training method in reinforcement learning that alleviates the constraints imposed by reward models by providing token-level supervision from a teacher model. Although on-policy distillation has been studied and applied across various settings, its fundamental design remains underexplored. In this paper, we introduce a new distillation reward, termed the delta signal, instead of directly imitating the teacher's output distribution. The delta signal is defined as the difference between the teacher model and its base model prior to instruction tuning for reasoning capability. It therefore captures the changes induced by reasoning tuning and provides a more direct signal for transferring reasoning capabilities. Using extensive empirical evidence, we show that the delta signal substantially improves on-policy distillation and refer to the new distillation method as On-Policy Delta Distillation (OPD^2). Experiments across mathematics, science, and code-reasoning benchmarks demonstrate that OPD^2 consistently outperforms conventional on-policy distillation, enabling reasoning LLMs to achieve strong performance with only a short post-training period. Code will be available at https://github.com/naver-ai/opd2

Mixture-of-Experts models, now popular for scaling capacity at fixed inference speed, switch experts at nearly every token. Once a model outgrows available GPU memory, this churn can render optimizations like offloading and pre-fetching ineffective. We make the case that the options framework in reinforcement learning is a perfect match to tackle this problem, and argue for temporally extended mixture-of-experts layers. Building on the option-critic framework with deliberation costs, we add a controller to each layer that learns when to switch expert sets and which to load. By applying this to gpt-oss-20b with low-rank adapters and a self-distillation reward, our method reduces switch rates from over 50% to below 5% while retaining up to 90% of base-model accuracy on MATH, MMLU, and MMMLU. This shows that even existing pre-trained models can be converted to temporally extended MoEs with lightweight training, with the deliberation cost allowing model trainers to trade off switching rates against capability. We hope this opens a principled path, grounded in the options framework, for memory-efficient serving and continual learning in ever-growing MoE models.

Masked diffusion language models (MDLMs) offer the potential for parallel token generation, but most open-source MDLMs decode fewer than 5 tokens per model forward pass even with sophisticated sampling strategies. As a result, their sampling speeds are often comparable to AR + speculative decoding schemes, limiting their advantage over mainstream autoregressive approaches. Existing distillation-based accelerators (dParallel, d3LLM) finetune MDLMs on trajectories generated by a base model, which can become off-policy during finetuning and restrict performance to the quality of the base model's samples. We propose dUltra, an on-policy reinforcement learning framework based on Group Relative Policy Optimization (GRPO) that learns unmasking strategies for efficient parallel decoding. dUltra introduces an unmasking planner head that predicts per-token unmasking likelihoods under independent Bernoulli distributions. We jointly optimize the base diffusion LLM and the unmasking order planner using reward signals combining verifiable reward, distillation reward, and the number of unmasking steps. Across mathematical reasoning and code generation tasks, dUltra improves the accuracy--efficiency trade-off over state-of-the-art heuristic and distillation baselines, moving towards achieving ``diffusion supremacy'' over autoregressive models.

Large-scale vision-language models (VLMs) have recently achieved remarkable multimodal understanding, but their massive size makes them impractical for deployment on mobile or edge devices. This raises the need for compact yet capable VLMs that can efficiently learn from powerful large teachers. However, distilling knowledge from a large teacher to a small student remains challenging due to their large size gap: the student often fails to reproduce the teacher's complex, high-dimensional representations, leading to unstable learning and degraded performance. To address this, we propose Masters (Masking Teacher and Reinforcing Student), a mask-progressive reinforcement learning (RL) distillation framework. Masters first masks non-dominant weights of the teacher to reduce unnecessary complexity, then progressively restores the teacher by gradually increasing its capacity during training. This strategy allows the student to learn richer representations from the teacher in a smooth and stable manner. To further refine knowledge transfer, Masters integrates an offline RL stage with two complementary rewards: an accuracy reward that measures the correctness of the generated responses, and a distillation reward that quantifies the ease of transferring responses from teacher to student. Unlike online think-answer RL paradigms that are computationally expensive and generate lengthy responses, our offline RL leverages pre-generated responses from masked teachers. These provide rich yet efficient guidance, enabling students to achieve strong performance without requiring the think-answer process.

While safety alignment for Multimodal Large Language Models (MLLMs) has gained significant attention, current paradigms primarily target malicious intent or situational violations. We propose shifting the safety frontier toward consequence-driven safety, a paradigm essential for the robust deployment of autonomous and embodied agents. To formalize this shift, we introduce OOD-MMSafe, a benchmark comprising 455 curated query-image pairs designed to evaluate a model's ability to identify latent hazards within context-dependent causal chains. Our analysis reveals a pervasive causal blindness among frontier models, with the highest 67.5% failure rate in high-capacity closed-source models, and identifies a preference ceiling where static alignment yields format-centric failures rather than improved safety reasoning as model capacity grows. To address these bottlenecks, we develop the Consequence-Aware Safety Policy Optimization (CASPO) framework, which integrates the model's intrinsic reasoning as a dynamic reference for token-level self-distillation rewards. Experimental results demonstrate that CASPO significantly enhances consequence projection, reducing the failure ratio of risk identification to 7.3% for Qwen2.5-VL-7B and 5.7% for Qwen3-VL-4B while maintaining overall effectiveness.

Score Distillation Sampling (SDS) has emerged as an effective technique for leveraging 2D diffusion priors for tasks such as text-to-3D generation. While powerful, SDS struggles with achieving fine-grained alignment to user intent. To overcome this, we introduce RewardSDS, a novel approach that weights noise samples based on alignment scores from a reward model, producing a weighted SDS loss. This loss prioritizes gradients from noise samples that yield aligned high-reward output. Our approach is broadly applicable and can extend SDS-based methods. In particular, we demonstrate its applicability to Variational Score Distillation (VSD) by introducing RewardVSD. We evaluate RewardSDS and RewardVSD on text-to-image, 2D editing, and text-to-3D generation tasks, showing significant improvements over SDS and VSD on a diverse set of metrics measuring generation quality and alignment to desired reward models, enabling state-of-the-art performance. Project page is available at https://itaychachy. github.io/reward-sds/.

Diffusion and flow models achieve State-Of-The-Art (SOTA) generative performance, yet many practically important behaviors such as fine-grained prompt fidelity, compositional correctness, and text rendering are weakly specified by score or flow matching pretraining objectives. Reinforcement Learning (RL) fine-tuning with external, black-box rewards is a natural remedy, but diffusion RL is often brittle. Trajectory-based methods incur high memory cost and high-variance gradient estimates; forward-process approaches converge faster but can suffer from distribution drift, and hence reward hacking. In this work, we present Centered Reward Distillation (CRD), a diffusion RL framework derived from KL-regularized reward maximization built on forward-process-based fine-tuning. The key insight is that the intractable normalizing constant cancels under within-prompt centering, yielding a well-posed reward-matching objective. To enable reliable text-to-image fine-tuning, we introduce techniques that explicitly control distribution drift: (i) decoupling the sampler from the moving reference to prevent ratio-signal collapse, (ii) KL anchoring to a CFG-guided pretrained model to control long-run drift and align with the inference-time semantics of the pre-trained model, and (iii) reward-adaptive KL strength to accelerate early learning under large KL regularization while reducing late-stage exploitation of reward-model loopholes. Experiments on text-to-image post-training with GenEval and OCR rewards show that CRD achieves competitive SOTA reward optimization results with fast convergence and reduced reward hacking, as validated on unseen preference metrics.

Distillation-based acceleration has become foundational for making autoregressive streaming video diffusion models practical, with distribution matching distillation (DMD) as the de facto choice. Existing methods, however, train the student to match the teacher's output indiscriminately, treating every rollout, frame, and pixel as equally reliable supervision. We argue that this caps distilled quality, since it overlooks two complementary axes of variance in DMD supervision: Inter-Reliability across student rollouts whose supervision varies in reliability, and Intra-Perplexity across spatial regions and temporal frames that contribute unequally to where quality can still be improved. The objective thus conflates two questions under a uniform weight: whether to learn from each rollout, and where to concentrate optimization within it. To address this, we propose Stream-R1, a Reliability-Perplexity Aware Reward Distillation framework that adaptively reweights the distillation objective at both rollout and spatiotemporal-element levels through a single shared reward-guided mechanism. At the Inter-Reliability level, Stream-R1 rescales each rollout's loss by an exponential of a pretrained video reward score, so that rollouts with reliable supervision dominate optimization. At the Intra-Perplexity level, it back-propagates the same reward model to extract per-pixel gradient saliency, which is factored into spatial and temporal weights that concentrate optimization pressure on regions and frames where refinement yields the largest expected gain. An adaptive balancing mechanism prevents any single quality axis from dominating across visual quality, motion quality, and text alignment. Stream-R1 attains consistent improvements on all three dimensions over distillation baselines on standard streaming video generation benchmarks, without architectural modification or additional inference cost.

Large language models (LMs) are capable of generating free-text rationales to aid question answering. However, prior work 1) suggests that useful self-rationalization is emergent only at significant scales (e.g., 175B parameter GPT-3); and 2) focuses largely on downstream performance, ignoring the semantics of the rationales themselves, e.g., are they faithful, true, and helpful for humans? In this work, we enable small-scale LMs (approx. 200x smaller than GPT-3) to generate rationales that not only improve downstream task performance, but are also more plausible, consistent, and diverse, assessed both by automatic and human evaluation. Our method, MaRio (Multi-rewArd RatIOnalization), is a multi-reward conditioned self-rationalization algorithm that optimizes multiple distinct properties like plausibility, diversity and consistency. Results on five difficult question-answering datasets StrategyQA, QuaRel, OpenBookQA, NumerSense and QASC show that not only does MaRio improve task accuracy, but it also improves the self-rationalization quality of small LMs across the aforementioned axes better than a supervised fine-tuning (SFT) baseline. Extensive human evaluations confirm that MaRio rationales are preferred vs. SFT rationales, as well as qualitative improvements in plausibility and consistency.

On-policy distillation (OPD), which aligns the student with the teacher's logit distribution on student-generated trajectories, has demonstrated strong empirical gains in improving student performance and often outperforms off-policy distillation and reinforcement learning (RL) paradigms. In this work, we first theoretically show that OPD is a special case of dense KL-constrained RL where the reward function and the KL regularization are always weighted equally and the reference model can by any model. Then, we propose the Generalized On-Policy Distillation (G-OPD) framework, which extends the standard OPD objective by introducing a flexible reference model and a reward scaling factor that controls the relative weight of the reward term against the KL regularization. Through comprehensive experiments on math reasoning and code generation tasks, we derive two novel insights: (1) Setting the reward scaling factor to be greater than 1 (i.e., reward extrapolation), which we term ExOPD, consistently improves over standard OPD across a range of teacher-student size pairings. In particular, in the setting where we merge the knowledge from different domain experts, obtained by applying domain-specific RL to the same student model, back into the original student, ExOPD enables the student to even surpass the teacher's performance boundary and outperform the domain teachers. (2) Building on ExOPD, we further find that in the strong-to-weak distillation setting (i.e., distilling a smaller student from a larger teacher), performing reward correction by choosing the reference model as the teacher's base model before RL yields a more accurate reward signal and further improves distillation performance. However, this choice assumes access to the teacher's pre-RL variant and incurs more computational overhead. We hope our work offers new insights for future research on OPD.

On-policy distillation (OPD) is widely used for LLM post-training. When pushed with a reward-extrapolation coefficient lambda > 1, the student can lift past the teacher in domain, but past a threshold lambda* the same step violates the output contract on structured-output tasks. In a single-position Bernoulli reduction, we derive a closed-form base-relative clip-safety threshold lambda*(p,b,c) determined by three measurable quantities: the teacher modal probability, the warm-start mass, and the importance-sampling clip strength. Above lambda*, the extrapolated fixed point exits the clip-safe region, changing training from format-preserving to format-collapsing. We extend the rule to calibrated K-ary listwise JSON tasks where a single binding equivalence class dominates the output contract and SFT retains parse headroom. On Amazon Fashion, three pre-registered tests--a fine-grid cliff interval, a budget-extension test, and a small-clip cross-prediction--fall within their locked prediction windows, with the small-clip value matching the closed-form prediction below grid resolution. Operating just below lambda*, ListOPD brings a 1.7B Qwen3 student to in-domain parity with an 8B-SFT baseline at one-fifth the parameters. The gain is driven primarily by format adherence: NDCG@1 on parsed outputs remains flat across lambda, while parse validity sharply changes at the predicted boundary. The cliff diagnostic is rubric-independent, whereas the parity claim uses a Gemini-graded rubric and inherits that evaluator's exposure.

Recent advances in few-step diffusion distillation have enabled efficient image generation, yet aligning these models with human preferences remains challenging. We propose Reward-Tilted Distribution Matching Distillation (RTDMD), a two-stage framework that unifies distribution matching distillation with reward-guided reinforcement learning for few-step flow generators. We show that minimizing the KL divergence to a reward-tilted teacher distribution naturally decomposes into a distribution matching term and a reward maximization term. In the first stage, we introduce Ambient-Consistent Distribution Matching Distillation (AC-DMD), which performs subinterval-wise distribution matching and augments the fake score objective with a consistency regularizer to help the fake score model track the shifting generator distribution under limited updates. In the second stage, we jointly optimize both terms: for the reward maximization term, we derive a hybrid policy gradient that combines a GRPO-style estimator for the stochastic intermediate transitions with direct reward backpropagation through the deterministic final step, and further introduce step-subset GRPO (SubGRPO) to reduce variance. Experiments on SD3, SD3.5, and FLUX.2 demonstrate that RTDMD establishes new state-of-the-art results across preference, aesthetic, and compositional metrics with only 4 inference steps, outperforming previous few-step text-to-image generation methods. Code and models are available at https://github.com/Harahan/RTDMD.

Flow-matching models are now a mainstream method to image generation, but its adaptation to diverse downstream scenarios typically relies on post-training, which may cause conflicts among task-specific optimization objectives. Reinforcement learning enables direct optimization of task-specific rewards beyond the original models, yet trajectory-level optimization may incur high-variance gradients and cross-task interference. On-policy distillation (OPD) offers dense and stable supervision on student rollouts, but conventional teacher matching remains imitation-based. We propose DreOPD, a Degraded-reference extrapolative OPD method for flow-matching models that bridges these two paradigms. Our DreOPD converts implicit reward extrapolation into closed-form velocity regression, enabling extrapolative post-training with the stability of OPD. It further uses a mildly degraded reference to strengthen the teacher-reference contrast, yielding a clearer extrapolation direction. Experiments on single- and multi-teacher settings show that DreOPD outperforms OPD and multi-task RL baselines in average performance, while surpassing specialized teachers on most metrics.

Recent years have witnessed remarkable progress in world models, which primarily aim to capture the spatio-temporal correlations between an agent's actions and the evolving environment. However, existing approaches often suffer from tight runtime coupling or depend on offline reward signals, resulting in substantial inference overhead or hindering end-to-end optimization. To overcome these limitations, we introduce WPT, a World-to-Policy Transfer training paradigm that enables online distillation under the guidance of an end-to-end world model. Specifically, we develop a trainable reward model that infuses world knowledge into a teacher policy by aligning candidate trajectories with the future dynamics predicted by the world model. Subsequently, we propose policy distillation and world reward distillation to transfer the teacher's reasoning ability into a lightweight student policy, enhancing planning performance while preserving real-time deployability. Extensive experiments on both open-loop and closed-loop benchmarks show that our WPT achieves state-of-the-art performance with a simple policy architecture: it attains a 0.11 collision rate (open-loop) and achieves a 79.23 driving score (closed-loop) surpassing both world-model-based and imitation-learning methods in accuracy and safety. Moreover, the student sustains up to 4.9x faster inference, while retaining most of the gains.

Recent advances in large multimodal models (LMMs) have enabled impressive reasoning and perception abilities, yet most existing training pipelines still depend on human-curated data or externally verified reward models, limiting their autonomy and scalability. In this work, we strive to improve LMM reasoning capabilities in a purely unsupervised fashion (without any annotated data or reward distillation). To this end, we propose a self-evolving framework, named EvoLMM, that instantiates two cooperative agents from a single backbone model: a Proposer, which generates diverse, image-grounded questions, and a Solver, which solves them through internal consistency, where learning proceeds through a continuous self-rewarding process. This dynamic feedback encourages both the generation of informative queries and the refinement of structured reasoning without relying on ground-truth or human judgments. When using the popular Qwen2.5-VL as the base model, our EvoLMM yields consistent gains upto sim3\% on multimodal math-reasoning benchmarks, including ChartQA, MathVista, and MathVision, using only raw training images. We hope our simple yet effective approach will serve as a solid baseline easing future research in self-improving LMMs in a fully-unsupervised fashion. Our code and models are available at https://github.com/mbzuai-oryx/EvoLMM.

On-policy self-distillation has emerged as a promising paradigm for post-training language models, in which the model conditions on environment feedback to serve as its own teacher, providing dense token-level rewards without external teacher models or step-level annotations. Despite its empirical success, what this reward actually measures and what kind of credit it assigns remain unclear. Under a posterior-compatibility interpretation of feedback conditioning, standard in the implicit-reward literature, we show that the self-distillation token reward is a Bayesian filtering increment whose trajectory sum is exactly the pointwise mutual information between the response and the feedback given the input. This pMI can be raised by input-specific reasoning or by input-generic shortcuts, so we further decompose the teacher log-probability along the input axis. Based on this analysis, we propose CREDIT (Contrastive REward from DIsTillation), which isolates the input-specific component with a batch-contrastive baseline. At the sequence level, CREDIT is a teacher-side surrogate for a contrastive pMI objective that also penalizes responses remaining likely under unrelated inputs. Across coding, scientific reasoning, and tool-use benchmarks on two model families, CREDIT delivers the strongest aggregate performance at negligible additional compute.

Generative models are transforming creative domains such as music generation, with inference-time strategies like Classifier-Free Guidance (CFG) playing a crucial role. However, CFG doubles inference cost while limiting originality and diversity across generated contents. In this paper, we introduce diversity-rewarded CFG distillation, a novel finetuning procedure that distills the strengths of CFG while addressing its limitations. Our approach optimises two training objectives: (1) a distillation objective, encouraging the model alone (without CFG) to imitate the CFG-augmented predictions, and (2) an RL objective with a diversity reward, promoting the generation of diverse outputs for a given prompt. By finetuning, we learn model weights with the ability to generate high-quality and diverse outputs, without any inference overhead. This also unlocks the potential of weight-based model merging strategies: by interpolating between the weights of two models (the first focusing on quality, the second on diversity), we can control the quality-diversity trade-off at deployment time, and even further boost performance. We conduct extensive experiments on the MusicLM (Agostinelli et al., 2023) text-to-music generative model, where our approach surpasses CFG in terms of quality-diversity Pareto optimality. According to human evaluators, our finetuned-then-merged model generates samples with higher quality-diversity than the base model augmented with CFG. Explore our generations at https://google-research.github.io/seanet/musiclm/diverse_music/.

Reinforcement Learning with Verifiable Rewards (RLVR) and on-policy distillation (OPD) have become two widely adopted paradigms for post-training large language models. However, RLVR suffers from sparse task-level feedback, while OPD provides dense token-level guidance but ignores trajectory correctness, limiting its performance to that of the teacher. Combining them is a promising direction: OPD supplies dense supervisory signals, while RLVR provides task-level correctness. Nevertheless, existing integrations often rely on weighted combination or heuristic switching, introducing extra hyperparameters and trade-offs. We propose On-policy Distillation with Verifiable Reward (OPDVR), a simple yet effective method that seamlessly combines OPD and RLVR without adding any hyperparameters. We first reformulate the implicit reward of sampled-token OPD based on trajectory correctness, then apply a ReLU gating mechanism to ensure that correct trajectories receive non-negative rewards and incorrect ones receive non-positive rewards---thereby aligning the distillation signal with task success while preserving the teacher's distributional guidance. Furthermore, our modification transforms sampled-token OPD into a proper RLVR method, making it readily combinable with any policy gradient algorithm, such as GRPO. Experiments on six reasoning benchmarks show that OPDVR consistently outperforms standard OPD. Our code is available at https://github.com/LeapLabTHU/OPDVR.

Diffusion probabilistic models have shown significant progress in video generation; however, their computational efficiency is limited by the large number of sampling steps required. Reducing sampling steps often compromises video quality or generation diversity. In this work, we introduce a distillation method that combines variational score distillation and consistency distillation to achieve few-step video generation, maintaining both high quality and diversity. We also propose a latent reward model fine-tuning approach to further enhance video generation performance according to any specified reward metric. This approach reduces memory usage and does not require the reward to be differentiable. Our method demonstrates state-of-the-art performance in few-step generation for 10-second videos (128 frames at 12 FPS). The distilled student model achieves a score of 82.57 on VBench, surpassing the teacher model as well as baseline models Gen-3, T2V-Turbo, and Kling. One-step distillation accelerates the teacher model's diffusion sampling by up to 278.6 times, enabling near real-time generation. Human evaluations further validate the superior performance of our 4-step student models compared to teacher model using 50-step DDIM sampling.

Reward models (RMs) play a pivotal role in aligning large language models (LLMs) with human preferences. Due to the difficulty of obtaining high-quality human preference annotations, distilling preferences from generative LLMs has emerged as a standard practice. However, existing approaches predominantly treat teacher models as simple binary annotators, failing to fully exploit the rich knowledge and capabilities for RM distillation. To address this, we propose RM-Distiller, a framework designed to systematically exploit the multifaceted capabilities of teacher LLMs: (1) Refinement capability, which synthesizes highly correlated response pairs to create fine-grained and contrastive signals. (2) Scoring capability, which guides the RM in capturing precise preference strength via a margin-aware optimization objective. (3) Generation capability, which incorporates the teacher's generative distribution to regularize the RM to preserve its fundamental linguistic knowledge. Extensive experiments demonstrate that RM-Distiller significantly outperforms traditional distillation methods both on RM benchmarks and reinforcement learning-based alignment, proving that exploiting multifaceted teacher capabilities is critical for effective reward modeling. To the best of our knowledge, this is the first systematic research on RM distillation from generative LLMs.

We introduce a novel approach to large language model (LLM) distillation by formulating it as a constrained reinforcement learning problem. While recent work has begun exploring the integration of task-specific rewards into distillation processes, existing methods typically rely on ad-hoc reward weighting. We propose a principled optimization framework that maximizes task-specific rewards while constraining the divergence from the teacher model to remain below a specified threshold. Our approach adapts constrained state augmented reinforcement learning to the distillation setting, introducing a modified reward function that maintains theoretical guarantees of constraint satisfaction without requiring state augmentation or teacher model access during deployment and without the computational overhead of the dual Lagrangian methods. Through extensive experiments on mathematical reasoning tasks, we demonstrate that our method achieves better constraint satisfaction rates and better reasoning compared to the soft Lagrangian relaxation baselines while maintaining competitive task performance. Our framework provides a theoretically grounded and practically efficient solution for reward-aware distillation in resource-constrained settings.

The landscape of high-performance image generation models is currently dominated by proprietary systems, such as Nano Banana Pro and Seedream 4.0. Leading open-source alternatives, including Qwen-Image, Hunyuan-Image-3.0 and FLUX.2, are characterized by massive parameter counts (20B to 80B), making them impractical for inference, and fine-tuning on consumer-grade hardware. To address this gap, we propose Z-Image, an efficient 6B-parameter foundation generative model built upon a Scalable Single-Stream Diffusion Transformer (S3-DiT) architecture that challenges the "scale-at-all-costs" paradigm. By systematically optimizing the entire model lifecycle -- from a curated data infrastructure to a streamlined training curriculum -- we complete the full training workflow in just 314K H800 GPU hours (approx. $630K). Our few-step distillation scheme with reward post-training further yields Z-Image-Turbo, offering both sub-second inference latency on an enterprise-grade H800 GPU and compatibility with consumer-grade hardware (<16GB VRAM). Additionally, our omni-pre-training paradigm also enables efficient training of Z-Image-Edit, an editing model with impressive instruction-following capabilities. Both qualitative and quantitative experiments demonstrate that our model achieves performance comparable to or surpassing that of leading competitors across various dimensions. Most notably, Z-Image exhibits exceptional capabilities in photorealistic image generation and bilingual text rendering, delivering results that rival top-tier commercial models, thereby demonstrating that state-of-the-art results are achievable with significantly reduced computational overhead. We publicly release our code, weights, and online demo to foster the development of accessible, budget-friendly, yet state-of-the-art generative models.

Streaming video understanding demands direct responses from the causally observed prefix of an unfolding video. Existing systems add inference-time memory, retrieval, and compression, yet a training-free sliding-window baseline already matches them. We therefore fix a memory-free recent-window protocol and ask how far post-training alone can go. Reinforcement learning with verifiable rewards fits this regime poorly, encouraging long ``think-then-answer'' generations, while on-policy distillation (OPD) supplies dense token-level teacher supervision on student trajectories but is stable only when both models train in thinking mode. These observations lead to StreamOPD, a recipe combining verifiable streaming-video data, thinking-mode OPD, and instruct-mode deployment. It raises StreamingBench from 77.9% to 83.9%---within 0.3 points of the 9B teacher---and improves OVO-Bench excluding its hallucination-detection subtask (HLD) by 9.1 points under unchanged inference. As a teacher-privilege extension, Spatio-Temporal CueGate (ST-CueGate) aggregates cue-versus-no-cue teacher likelihood ratios into a group-relative response score that reweights OPD. It reaches 71.9% on OVO-Bench (excluding HLD) and 64.9% on Video-MME, and is the only variant that stays above the base model on all four benchmarks. Replacing the teacher with a frozen copy of the student's initial policy---on-policy self-distillation---retains most of these gains and lifts HLD to 57.0%, above both the untrained student and the 9B teacher, so abstention loss is not intrinsic to the recipe. We provide a transparent and reproducible reference for open-source streaming-video research.

Reward fine-tuning aims to update a pre-trained flow-based generative model to improve the downstream reward of its generated samples. Existing methods typically formulate this problem as sampling from a reward-tilted distribution, the solution to a KL-regularized reward-maximization problem. Here, we introduce an optimal transport regularizer built directly from the pre-trained drift. Unlike KL reward tilting, the resulting objective transports individual samples toward higher reward rather than reweighting the base distribution. We show that the resulting problem is equivalent to a deterministic optimal control problem on the flow. Given a pre-trained flow map, this equivalence yields a simulation-free reinforcement learning algorithm for fine-tuning generative flows. We call the resulting framework Wasserstein-Tilted Flow Maps (WTF), the first end-to-end fine-tuning recipe native to flow maps. The output is a fine-tuned flow map that retains strong reward-aligned performance at few-step inference budgets without post-hoc distillation. Experiments on ImageNet-256 and text-to-image show that WTF achieves higher reward with comparable or higher diversity than baselines, while requiring up to 280times less training compute. More broadly, we argue that accelerated samplers such as flow maps are essential infrastructure for efficient post-training, and that the dominant KL-regularized formulation is only one of many choices worth revisiting.

Existing Large Reasoning Models (LRMs) have shown the potential of reinforcement learning (RL) to enhance the complex reasoning capabilities of Large Language Models~(LLMs). While they achieve remarkable performance on challenging tasks such as mathematics and coding, they often rely on their internal knowledge to solve problems, which can be inadequate for time-sensitive or knowledge-intensive questions, leading to inaccuracies and hallucinations. To address this, we propose R1-Searcher, a novel two-stage outcome-based RL approach designed to enhance the search capabilities of LLMs. This method allows LLMs to autonomously invoke external search systems to access additional knowledge during the reasoning process. Our framework relies exclusively on RL, without requiring process rewards or distillation for a cold start. % effectively generalizing to out-of-domain datasets and supporting both Base and Instruct models. Our experiments demonstrate that our method significantly outperforms previous strong RAG methods, even when compared to the closed-source GPT-4o-mini.

Training reasoning language models (LMs) with reinforcement learning (RL) for one-hot correctness inherently relies on the LM being able to explore and solve its task with some chance at initialization. Furthermore, a key use case of reasoning LMs is to act as teachers for distilling new students and cold-starting future RL iterations rather than being deployed themselves. From these considerations, we introduce a new framework that avoids RL's exploration challenge by training a new class of Reinforcement-Learned Teachers (RLTs) focused on yielding the most effective downstream distillation. RLTs are prompted with both the question and solution to each problem, and tasked to simply "connect-the-dots" with detailed explanations tailored for their students. We train RLTs with dense rewards obtained by feeding each explanation to the student and testing its understanding of the problem's solution. In practice, the raw outputs of a 7B RLT provide higher final performance on competition and graduate-level tasks than existing distillation and cold-starting pipelines that collect and postprocess the reasoning traces of orders of magnitude larger LMs. Furthermore, RLTs maintain their effectiveness when training larger students and when applied zero-shot to out-of-distribution tasks, unlocking new levels of efficiency and re-usability for the RL reasoning framework.

Latent Consistency Distillation (LCD) has emerged as a promising paradigm for efficient text-to-image synthesis. By distilling a latent consistency model (LCM) from a pre-trained teacher latent diffusion model (LDM), LCD facilitates the generation of high-fidelity images within merely 2 to 4 inference steps. However, the LCM's efficient inference is obtained at the cost of the sample quality. In this paper, we propose compensating the quality loss by aligning LCM's output with human preference during training. Specifically, we introduce Reward Guided LCD (RG-LCD), which integrates feedback from a reward model (RM) into the LCD process by augmenting the original LCD loss with the objective of maximizing the reward associated with LCM's single-step generation. As validated through human evaluation, when trained with the feedback of a good RM, the 2-step generations from our RG-LCM are favored by humans over the 50-step DDIM samples from the teacher LDM, representing a 25 times inference acceleration without quality loss. As directly optimizing towards differentiable RMs can suffer from over-optimization, we overcome this difficulty by proposing the use of a latent proxy RM (LRM). This novel component serves as an intermediary, connecting our LCM with the RM. Empirically, we demonstrate that incorporating the LRM into our RG-LCD successfully avoids high-frequency noise in the generated images, contributing to both improved FID on MS-COCO and a higher HPSv2.1 score on HPSv2's test set, surpassing those achieved by the baseline LCM.

On-policy distillation (OPD) has recently emerged as a popular post-training paradigm for large language models (LLMs), providing an efficient way to transfer the knowledge and capabilities of teacher models into student models. However, teacher guidance on student-generated prefixes is not always reliable. Training should optimize the model to generate responses that are more likely to be correct, or equivalently, to get higher outcome rewards. But during OPD, the teacher model may provide guidance that discourages the student from moving toward correct trajectories or moves the student toward incorrect ones, which is misaligned with outcome reward. Such misaligned guidance is unreliable, as it would mislead the optimization process and ultimately degrade model performance. To mitigate misaligned teacher guidance, we propose Reward-Aligned On-Policy Distillation (RA-OPD). The key insight is to keep only trajectories whose induced updates move the student toward correct trajectories or discourage the student from moving toward incorrect ones. Specifically, for each sampled trajectory, RA-OPD checks whether its trajectory-level distillation return is consistent with its outcome reward and then filters out the misaligned trajectories. RA-OPD selects more reliable trajectories to improve student model performance without requiring additional computational cost. We evaluate RA-OPD on math and code benchmarks using models from the Qwen3 family and the DeepSeek-R1 family. Across seven math benchmarks and three code benchmarks, RA-OPD significantly outperforms standard OPD and other tested OPD variants.

Audio-language models (ALMs) can exploit textual shortcuts to answer questions while overlooking acoustic evidence, weakening audio understanding. On-policy distillation (OPD) trains compact ALMs by supervising student-generated responses with teacher predictions, but does not explicitly distinguish acoustic support from linguistic predictability. We propose Reward-Tilted On-Policy Distillation (RT-OPD) to strengthen acoustic grounding. Given the same question and student-generated text, a frozen teacher predicts the next token with and without audio inputs. Their log-probability contrast defines a reward that reshapes the teacher distribution for reverse-KL distillation, emphasizing the additional evidence provided by audio. Across two compact students and three benchmarks, RT-OPD consistently outperforms Vanilla OPD. Experiments with silenced and replacement audio further suggest that RT-OPD strengthens the student's reliance on acoustic evidence. Our 3B model achieves 72.72% accuracy on MMAU, the highest among the compared 3B models and competitive with several 7B and 8B models. Code and model checkpoints are available at https://github.com/KaiyangLi1992/RT-OPD.

One-step generators enable high-quality visual generation with a single network evaluation, but their post-training is difficult: general implicit generators provide neither tractable likelihoods nor denoising trajectories, and many rewards are non-differentiable. We introduce Reward-Weighted Transport Distillation (RWTD), a post-training method that requires only generated samples and scalar reward evaluations. Rather than aligning solely to the conventional reward-tilted reference distribution, RWTD constructs an adaptive target that mixes separately tilted current and reference distributions. The current component incorporates improvements discovered during training, while the reference component anchors the target to the pretrained generator. RWTD realizes this target through feature-space optimal transport and fixed-point regression. Theoretical analysis shows that the fixed-point distributions of RWTD interpolate between off-policy reward tilting of the reference and on-policy tilting of the current model, providing a principled approach to balancing reward adaptation with retention of prior knowledge. Empirically, RWTD substantially improves the GenEval score of the one-step SANA Sprint 1.6B backbone from 0.73 to 0.80, while separate preference alignment experiments demonstrate strong cross-reward generalization that yields balanced improvements and preservation of compositional capabilities.

Current post-training methods in verifiable settings fall into two categories. Reinforcement learning (RLVR) relies on binary rewards, which are broadly applicable and powerful, but provide only sparse supervision during training. Distillation provides dense token-level supervision, typically obtained from an external teacher or using high-quality demonstrations. Collecting such supervision can be costly or unavailable. We propose Self-Distillation Zero (SD-Zero), a method that is substantially more training sample-efficient than RL and does not require an external teacher or high-quality demonstrations. SD-Zero trains a single model to play two roles: a Generator, which produces an initial response, and a Reviser, which conditions on that response and its binary reward to produce an improved response. We then perform on-policy self-distillation to distill the reviser into the generator, using the reviser's token distributions conditioned on the generator's response and its reward as supervision. In effect, SD-Zero trains the model to transform binary rewards into dense token-level self-supervision. On math and code reasoning benchmarks with Qwen3-4B-Instruct and Olmo-3-7B-Instruct, SD-Zero improves performance by at least 10% over the base models and outperforms strong baselines, including Rejection Fine-Tuning (RFT), GRPO, and Self-Distillation Fine-Tuning (SDFT), under the same question set and training sample budget. Extensive ablation studies show two novel characteristics of our proposed algorithm: (a) token-level self-localization, where the reviser can identify the key tokens that need to be revised in the generator's response based on reward, and (b) iterative self-evolution, where the improving ability to revise answers can be distilled back into generation performance with regular teacher synchronization.

We aim to improve model performance in multi-reward reinforcement learning training process. Existing Group reward-Decoupled Normalization Policy Optimization (GDPO) has mitigated the issue of reward signals masking one another during direct scalarization by normalizing each reward dimension separately before aggregation. However, our experiments show that GDPO still struggles to balance reward signals with different granularities. Specifically, in some particular training tasks, the model may receive a dense reward that assigns fine-grained scores ranging from 0.1 to 1.0, together with a sparse reward that provides only binary feedback of either 0 or 1. In such cases, we find that the sparse reward may provide an insufficient optimization signal, preventing its corresponding capability from being effectively reinforced. Therefore, how can we strengthen the optimization signal from the sparse reward without sacrificing the capability already learned from the fine-grained reward? To overcome this limitation, we propose Specialize-and-Merge Online Policy Distillation (SMOPD), a two-stage training method for multi-reward optimization. Stage1-Specialize: SMOPD first employs reward-priority configurations to train multiple reward-specialized teachers, allowing each reward to be learned under conditions where its signal can effectively drive optimization. Stage2-Merge: SMOPD then utilizes online policy distillation to combine the reward-specialized capabilities of these teachers into a single student policy, while maintaining balanced task-level optimization. To validate our method, we conduct experiments on two multi-reward settings: complementary rewards(tool-calling accuracy and format) and conflicting rewards (helpful and harmless rewards). Based on above settings, SMOPD outperforms GDPO across 1.5B, 3B and 7B backbones.

Efficient streaming video generation is critical for simulating interactive and dynamic worlds. Existing methods distill few-step video diffusion models with sliding window attention, using initial frames as sink tokens to maintain attention performance and reduce error accumulation. However, video frames become overly dependent on these static tokens, resulting in copied initial frames and diminished motion dynamics. To address this, we introduce Reward Forcing, a novel framework with two key designs. First, we propose EMA-Sink, which maintains fixed-size tokens initialized from initial frames and continuously updated by fusing evicted tokens via exponential moving average as they exit the sliding window. Without additional computation cost, EMA-Sink tokens capture both long-term context and recent dynamics, preventing initial frame copying while maintaining long-horizon consistency. Second, to better distill motion dynamics from teacher models, we propose a novel Rewarded Distribution Matching Distillation (Re-DMD). Vanilla distribution matching treats every training sample equally, limiting the model's ability to prioritize dynamic content. Instead, Re-DMD biases the model's output distribution toward high-reward regions by prioritizing samples with greater dynamics rated by a vision-language model. Re-DMD significantly enhances motion quality while preserving data fidelity. We include both quantitative and qualitative experiments to show that Reward Forcing achieves state-of-the-art performance on standard benchmarks while enabling high-quality streaming video generation at 23.1 FPS on a single H100 GPU.

Diffusion models achieve state-of-the-art generative performance but are fundamentally bottlenecked by their slow, iterative sampling process. While diffusion distillation techniques enable high-fidelity, few-step generation, traditional objectives often restrict the student's performance by anchoring it solely to the teacher. Recent approaches have attempted to break this ceiling by integrating Reinforcement Learning (RL), typically through a simple summation of distillation and RL objectives. In this work, we propose a novel paradigm by re-conceptualizing distribution matching as a reward, denoted as R_dm. This unified perspective bridges the algorithmic gap between Diffusion Matching Distillation (DMD) and RL, providing several primary benefits. (1) Enhanced Optimization Stability: We introduce Group Normalized Distribution Matching (GNDM), which adapts standard RL group normalization to stabilize R_dm estimation. By leveraging group-mean statistics, GNDM establishes a more robust and effective optimization direction. (2) Seamless Reward Integration: Our reward-centric formulation inherently supports adaptive weighting mechanisms, allowing for the fluid combination of DMD with external reward models. (3) Improved Sampling Efficiency: By aligning with RL principles, the framework readily incorporates Importance Sampling (IS), leading to a significant boost in sampling efficiency. Extensive experiments demonstrate that GNDM outperforms vanilla DMD, reducing the FID by 1.87. Furthermore, our multi-reward variant, GNDMR, surpasses existing baselines by striking an optimal balance between aesthetic quality and fidelity, achieving a peak HPS of 30.37 and a low FID-SD of 12.21. Ultimately, R_dm provides a flexible, stable, and efficient framework for real-time, high-fidelity synthesis. Codes are coming soon.

Agentic on-policy distillation (OPD) trains a student agent on its own trajectories while a teacher scores every token. A natural explanation of where it goes wrong is compounding error: the student's mistakes carry its trajectories beyond the region where the teacher is reliable, so recent methods keep training inside that region, through episode-length curricula, teacher-guided turns, prefix replay, or teacher intervention where the two disagree. We show this explanation is incomplete. From the states where the student departs most from the teacher, the teacher's own takeovers still finish the task, yet the student trained on its token scores keeps running out of turns. We attribute this to the myopia of next-token scoring. The teacher rates each token without seeing where the episode ends, so its scores let through the safe, delaying moves that outcome-based RL would punish with zero reward, a penalty agentic OPD lacks. We call this difference between what the teacher does and what it scores the rollout-scorer gap. We propose Rollout-to-Scorer Distillation (R2S), which repairs the scorer before distillation by mining high-disagreement states from the untrained student, running teacher takeovers from them, keeping the states where enough takeovers pass a reward-free check such as consensus or an LLM judge, and fine-tuning the teacher to give its takeover-informed scores on every assistant token from the divergence turn onward, without the takeover in context. The repaired teacher is frozen as the sole scorer for standard OPD, requiring no task reward during repair or training. Across three environments, two teachers and four students, R2S outperforms three state-of-the-art agentic-OPD methods in all twelve teacher-student-environment cells, by 12.8 success points on average over the best of them in each cell. Project page: https://tree-shu-zhao.github.io/r2s/.

In settings where labeled verifiable training data is the binding constraint, each checked example should be allocated carefully. The standard practice is to use this data directly on the model that will be deployed, for example by running GRPO on the deployment student. We argue that this is often an inefficient allocation because it overlooks a reward-density principle: sparse sequence-level reward should train models where exploration is productive, while dense token-level teacher reward should be used where the aim is to compress behavior into a smaller model. In this view, GRPO-style sparse RL and OPD-style dense teacher supervision are not separate recipes; they are different reward-density regimes. The allocation rule is simple: use scarce labeled training data upstream on the strongest model that can turn it into reward-shaped behavior, then transfer that behavior downstream as dense supervision. We evaluate this rule on verifiable math with Qwen3 and Llama models. At fixed Qwen3-1.7B deployment-student size, an RL-improved 8B teacher distilled through the dense bridge outperforms direct GRPO on the same student, while transfer from the same teacher before RL underperforms. The bridge is important: a forward-KL warmup on teacher rollouts followed by OPD on student rollouts is consistently strongest on MATH before any post-bridge student-side sparse RL, and also gives the best pre-Stage~3 AIME endpoints for the canonical 8B/14B teachers. The bridge also makes later student-side sparse RL effective: GRPO that is weak on a cold student lifts MATH from 75.4% to 78.5% after the bridge and outperforms a matched replay control by 2.8 points. The operational principal is to avoid using scarce labeled data on the least prepared policy: use sparse reward for teacher-side discovery, dense transfer for student compression, and student-side sparse reward only after the bridge.

Aligning generated images to complicated text prompts and human preferences is a central challenge in Artificial Intelligence-Generated Content (AIGC). With reward-enhanced diffusion distillation emerging as a promising approach that boosts controllability and fidelity of text-to-image models, we identify a fundamental paradigm shift: as conditions become more specific and reward signals stronger, the rewards themselves become the dominant force in generation. In contrast, the diffusion losses serve as an overly expensive form of regularization. To thoroughly validate our hypothesis, we introduce R0, a novel conditional generation approach via regularized reward maximization. Instead of relying on tricky diffusion distillation losses, R0 proposes a new perspective that treats image generations as an optimization problem in data space which aims to search for valid images that have high compositional rewards. By innovative designs of the generator parameterization and proper regularization techniques, we train state-of-the-art few-step text-to-image generative models with R0 at scales. Our results challenge the conventional wisdom of diffusion post-training and conditional generation by demonstrating that rewards play a dominant role in scenarios with complex conditions. We hope our findings can contribute to further research into human-centric and reward-centric generation paradigms across the broader field of AIGC. Code is available at https://github.com/Luo-Yihong/R0.

Reinforcement learning can align diffusion models with human preferences and task-specific objectives, but endpoint rewards do not specify how an intermediate denoising prediction should change. We introduce DiffusionOPSD as an on-policy self-distillation framework that converts image-level reward guidance into explicit targets for clean-output predictions at sampled queries. At each outer iteration, a frozen behavior policy generates trajectories and supplies query states and anchors. Reward gradients construct bounded positive and negative targets around each anchor. The trainable policy fits these targets as detached supervision through finite fitting before an exponential moving average update refreshes the behavior policy. This setup lets us measure target construction and finite realization separately. Controlled same-query experiments show that larger target-construction gains do not necessarily translate into larger realized gains after a single fitting update. Across SD 3.5-M and the step-distilled Z-Image-Turbo, our approach achieves the best final held-out scores in 19 of 20 reward-matched settings across two backbones and ten evaluators. It outperforms the strongest competing method by up to 44.0% and reduces training GPU-hours relative to DiffusionNFT by 40% on SD 3.5-M and 63% on Z-Image-Turbo. These results support on-policy self-distillation as an efficient and analyzable approach to diffusion post-training by converting image-level reward guidance into explicit and continually refreshed intermediate supervision, thereby opening a path toward more efficient and diagnosable alignment.

Deep search agents operate over trajectories spanning dozens of steps, yet standard reinforcement learning provides only a single outcome reward per trajectory, which is far too sparse for effective credit assignment. On-policy self-distillation (OPSD) addresses this by using the model's own logits as dense token-level teachers, but extending it to search agents introduces a fundamental tension: the teacher, having access to privileged information such as the correct answer, produces a distribution that differs systematically from the student's exploration-based reasoning, and naive distillation causes the student to inherit this information asymmetry rather than learn better search strategies. We resolve this tension through two contributions. First, we construct Evidence Anchors, which are concise, step-level evidence snippets extracted from the web, as privileged information that captures key reasoning steps without revealing the entire answer path. Second, we propose Step-Level Self-Distilled Policy Optimization (SSPO), which converts teacher-student disagreement into step-level advantage weights within GRPO, applied exclusively to incorrect trajectories. This design decouples what to update from how much to update: the outcome reward determines the direction of policy change, while the teacher modulates its magnitude at each step. Correct trajectories are left untouched, preserving their diversity. On Qwen3-8B, SSPO consistently outperforms GRPO across BrowseComp, GAIA, and FRAMES, surpassing or matching GRPO trained with twice as many gradient steps while adding only about 5 percent overhead per step from a single additional forward pass.

Reinforcement learning for multimodal large language models (MLLMs) is often hindered by severe reward sparsity in complex reasoning tasks. This challenge is particularly pronounced in human-centered scenarios involving states, emotions, intentions, and behaviors, where heterogeneous multimodal signals and subjective human factors make high-quality chain-of-thought (CoT) annotations expensive and difficult to obtain. Although many multimodal datasets provide expert-annotated ground-truth labels, directly using these labels for supervised fine-tuning may encourage shortcut learning in multimodal perception and provides limited transparency for safety-critical human--AI interaction. To address these limitations, we propose OmniOPSD, a Rationale-Privileged On-Policy Self-Distillation framework that uses frontier-generated rationales as teacher-side privileged evidence rather than student imitation targets. OmniOPSD uses frontier-generated evidence-aware rationales only as training-time privileged evidence context for a local teacher. The student samples its own rollout from the original multimodal input, while the rationale-privileged teacher scores the same tokens and provides dense token-level supervision. Thus, the student learns on its own trajectory distribution without directly imitating frontier-model completions, and inference requires no labels, rationales, CoT annotations, or closed-source model access. Experiments on MER-UniBench show that OmniOPSD achieves state-of-the-art performance with an average score of 84.19, and ablations further support the value of rationale-privileged teacher guidance.

Deciding whether a trajectory actually fulfills its instruction governs how we measure computer-use agents on long-horizon graphical-user-interface tasks and how we train them with reinforcement learning. This judgment has long relied on rule-based evaluation, which struggles to align with human intention and goes stale when an app updates or its online content drifts. Existing model-based judges attempt to address these problems but still leave a performance gap to the rule-based evaluation. We propose the SeekJudge framework, in which four role-specialized agents, a Condense, a Ground, a Seek and an Analyze agent, reach a verdict through a Seek--Analyze loop over the trajectory. A seed-calibrated distillation pipeline trains one specialized 9B model to serve as the shared backbone for all four agents. Measured by downstream success rate on held-out RL test goals, SeekJudge is the first practical model-based reward to match or surpass native rule-based supervision in online RL. Beyond accuracy, SeekJudge provides step-level judgments, runs far cheaper than a closed-source large model, and keeps a small per-call context that scales to much longer trajectories. We further contribute a general architectural improvement to the reward server that speeds up judging in RL. Together these make model-based reward a practical drop-in for rule-based supervision in CUA reinforcement learning.

On-policy distillation (OPD), which leverages a pre-trained, specialized teacher model to provide dense supervisory signals, has achieved significant success in Large Language Models (LLMs) and has recently been adapted to flow matching models. However, this paradigm suffers from two major issues: First, training a separate, task-specific teacher for every new objective incurs high computational costs. Second, the discrepancy between teacher and student distributions often leads to compounding errors along the generation trajectory. In this paper, we introduce Self-OPD, a teacher-free OPD framework for flow matching models that turns the student's own self-exploration into step-wise supervision. At each timestep, Self-OPD branches the deterministic next-state prediction into K stochastic SDE candidates, rolls them out with the ODE sampler, and compares their rewards against a deterministic self-reference baseline to obtain normalized advantages. The velocity field is optimized with an all-branch pull-push objective, where high-advantage branches attract the student and low-advantage branches repel it under direction-aware attenuation and SDE-variance normalization. For multi-objective alignment, Self-OPD fuses normalized scores at the reward level, avoiding direct gradient conflict. Experiments on single and mixed reward benchmarks show that Self-OPD outperforms prior RL and OPD methods without task-specific teachers.

On-policy distillation (OPD) can succeed with surprisingly few training questions, motivating a view of questions as carriers of distributional knowledge transfer. How can we assess whether a newly generated question is valuable for distillation, and learn to generate better questions---without access to its correct answer? We introduce Ask-to-Distill (\method), a framework that treats synthetic questions as distribution elicitors and optimizes the teacher supervision they expose along student-generated trajectories. A student-initialized synthesizer is optimized using a distributional surrogate for question value, computed from a frozen student--teacher pair without reference answers, correctness rewards, or per-candidate student updates. The reward combines teacher--student disagreement with teacher probability mass on the student's candidate support, together with length and diversity controls. The generated questions are filtered and deduplicated into a prompt-only bank for standard OPD. Across mathematical reasoning, code generation, and agentic tasks, our method improves average performance and teacher-gap recovery over unoptimized student-generated prompts and KL-filtered real questions, while matching or exceeding real-bank OPD in the evaluated settings. Signal analyses further show that optimized questions can elicit denser supervision and more total signal from shorter student responses, supporting elicited distributional supervision as an objective for synthetic distillation data. Our code, data and models are publicly available at https://github.com/GasolSun36/Ask_to_Distill.

Diffusion-based dataset distillation (DD) suffers from a fundamental objective mismatch: likelihood-driven diffusion models prioritize density approximation over the discriminative decision boundaries required for downstream tasks. Beyond semantic mismatch, relying solely on density also leads to geometric coverage loss, where generated samples collapse into a few high-density modes and fail to cover the manifold's structural diversity. We propose Manifold-Guided Policy Optimization (MGPO), which reformulates DD as a multi-objective reinforcement learning problem and achieves Dual-Space Alignment via a pixel-space discriminative reward and a latent-space geometric reward guided by a class-wise Minimum Spanning Tree (MST). The discriminative reward enforces class separability, while the MST-based geometric reward encourages generated latents to cover a sparse geometric skeleton of each class, jointly addressing both failure modes. We further provide an idealized analysis that motivates the MST-based reward, including a Hausdorff approximation bound and a subsampling bound independent of the dataset size. The reward-modular design extends to structured tasks such as object detection and segmentation by substituting the frozen task reward model. Extensive experiments show MGPO consistently outperforms existing methods, including a +8.0% mIoU gain on segmentation under low-budget settings.

Score Distillation Sampling (SDS) enables high-quality text-to-3D generation by supervising 3D models through the denoising of multi-view 2D renderings, using a pretrained text-to-image diffusion model to align with the input prompt and ensure 3D consistency. However, existing SDS-based methods face two fundamental limitations: (1) their reliance on CLIP-style text encoders leads to coarse semantic alignment and struggles with fine-grained prompts; and (2) 2D diffusion priors lack explicit 3D spatial constraints, resulting in geometric inconsistencies and inaccurate object relationships in multi-object scenes. To address these challenges, we propose VLM3D, a novel text-to-3D generation framework that integrates large vision-language models (VLMs) into the SDS pipeline as differentiable semantic and spatial priors. Unlike standard text-to-image diffusion priors, VLMs leverage rich language-grounded supervision that enables fine-grained prompt alignment. Moreover, their inherent vision language modeling provides strong spatial understanding, which significantly enhances 3D consistency for single-object generation and improves relational reasoning in multi-object scenes. We instantiate VLM3D based on the open-source Qwen2.5-VL model and evaluate it on the GPTeval3D benchmark. Experiments across diverse objects and complex scenes show that VLM3D significantly outperforms prior SDS-based methods in semantic fidelity, geometric coherence, and spatial correctness.

Reinforcement learning with verifiable rewards (RLVR) has substantially improved the reasoning capabilities of large language models on tasks such as mathematical reasoning and code generation. However, most RLVR methods assign a scalar outcome reward to an entire trajectory, resulting in sparse supervision and limited token-level credit assignment. On-policy distillation (OPD) provides denser supervision by distilling token-level distributions from a stronger teacher model, but requires an additional teacher and typically assumes a shared vocabulary. On-policy self-distillation (OPSD) removes this dependency by conditioning the same model on privileged information to construct a teacher policy. However, directly matching the teacher distribution may cause information leakage and unstable optimization. RLSD avoids direct matching by using the teacher signal only to modulate update magnitudes, but it cannot provide an explicit correction direction when the sampled reasoning fails. To address this tradeoff, we introduce H^{2}SD, a hybrid hindsight self distillation framework that uses the teacher differently according to trajectory correctness. For successful trajectories, the teacher receives the student response confirmed as correct together with a rephrasing instruction, and its probabilities on the original response tokens are used to modulate update magnitudes without changing the direction determined by the reward. For failed trajectories, we condition the teacher on a reference hint containing key reasoning steps and a verified answer, and minimize the reverse KL divergence from the student to the teacher. Experiments on multiple challenging reasoning benchmarks show that H^2SD consistently outperforms representative RLVR, OPSD, and RLSD baselines while maintaining stable optimization and favorable generation efficiency.

Reinforcement learning with verifiable rewards (RLVR), along with recent selfdistillation variants such as SDPO, evaluates each rollout against a verifier and updates the policy from that episode-level signal. However, the richer procedural information in the rollout is rarely retained or reused. Across episodes and epochs, the model repeatedly encounters related problems under a changing policy, producing cross-episode signals that episode-local updates cannot capture: which strategies consistently pass verification, which failure modes persist, which patterns recur. We propose Procedural Memory Distillation (PMD), which converts these crossepisode signals into reusable procedural memory and distills it into the policy's weights during training. This memory functions as a training scaffold, absorbed into the policy itself, yielding a memory-free model at inference. PMD organizes the memory at three levels of abstraction: raw trajectories, self-reflected strategies and lessons, and higher-level behavioral patterns that recur across problems, all extracted online from the model's own trajectories. A memory-conditioned self-teacher draws on the accumulated experience to supervise the student on its own rollouts, enabling student to progressively internalize procedural knowledge within its parameters. The central design principle is co-evolution: the policy generates rollouts that update the memory, and memory shapes the supervision that updates the policy. Empirically, across Qwen3-8B and OLMo3-Instruct-7B, PMD improves over SDPO by 3.8-5.5% on SCIKNOWEVAL and 7.9-13.6% on LIVECODEBENCH. Co-evolution powers these gains: freezing either the memory or the policy trails PMD by more than 10% across SCIKNOWEVAL domains.

Reinforcement learning (RL) with verifiable rewards constructs trajectory-level advantage estimates, yet it often fails to credit the few pivotal decisions that determine outcomes in long-horizon, multi-turn agentic tasks. Recent work introduces privileged self-distillation for credit assignment, providing denser supervision, but it remains unclear how such local signals should represent sequential credit. We propose AgentOPSD, a critic-free, recursive method for turn-level credit assignment in agentic reinforcement learning. AgentOPSD aggregates token-level teacher-student log-probability gaps into turn-level evidence and recursively updates a Bayesian belief state in log-odds space. This yields a principled reweighting scheme that converts sparse outcome supervision into turn-level credit signals and identifies pivotal turns through the marginal belief revision between consecutive states. The method is fully compatible with standard policy optimization and requires neither an additional critic nor extra rollouts. We evaluate AgentOPSD on ALFWorld, WebShop, and Search-QA using Qwen2.5 models at two scales (3B and 7B). AgentOPSD outperforms GRPO and strong self-distillation baselines, achieving 89.1% success on ALFWorld with Qwen2.5-7B. Ablation studies attribute the gains to turn-level aggregation and history-dependent recursive belief updates.

Reinforcement learning with verifiable rewards (RLVR) turns agent experience into learning signals primarily through scalar outcome rewards after interaction. For group-relative objectives, however, this signal vanishes when all rollouts receive the same reward, even though their trajectories may reveal useful information about what the task requires and how the agent fails. We ask a complementary question: can hindsight teach an agent what it could have anticipated before acting? We introduce prospective learning, which uses post-hoc experience to supervise foresight predictions from the pre-interaction view, and instantiate it with Self-Retrospection Distillation (SRD). Intuitively, a completed trajectory reveals knowledge that would have been useful and pitfalls that should be avoided; SRD distills this privileged hindsight into trajectory-blind foresight of the same policy. Foresight serves only as a training target and need not be explicitly generated at inference time. Across 10 tool-integrated reasoning and long-horizon agentic tasks, SRD complements RLVR and self-distillation baselines with gains of up to 24.2 pp. Its advantage is especially pronounced when reward contrast is scarce: when 37--98% of rollout groups are reward-uniform across model scales, yet SRD can still exploit learning signal from sampled trajectories. In the 2B setting, where 98% of groups are all-failure, the RLVR training ends up at 0.0% success, while adding SRD reaches 60.6% under the same rollout budget. Our results suggest that post-hoc agent experience is useful not only for evaluating or improving behavior, but also for shaping predictive representations before available interaction.

Reinforcement learning with verifiable rewards has become a scalable post-training paradigm for large language models. However, training agents over long-horizon interactions presents two challenges: using only terminal rewards makes turn-level credit assignment sparse and high-variance, while sequential mini-batch updates on a fixed rollout batch introduce policy staleness as the current policy drifts from the behavior policy. Existing approaches obtain denser credit via learned critics, process reward models, or on-policy distillation, but they rely on a value network, step-level annotations, or a stronger external teacher that increase training complexity. Meanwhile, importance sampling and clipping mitigate policy drift, but they do not specify a finite target that stops sequential updates from overshooting. In this work, we introduce TreeOPSD, a critic-free, tree-aware posterior self-distillation framework for long-horizon agentic reinforcement learning. Treating each interaction turn as a tree node, TreeOPSD branches alternative continuations from shared prefixes and recursively backs up terminal rewards into implicit turn-level credit. It then maps each credit to a closed-form, bounded teacher distribution parameterized by the likelihood ratio relative to the frozen behavior policy. Minimizing the divergence between this teacher and the student policy recovers the policy-gradient direction as the tangent at the behavior policy while bounding updates toward a finite target across sequential mini-batches. Empirically, TreeOPSD improves agentic capabilities across several competitive benchmarks and model scales, especially for long-horizon tasks, e.g., raising the average accuracy of Qwen3-30B-A3B-Thinking-2507 by 4.97 points over GRPO on the Deep Research setting.

Reinforcement Learning with Verifiable Rewards (RLVR) improves reasoning in language models, but scalar outcomes provide coarse supervision over response trajectories. Self-distillation offers denser token-level guidance by comparing an unassisted policy with a feedback-informed self-teacher, yet the teacher evolves with the student, creating unstable targets and motivating additional regularization. We introduce Function-Space Policy Optimization (FSPO), a single-policy framework that combines current privileged adaptation with function-space temporal consolidation. FSPO stores sparse historical next-token distributions as fixed temporal targets, providing temporal separation without maintaining a persistent teacher model. It further introduces privileged disagreement correction to explicitly suppress locally overvalued actions, reliability-adaptive grounding that strengthens reference supervision when self-generated evidence repeatedly fails, and quality and coverage-aware temporal memory that retains verified behaviors and improves privileged demonstrations. Across reasoning and tool-use benchmarks, FSPO achieves the highest average accuracy under a 5-hour training budget, improving over the strongest baseline by approximately 4% points on average while reducing peak memory from about 70 GB to 56 GB on Olmo3-7B-Instruct.

Reward modeling is essential for aligning large language models (LLMs) with human preferences, especially through reinforcement learning from human feedback (RLHF). To provide accurate reward signals, a reward model (RM) should stimulate deep thinking and conduct interpretable reasoning before assigning a score or a judgment. However, existing RMs either produce opaque scalar scores or directly generate the prediction of a preferred answer, making them struggle to integrate natural language critiques, thus lacking interpretability. Inspired by recent advances of long chain-of-thought (CoT) on reasoning-intensive tasks, we hypothesize and validate that integrating reasoning capabilities into reward modeling significantly enhances RM's interpretability and performance. In this work, we introduce a new class of generative reward models -- Reasoning Reward Models (ReasRMs) -- which formulate reward modeling as a reasoning task. We propose a reasoning-oriented training pipeline and train a family of ReasRMs, RM-R1. The training consists of two key stages: (1) distillation of high-quality reasoning chains and (2) reinforcement learning with verifiable rewards. RM-R1 improves LLM rollouts by self-generating reasoning traces or chat-specific rubrics and evaluating candidate responses against them. Empirically, our models achieve state-of-the-art or near state-of-the-art performance of generative RMs across multiple comprehensive reward model benchmarks, outperforming much larger open-weight models (e.g., Llama3.1-405B) and proprietary ones (e.g., GPT-4o) by up to 13.8%. Beyond final performance, we perform thorough empirical analysis to understand the key ingredients of successful ReasRM training. To facilitate future research, we release six ReasRM models along with code and data at https://github.com/RM-R1-UIUC/RM-R1.

Aligning large vision-language models (LVLMs) with human preferences is challenging due to the scarcity of fine-grained, high-quality, and multimodal preference data without human annotations. Existing methods relying on direct distillation often struggle with low-confidence data, leading to suboptimal performance. To address this, we propose CAREVL, a novel method for preference reward modeling by reliably using both high- and low-confidence data. First, a cluster of auxiliary expert models (textual reward models) innovatively leverages image captions as weak supervision signals to filter high-confidence data. The high-confidence data are then used to fine-tune the LVLM. Second, low-confidence data are used to generate diverse preference samples using the fine-tuned LVLM. These samples are then scored and selected to construct reliable chosen-rejected pairs for further training. CAREVL achieves performance improvements over traditional distillation-based methods on VL-RewardBench and MLLM-as-a-Judge benchmark, demonstrating its effectiveness. The code will be released soon.

Black-box distillation creates student large language models (LLMs) by learning from a proprietary teacher model's text outputs alone, without access to its internal logits or parameters. In this work, we introduce Generative Adversarial Distillation (GAD), which enables on-policy and black-box distillation. GAD frames the student LLM as a generator and trains a discriminator to distinguish its responses from the teacher LLM's, creating a minimax game. The discriminator acts as an on-policy reward model that co-evolves with the student, providing stable, adaptive feedback. Experimental results show that GAD consistently surpasses the commonly used sequence-level knowledge distillation. In particular, Qwen2.5-14B-Instruct (student) trained with GAD becomes comparable to its teacher, GPT-5-Chat, on the LMSYS-Chat automatic evaluation. The results establish GAD as a promising and effective paradigm for black-box LLM distillation.

While online Reinforcement Learning has emerged as a crucial technique for aligning flow matching models with human preferences, current approaches are hindered by inefficient exploration during training rollouts. Relying on undirected stochasticity and sparse outcome rewards, these methods struggle to discover high-reward samples, resulting in data-inefficient and slow optimization. To address these limitations, we propose Euphonium, a novel framework that steers generation via process reward gradient guided dynamics. Our key insight is to formulate the sampling process as a theoretically principled Stochastic Differential Equation that explicitly incorporates the gradient of a Process Reward Model into the flow drift. This design enables dense, step-by-step steering toward high-reward regions, advancing beyond the unguided exploration in prior works, and theoretically encompasses existing sampling methods (e.g., Flow-GRPO, DanceGRPO) as special cases. We further derive a distillation objective that internalizes the guidance signal into the flow network, eliminating inference-time dependency on the reward model. We instantiate this framework with a Dual-Reward Group Relative Policy Optimization algorithm, combining latent process rewards for efficient credit assignment with pixel-level outcome rewards for final visual fidelity. Experiments on text-to-video generation show that Euphonium achieves better alignment compared to existing methods while accelerating training convergence by 1.66x.

Due to constraints on privacy, cost, and latency, on-premise deployment of small models is increasingly common. However, most practical pipelines stop at supervised fine-tuning (SFT) and fail to reach the reinforcement learning (RL) alignment stage. The main reason is that RL alignment typically requires either expensive human preference annotation or heavy reliance on high-quality reward models with large-scale API usage and ongoing engineering maintenance, both of which are ill-suited to on-premise settings. To bridge this gap, we propose a positive-unlabeled (PU) RL distillation method for on-premise small-model deployment. Without human-labeled preferences or a reward model, our method distills the teacher's preference-optimization capability from black-box generations into a locally trainable student. For each prompt, we query the teacher once to obtain an anchor response, locally sample multiple student candidates, and perform anchor-conditioned self-ranking to induce pairwise or listwise preferences, enabling a fully local training loop via direct preference optimization or group relative policy optimization. Theoretical analysis justifies that the induced preference signal by our method is order-consistent and concentrates on near-optimal candidates, supporting its stability for preference optimization. Experiments demonstrate that our method achieves consistently strong performance under a low-cost setting.

Reinforcement learning can train LLM agents from sparse task rewards, but long-horizon credit assignment remains challenging: a single success-or-failure signal must be distributed across many actions. Existing methods rely on trajectory-level rewards or proxy signals, without fully leveraging per-step environmental feedback. Multi-turn agent settings are underexplored, where feedback can include error messages, page changes, observations, or reference trajectories. We systematically study five feedback sources and two insertion granularities and introduce SERL, a selective environment-reweighted learning framework. SERL uses the task reward to determine update direction, while environment feedback adjusts placement and magnitude, focusing on critical actions. On ALFWorld and WebShop, SERL achieves 90.0% and 80.1% success, outperforming strong RL and distillation baselines. Analysis shows that grounded, action-relevant feedback at meaningful points consistently outperforms indiscriminate use of longer or richer context.

Aligning diffusion models with human preferences usually relies on a sparse terminal reward evaluated on the final generated samples, which creates a severe temporal credit-assignment problem across the denoising process. We propose Latent Reward Registers, a mechanism that estimates terminal preference directly from intermediate noisy latents. Learnable, position-free register tokens are appended as an auxiliary read path to a frozen Diffusion Transformer (DiT), extracting preference signals without altering the generator's hidden states or velocity field. The resulting dense, differentiable reward field spans the full denoising trajectory and supports two alignment strategies. For training, Reward-Gradient On-Policy Distillation (RG-OPD) converts this dense reward field into per-step targets at states visited by the current generator, replacing rollout-intensive policy gradients with direct on-policy distillation. For inference, Reward-Guided Sampling (RGS) steers trajectories with magnitude-matched reward-gradient corrections and no parameter updates. Empirically, at high noise levels (t=0.8) the registers reach the highest pairwise accuracy among the evaluated latent reward models. RG-OPD outperforms online reinforcement learning baselines while reducing GPU hours by up to 33x. RGS achieves significant reward improvement with a favorable reward-quality balance against training-free baselines. Code and weights are to be available at https://github.com/Guanys-dar/latent-reward-register

On-policy distillation (OPD) has become an important approach to language model post-training. However, despite its performance gains, OPD can also collapse into excessively long and repetitive generation, and the mechanism underlying these divergent outcomes remains poorly understood. We explain these outcomes through a reinforcement learning perspective: the teacher implicitly rewards student behaviors, even those it rarely exhibits itself. From this perspective, our experiments show that OPD improves performance without expanding the student's capabilities. When the implicit reward model is reliable, OPD makes correct responses easier to sample. In contrast, when the preference misaligns with quality, reward hacking happens: the implicit reward model amplifies overlong, repetitive student rollouts, even though it rarely generates such text itself. Guided by this diagnosis, we find that masking unhealthy responses during training and using SFT initialization can each effectively mitigate the collapse. Together, these findings show that OPD amplifies student behaviors favored by the teacher's implicit feedback, shifting the focus from how well the teacher generates to how reliably it evaluates student rollouts. Our code is available at https://github.com/HancCui/opd_hacking.

Outcome-based reinforcement learning provides coarse supervision for long-horizon LLM agents: failed trajectories may contain correct steps, while successful trajectories can contain errors. Trajectory-level skills can also shift teacher preferences at unrelated steps, introducing off-target distillation signals. We introduce SCOPE, a framework for Selective Critical-Step On-Policy Self-Distillation. After a supervised warm-up, the policy analyzes its own trajectories, generates step-specific hindsight skills, and distills their guidance only at selected error steps. A bounded signed teacher-student coefficient stabilizes auxiliary updates while preserving positive reinforcement and negative correction. A success-relative pairwise length reward further encourages efficient task completion. Experiments on ALFWorld and WebShop demonstrate consistent improvements across multiple model backbones. With Qwen2.5-3B, SCOPE achieves 97.7% success on ALFWorld and 85.2% on WebShop, outperforming SEED by 5.9 and 6.3 percentage points, respectively. On unseen ALFWorld tasks, SCOPE reaches 89.6% success, improving over GRPO by 23.2 percentage points with 46.7% fewer interactions. Scaling to Qwen3-32B, SCOPE achieves 99.2% success on seen ALFWorld tasks and 94.0% on unseen tasks.

Distilling large language models (LLMs) typically involves transferring the teacher model's responses through supervised fine-tuning (SFT). However, this approach neglects the potential to distill both data (output content) and reward signals (quality evaluations). Extracting reliable reward signals directly from teacher models is challenging, as LLMs are optimized for generation rather than evaluation, often resulting in biased or inconsistent assessments. To address this limitation, we propose a novel distillation pipeline that transfers both responses and rewards. Our method generates pseudo-rewards through a self-supervised mechanism that leverages the inherent structure of both teacher and student responses, enabling reward learning without explicit external evaluation. The reward model subsequently guides reinforcement learning (RL), allowing iterative refinement of the student model after an SFT warm-up phase. Experiments on GSM8K and MMLU-PRO demonstrate that our method consistently outperforms traditional SFT-based approaches, enabling student models to surpass the performance of their teachers. This work highlights the potential for scalable, efficient distillation through structured self-supervised reward learning, reducing dependence on external reward supervision.

Large language models are increasingly post-trained with reinforcement learning in verifiable domains such as code and math. Yet, current methods for reinforcement learning with verifiable rewards (RLVR) learn only from a scalar outcome reward per attempt, creating a severe credit-assignment bottleneck. Many verifiable environments actually provide rich textual feedback, such as runtime errors or judge evaluations, that explain why an attempt failed. We formalize this setting as reinforcement learning with rich feedback and introduce Self-Distillation Policy Optimization (SDPO), which converts tokenized feedback into a dense learning signal without any external teacher or explicit reward model. SDPO treats the current model conditioned on feedback as a self-teacher and distills its feedback-informed next-token predictions back into the policy. In this way, SDPO leverages the model's ability to retrospectively identify its own mistakes in-context. Across scientific reasoning, tool use, and competitive programming on LiveCodeBench v6, SDPO improves sample efficiency and final accuracy over strong RLVR baselines. Notably, SDPO also outperforms baselines in standard RLVR environments that only return scalar feedback by using successful rollouts as implicit feedback for failed attempts. Finally, applying SDPO to individual questions at test time accelerates discovery on difficult binary-reward tasks, achieving the same discovery probability as best-of-k sampling or multi-turn conversations with 3x fewer attempts.

Reinforcement learning with verifiable rewards (RLVR) has become a standard paradigm for post-training large language models. While Group Relative Policy Optimization (GRPO) is widely adopted, its coarse credit assignment uniformly penalizes failed rollouts, lacking the token-level focus needed to efficiently address specific deviations. Self-Distillation Policy Optimization (SDPO) addresses this by providing denser, more targeted logit-level supervision that facilitates rapid early improvement, yet it frequently collapses during prolonged training. We trace this late-stage instability to two intrinsic flaws: self-distillation on already-correct samples introduces optimization ambiguity, and the self-teacher's signal reliability progressively degrades. To resolve these issues, we propose Sample-Routed Policy Optimization (SRPO), a unified on-policy framework that routes correct samples to GRPO's reward-aligned reinforcement and failed samples to SDPO's targeted logit-level correction. SRPO further incorporates an entropy-aware dynamic weighting mechanism to suppress high-entropy, unreliable distillation targets while emphasizing confident ones. Evaluated across five benchmarks and two model scales, SRPO achieves both the rapid early improvement of SDPO and the long-horizon stability of GRPO. It consistently surpasses the peak performance of both baselines, raising the five-benchmark average on Qwen3-8B by 3.4% over GRPO and 6.3% over SDPO, while simultaneously yielding moderate response lengths and lowering per-step compute cost by up to 17.2%.

Reinforcement learning from verifiable rewards (RLVR) suffers from sparse outcome signals, creating severe exploration bottlenecks on complex reasoning tasks. Recent on-policy self-distillation methods attempt to address this by utilizing language feedback to generate dense, token-level supervision. However, these approaches rely on a fixed, passive teacher to interpret the feedback. As the student policy improves, the teacher's zero-shot assessment capabilities plateau, ultimately halting further learning. To overcome this, we propose Variational Policy Distillation (VPD), a framework that formalizes learning from language feedback as a Variational Expectation-Maximization (EM) problem. VPD co-evolves both policies: in the E-step, the teacher is actively refined on trajectory outcomes via an adaptive trust-region update, translating textual feedback into a dynamically improved target token distribution. In the M-step, the student internalizes this dense distributional guidance on its own on-policy rollouts. By continuously improving the teacher's ability to extract actionable signals from textual critique, VPD overcomes the limitations of passive distillation. Evaluated across diverse sources of diagnostic feedback on scientific reasoning and code generation tasks, VPD consistently outperforms both standard RLVR and existing self-distillation baselines. Finally, by stress-testing our framework on rigid mathematical reasoning and cold-start regimes, we illuminate the fundamental bounds of feedback-driven self-distillation compared to pure environment-driven RL.

On-policy distillation (OPD) leverages dense teacher rewards to enhance reasoning models. However, scaling OPD to long-horizon tasks exposes a critical flaw: as the student's generated prefix inevitably diverges from the teacher's thought process, the teacher's dense reward loses local exploitability. Continuing to generate and evaluate tokens on these ``drifted'' trajectories not only degrades reward quality but also incurs massive computational waste. To address this, we introduce Prune-OPD, a framework that dynamically aligns training budgets with supervision quality. By continuously monitoring the local compatibility between student and teacher predictions (e.g., via top-k overlap), Prune-OPD detects prefix-drift events in real time. Upon detecting severe drift, it monotonically down-weights subsequent unreliable rewards and triggers dynamic rollout truncation. This allows the training process to halt futile generation and reallocate compute strictly to reliable teacher supervision. Across diverse teacher-student combinations, Prune-OPD consistently aligns computation with supervision reliability. When prefix drift makes dense teacher rewards unreliable, it reduces training time by 37.6\%--68.0\% while preserving, and often improving, performance on challenging benchmarks (AMC, AIME, HMMT). When student-teacher compatibility remains high, it automatically preserves long-context supervision by expanding the training window. These results suggest that Prune-OPD improves OPD not by blindly shortening rollouts, but by reallocating computation toward locally exploitable teacher rewards.

Reward models are central to text-to-image post-training, but visual preference is subjective and better represented as a distribution over rubric scores than as a deterministic scalar. Existing scalar, score-token, and pairwise reward models over-compress uncertainty and fine-grained score differences, while reasoning-based generative rewards provide stronger judgments but are costly to deploy and difficult to use as direct optimization signals. We propose Z-Reward, a teacher-student reward modeling framework that decouples reasoning-heavy judgment from efficient reward deployment. The teacher is a large VLM that uses reasoning to infer rubric-aligned score distributions, and is trained with Group-wise Direct Score Optimization (GDSO), which combines policy-gradient rewards from distribution expectations with direct pointwise and pairwise supervision on score distributions and score gaps. The student is trained with Reasoning-Internalized Score Distillation (RISD), which transfers the teacher's reasoning-conditioned score distribution into a compact VLM without requiring explicit reasoning chains at inference time. On our internally annotated evaluation set, the 27B GDSO teacher reaches 89.6% human preference accuracy, outperforming SFT, RewardDance, and GRPO, while the 9B RISD student reaches 88.6%, outperforming the OPD baseline and closely matching the larger teacher. We further show that Z-Reward can serve as a differentiable reward signal for text-to-image optimization, yielding a 41.3% net human-preference improvement over the SFT baseline.

Reinforcement learning with verifiable rewards (RLVR) often relies on sparse outcome rewards, providing coarse supervision for long-horizon agents. On-policy self-distillation (OPSD) complements this signal with dense privileged feedback. However, we identify Decision--Timestamp Mismatch: privileged guidance may be misaligned with the student's functional decision because the corresponding decision can occur at a different timestep, while the student's decision itself may span multiple timesteps rather than being tied to a single timestamp. Thus, timestamp-local supervision can misalign both the context and the temporal scope of credit. To address this mismatch, we introduce AlignOPSD, following the principle of aligning supervision before assigning credit. Decision-Aligned Supervision Rectification re-scores the same student-sampled response in functionally matched contexts across sibling rollouts to calibrate local teacher evidence. Semi-Markov Hierarchical Credit Assignment then derives variable-duration decision spans from correspondence changes and uses rectified evidence to allocate outcome-grounded credit across spans and their constituent turns. We evaluate AlignOPSD with Qwen2.5-3B and Qwen2.5-7B on ALFWorld, WebShop, and Search-QA against representative baselines. AlignOPSD outperforms both GRPO and StepOPSD across all eight backbone--aggregate-metric comparisons, improving on GRPO by 5.5--8.7 \% and ranking first in six. Additional analyzes examine the two alignment stages and hyperparameter sensitivity between tasks. Our code is avaliable at https://github.com/mingju-c/Align-OPSD

Reinforcement learning with verifiable rewards (RLVR) via Group-Relative Policy Optimization (GRPO) is widely used for multi-turn VLM agent training, yet it suffers from zero-gradient silence on uniform failures and coarse episode-level credit assignment. While On-Policy Distillation (OPD) and On-Policy Self-Distillation (OPSD) mitigate sparse rewards using hindsight information, their underlying mechanisms remain poorly understood. Through controlled counterfactual rollback probes across five multi-turn VLM agent benchmarks, we reveal that performance gains in OPSD/OPD are largely driven by physical state rollback at the pivot step, defined as the first unrecoverable action without remaining step budget. However, physical state rollbacks are computationally prohibitive and infeasible in real-world environments. To bridge this gap, we present Pivot-Aware Internalized Visual On-Policy Training (PIVOT), an RL framework that internalizes pivot localization and state restoration directly into token-level parameter updates, eliminating environment rollbacks during RL training and additional skill hints at test time. PIVOT unifies three functional roles within a single architecture: a failure Analyzer non-invasively localizes the pivot step and diagnoses failure modes from visual trajectory collages and action logs; a detached Teacher re-scores failed tokens under this privileged diagnostic context; and a Student optimizes joint GRPO and confidence-gated OPD objectives. At test time, both Teacher and Analyzer branches are stripped. Evaluated on five multi-turn VLM agent tasks across cognitive grid puzzles, 3D embodied control and navigation, and generative reasoning, PIVOT achieves 0.90 overall accuracy on Qwen2.5-VL-3B (+8% over SFT+GRPO baseline and +5% over previous SOTA) and scales to 0.92 on Qwen3-VL-2B (+12% over SFT+GRPO baseline).

In the field of large language models (LLMs), Knowledge Distillation (KD) is a critical technique for transferring capabilities from teacher models to student models. However, existing KD methods face limitations and challenges in distillation of LLMs, including efficiency and insufficient measurement capabilities of traditional KL divergence. It is shown that LLMs can serve as an implicit reward function, which we define as a supplement to KL divergence. In this work, we propose Direct Preference Knowledge Distillation (DPKD) for LLMs. DPKD utilizes distribution divergence to represent the preference loss and implicit reward function. We re-formulate KD of LLMs into two stages: first optimizing and objective consisting of implicit reward and reverse KL divergence and then improving the preference probability of teacher outputs over student outputs. We conducted experiments and analysis on various datasets with LLM parameters ranging from 120M to 13B and demonstrate the broad applicability and effectiveness of our DPKD approach. Meanwhile, we prove the value and effectiveness of the introduced implicit reward and output preference in KD through experiments and theoretical analysis. The DPKD method outperforms the baseline method in both output response precision and exact match percentage. Code and data are available at https://aka.ms/dpkd.

Reasoning abilities, especially those for solving complex math problems, are crucial components of general intelligence. Recent advances by proprietary companies, such as o-series models of OpenAI, have made remarkable progress on reasoning tasks. However, the complete technical details remain unrevealed, and the techniques that are believed certainly to be adopted are only reinforcement learning (RL) and the long chain of thoughts. This paper proposes a new RL framework, termed OREAL, to pursue the performance limit that can be achieved through Outcome REwArd-based reinforcement Learning for mathematical reasoning tasks, where only binary outcome rewards are easily accessible. We theoretically prove that behavior cloning on positive trajectories from best-of-N (BoN) sampling is sufficient to learn the KL-regularized optimal policy in binary feedback environments. This formulation further implies that the rewards of negative samples should be reshaped to ensure the gradient consistency between positive and negative samples. To alleviate the long-existing difficulties brought by sparse rewards in RL, which are even exacerbated by the partial correctness of the long chain of thought for reasoning tasks, we further apply a token-level reward model to sample important tokens in reasoning trajectories for learning. With OREAL, for the first time, a 7B model can obtain 94.0 pass@1 accuracy on MATH-500 through RL, being on par with 32B models. OREAL-32B also surpasses previous 32B models trained by distillation with 95.0 pass@1 accuracy on MATH-500. Our investigation also indicates the importance of initial policy models and training queries for RL. Code, models, and data will be released to benefit future researchhttps://github.com/InternLM/OREAL.

While long Chain-of-Thought (CoT) distillation effectively transfers reasoning capability to smaller language models, the reasoning process often remains redundant and computational budget uncontrollable, leading to inefficient resource usage. To address this limitation, we propose Budget-Aware Reasoning Distillation (BARD), a novel framework that simultaneously distills reasoning capability and enables fine-grained control over the reasoning length. BARD uses the thinking budget as a user-specified control signal, allowing the model to dynamically balance reasoning performance and computational efficiency. To achieve this concept, BARD introduces a two-phase training regimen. The first phase, Supervised Fine-Tuning (SFT) on teacher-generated long CoT data compressed to various budget levels, bootstrapping the model's understanding of budget constraints. The second phase leverages Reinforcement Learning (RL) from a reward signal in consideration of reasoning performance and budget fidelity simultaneously. Incorporating the two-phase regimen is crucial to avoiding policy degradation and ensuring that both objectives are optimized jointly. Extensive experiments demonstrate that our method empowers an 8B student model to achieve strong performance on challenging reasoning benchmarks (AIME24, AIME25, GPQA) while providing precise and adaptive control over its reasoning length across a wide range of budgets.

Web agents hold great potential for automating complex computer tasks, yet their interactions involve long-horizon, sequential decision-making with irreversible actions. In such settings, outcome-based supervision is sparse and delayed, often rewarding incorrect trajectories and failing to support inference-time scaling. This motivates the use of Process Reward Models (WebPRMs) for web navigation, but existing approaches remain limited: scalar WebPRMs collapse progress into coarse, weakly grounded signals, while checklist-based WebPRMs rely on brittle template matching that fails under layout or semantic changes and often mislabels superficially correct actions as successful, providing little insight or interpretability. To address these challenges, we introduce WebArbiter, a reasoning-first, principle-inducing WebPRM that formulates reward modeling as text generation, producing structured justifications that conclude with a preference verdict and identify the action most conducive to task completion under the current context. Training follows a two-stage pipeline: reasoning distillation equips the model with coherent principle-guided reasoning, and reinforcement learning corrects teacher biases by directly aligning verdicts with correctness, enabling stronger generalization. To support systematic evaluation, we release WebPRMBench, a comprehensive benchmark spanning four diverse web environments with rich tasks and high-quality preference annotations. On WebPRMBench, WebArbiter-7B outperforms the strongest baseline, GPT-5, by 9.1 points. In reward-guided trajectory search on WebArena-Lite, it surpasses the best prior WebPRM by up to 7.2 points, underscoring its robustness and practical value in real-world complex web tasks.

When a model produces a correct solution under reinforcement learning with verifiable rewards (RLVR), every token receives the same reward signal regardless of whether it was a decisive reasoning step or a grammatical filler. A natural fix is to condition the model on the correct answer as a teacher, identifying tokens it would have generated differently had it known the answer. Prior work shows this either corrupts training by leaking the answer into the gradient, or produces a weak signal that cannot distinguish decisive steps from filler, since both look equally surprising relative to the model's baseline. We propose Contrastive Evidence Policy Optimization (CEPO), which asks a sharper question at every token: not just "does the correct answer favor this token?" but "does the correct answer favor it while the wrong answer disfavors it?" A token satisfying both is a genuine reasoning step; one satisfying neither is filler. The wrong-answer teacher is constructed from rejected rollouts already in the training batch, incurring no additional sampling cost. We prove CEPO inherits all structural safety guarantees of the prior state of the art while strictly sharpening credit at decisive tokens, with the improvement vanishing exactly at filler positions. Empirically, CEPO achieves 43.43% and 60.56% average accuracy across five multimodal mathematical reasoning benchmarks at 2B and 4B scale, respectively, versus 41.17% and 57.43% for GRPO under identical training budgets. Distribution-matching self-distillation methods (OPSD, SDPO) fall below the untrained baseline, empirically confirming the information leakage our theory predicts. Our code is available at https://github.com/ahmedheakl/CEPO.

Deep search agents have emerged as a promising paradigm for addressing complex information-seeking tasks, but their training remains challenging due to sparse rewards, weak credit assignment, and limited labeled data. Self-play offers a scalable route to reduce data dependence, but conventional self-play optimizes students only through sparse outcome rewards, leading to low learning efficiency. In this work, we observe that self-play naturally produces a question construction path (QCP) during task generation, an intermediate artifact that captures the reverse solution process. This reveals a new source of privileged information: self-play can provide high-quality privileged information for the self-distillation at low cost and at scale, without relying on human feedback or curated privileged information. Leveraging this insight, we propose Privileged Information Self-Play (π-Play), a novel multi-agent self-evolution framework combining self-play and self-distillation. In π-Play, an examiner generates tasks together with QCPs, and a teacher employs QCP as privileged context to densely supervise a student via self-distillation. This design transforms sparse-reward self-play into a dense-feedback co-evolution. Extensive experiments show that data-free π-Play surpasses fully supervised search agents and improves evolutionary efficiency by 2-3times over conventional self-play. Code is available at https://github.com/zhyaoch/pi-play.

Conditioning a language model on additional context, such as feedback on a previous attempt, typically improves its response. Self-distillation trains the model to retain this improvement when the context is not present. The method works by matching the model's output distribution under two settings: a student that sees only the question, and a self-teacher that also sees the context. What the model learns therefore depends on what context the self-teacher receives, yet the design of this context remains largely unexplored. We study context design for self-distillation by training a solver on feedback from a frozen critic. We compare three conditions: (i) a binary reward (GRPO), (ii) the reference solution, and (iii) a step-by-step critique aligned to the solver's reasoning trace. Step-aligned critique yields the largest gains, outperforming GRPO by 16.11 points and reference-solution-conditioned self-distillation by 5.27 points (Avg@12). Per-token advantage analysis reveals why: step-aligned feedback targets only the tokens where reasoning fails, leaving correct behavior intact. Conditioning on the reference solution, by contrast, pressures the model to change its behavior at every token (even correct steps) because an alternative derivation inevitably differs in phrasing and approach. This suggests that structural alignment between feedback and the solver's reasoning is a key driver of self-distillation effectiveness.

Realistic and diverse traffic simulation is essential to autonomous driving development. Yet prevailing benchmarks predominantly reward realism, and recent methods have optimized accordingly, leaving diversity underexplored. We introduce Flow-ERD, a multi-agent simulator that pursues realism and diversity jointly. Its backbone, Agent-Type Aware Flow Matching (AFM), couples flow matching's multi-modal expressiveness with type-specific kinematic execution. It preserves fine-grained diversity while keeping motions consistent with each agent type. A second stage, Entropy-Regularized Distillation (ERD), fine-tunes the closed-loop rollout distribution with an entropy-regularized reverse-KL objective. This mitigates covariate shift while explicitly preventing collapse onto high-density modes. We evaluate Flow-ERD with a log-free diversity metric alongside standard realism scores. Flow-ERD ranks first on the WOSAC test benchmark and dominates the realism--diversity Pareto front among reproducible baselines. Our project page is available https://seulbinhwang.github.io/flow-erd-project-page/{here}.

Training small language-model agents for long-horizon interactive tasks requires both fast imitation and reward-driven improvement. On-policy distillation (OPD) provides dense teacher guidance and typically improves rapidly in the early stage, but its gains saturate once the student approaches the teacher, limiting the final performance ceiling. Reinforcement learning (RL) directly optimizes environment rewards and encourages exploratory improvement toward a higher reward-defined ceiling, but sparse and delayed feedback makes early-stage learning much less efficient than OPD. In this paper, we propose ATOD (Annealed Turn-aware On-policy Distillation), a hybrid online distillation algorithm that explicitly exploits this complementarity. (1) ATOD uses an annealed OPD-RL schedule: OPD dominates early training to approach teacher-level behavior, while RL is gradually strengthened to drive reward-based exploration. (2) ATOD introduces Turn-level Disagreement-Uncertainty Reweighting (T-DUR), which softly amplifies high-utility turns and improves dense supervision in long trajectories. Experiments on ALFWorld, WebShop, and Search-QA show that ATOD consistently outperforms competing post-training baselines: across the three student sizes, ATOD improves average success rate by 3.03 points over OPD and 23.62 points over GRPO, while surpassing the corresponding teacher models by 2.16 points.

Reinforcement learning with verifiable rewards (RLVR) broadcasts a single response-level reward to every token, while on-policy distillation (OPD) scores each token against a stronger teacher for a dense advantage but caps performance at teacher quality and discourages exploration beyond it. Their complementarity makes combining RLVR and OPD promising, but we find that fusing the two advantages with a fixed coefficient triggers entropy collapse from two miscalibrations: a magnitude mismatch, where token-level OPD advantages can spike far beyond the bounded RLVR advantage and erase its signal, and a temporal mismatch, where sustained full-strength OPD keeps pulling the student toward the teacher and limits exploration needed to surpass it. We propose SAF, a Stable Advantage Fusion framework that resolves both issues via a lightweight, four-stage pipeline applied only to the OPD advantage: a sparsify-then-compress mechanism for magnitude control paired with a warm-up-then-anneal mechanism for temporal control, with each stage independently switchable and adding negligible overhead. Instantiating RLVR with GRPO, we evaluate SAF across seven mathematical reasoning and code generation benchmarks with Qwen3-1.7B/4B/8B: SAF avoids entropy collapse and consistently outperforms fixed-coefficient GRPO+OPD fusion, improving the aggregate score by 0.51-2.70% across all six model-domain settings while achieving more stable training.

Reinforcement learning with verifiable rewards (RLVR) has emerged as a powerful paradigm for large language model (LLM) post-training, but its reliance on coarse outcome rewards leads to limited guidance on intermediate reasoning processes. Existing approaches such as process reward modeling and on-policy distillation introduce additional constraints, such as reliance on a specialized reward model or assuming identical reasoning patterns between teacher and student. Nevertheless, we observe that once a reasoning process first goes wrong, evaluating the subsequent reasoning provides limited additional information, as it is already conditioned on an invalid prefix. Therefore, we propose Cliff, a reward shaping strategy that utilizes an off-the-shelf LLM as a teacher to identify the first mistake in each rollout. As a result, the rollout is naturally decomposed into two parts: a correct prefix and an incorrect suffix. Cliff then converts this signal into token-level advantages, assigning positive advantages for the correct prefix and negative feedback afterward. Experiments across 12 different scenarios demonstrate that Cliff consistently improves reasoning performance, outperforming on-policy distillation by 15% and standard GRPO by 7%, even with teachers of modest capability. Furthermore, we analyse the role of ``ground truth'' in Cliff and investigate its training dynamics. These results establish Cliff as a simple, general and effective approach for improving RLVR with richer, fine-grained supervision.

Efficient text-to-image generation requires both reinforcement-learning (RL)-based reward alignment and few-step distillation, yet these procedures are typically performed sequentially, increasing training cost and risking the loss of reward gains during compression. We take an RL-native perspective: diffusion RL already generates reward-scored finite-step trajectories, whose intermediate states provide distillation supervision. Based on this insight, we propose REST (Reward-Enhanced Scored-Trajectory Distillation), a single-stage co-training framework in which a decoupled student learns from the evolving RL teacher's trajectories without changing teacher optimization. Advantage-Modulated Distillation (AMD) transforms rollout advantages into signed weights, strengthening imitation of preferred trajectories and aligning distillation priorities with task value. The resulting framework is general and lightweight, requires no extra image rollouts, no separate distillation dataset, and no adversarial training. Experiments on compositional generation, visual text rendering, and human-preference alignment demonstrate competitive few-step, CFG-free generation with RAM or DiffusionNFT teachers. With only four sampling steps, REST-RAM achieves a DrawBench PickScore of 23.97, outperforming both the 40-step RAM teacher (23.95) and RTDMD (23.71).

Reward-specialized post-training produces strong experts for flow-based generative models, while multi-teacher on-policy distillation (OPD) consolidates their capabilities into a single student. Existing methods, however, route each prompt to a single teacher according to its semantic category, implicitly binding the desired capability to prompt content. This coupling makes capability invocation vulnerable to prompt perturbations and prevents users from explicitly adjusting the strength of the desired capability at inference time. In this work, we introduce CapField-OPD, an OPD framework that integrates multiple teachers into a continuous capability field through explicit capability coordinates. We use teacher models as anchors to construct this field, with the coordinates determining how their outputs are combined. Each capability configuration thus receives a unique supervision target, and capability control no longer depends on prompt semantics. Since the training anchors may not be optimal at inference time, we further profile the learned field on a small calibration set. The coordinate with the highest mean reward serves as the recommended default, while coordinates that are frequently optimal offer a promising candidate set for test-time scaling. Extensive experiments on compositional generation, text rendering, and visual aesthetics demonstrate that CapField-OPD consolidates multiple specialized teachers into a single student while preserving or surpassing their performance, reliably invokes the desired capabilities under semantics-preserving prompt variations, and supports continuous capability control and coordinate-based test-time scaling.

Looped Language Models (LoopLMs) offer a parameter efficient approach to scaling reasoning by reusing shared parameters across recurrent computation steps. Despite their promise, effective post-training of LoopLMs remains challenging. Existing approaches either provide reward based supervision that is sparse or costly to extend across loops, or rely on external teachers or privileged information, leading to limited teacher availability or teacher-student context mismatch. To address these limitations, we introduce LoopOPD, a cross-loop on-policy distillation framework that uses additional recurrent computation within a LoopLM as its own source of supervision. LoopOPD uses a frozen terminal loop policy as a compute privileged teacher for an intermediate loop student on student generated rollouts, providing dense supervision without an external teacher or privileged information. We further propose Dynamic LoopOPD (D-LoopOPD), which continually refreshes the terminal loop teacher as the shared model parameters are updated, enabling recurrent self-improvement. We characterize how distillation updates propagate across loop depths and derive sufficient conditions under which a single update yields simultaneous local improvement at both loop depths. Experiments on Ouro-Thinking models show that LoopOPD improves mathematical reasoning, while D-LoopOPD yields further gains through dynamic teacher updates. Despite being trained only on mathematical data, the resulting models also improve on general reasoning and code generation benchmarks, demonstrating that recurrent computation can serve as an effective source of supervision for LoopLMs. Our code and model checkpoints will be released upon acceptance.

Self-Distillation Policy Optimization (SDPO) provides dense token-level credit assignment for reinforcement learning with large language models by leveraging the model's own feedback-conditioned predictions as a self-teacher. Unlike GRPO, however, whose group-relative advantage naturally concentrates learning on a sweet spot of intermediate-difficulty questions, SDPO's KL-based advantage lacks an implicit notion of difficulty awareness. We analyze this gap through the lens of GRPO's advantage normalization. Extending the learnability framework to normalized rewards, we show that normalization absorbs the variance term p(1-p), equalizing leading-order learnability across questions and leaving p(1-p) as the sole residual scaling factor in the per-question gradient. This analysis yields a simple prescription: weight each question's SDPO loss by [p(1-p)]^{1/2}, resulting in SC-SDPO, a scale-consistent variant of SDPO. The proposed weights are obtained as a zero-cost byproduct of on-policy rollouts with batch-adaptive normalization, inducing an implicit curriculum that dynamically tracks the model's evolving competence. Experiments on scientific reasoning and tool-use benchmarks demonstrate that SC-SDPO consistently improves over SDPO, yielding gains of +3.2/+4.3 (mean@16/maj@16) on Qwen3-8B and +1.8/+3.0 on OLMo-3-7B, while preserving stable training dynamics throughout optimization.

In modern large language models (LLMs), LLM alignment is of crucial importance and is typically achieved through methods such as reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO). However, in most existing methods for LLM alignment, all tokens in the response are optimized using a sparse, response-level reward or preference annotation. The ignorance of token-level rewards may erroneously punish high-quality tokens or encourage low-quality tokens, resulting in suboptimal performance and slow convergence speed. To address this issue, we propose AlignDistil, an RLHF-equivalent distillation method for token-level reward optimization. Specifically, we introduce the reward learned by DPO into the RLHF objective and theoretically prove the equivalence between this objective and a token-level distillation process, where the teacher distribution linearly combines the logits from the DPO model and a reference model. On this basis, we further bridge the accuracy gap between the reward from the DPO model and the pure reward model, by building a contrastive DPO reward with a normal and a reverse DPO model. Moreover, to avoid under- and over-optimization on different tokens, we design a token adaptive logit extrapolation mechanism to construct an appropriate teacher distribution for each token. Experimental results demonstrate the superiority of our AlignDistil over existing methods and showcase fast convergence due to its token-level distributional reward optimization.

We introduce an exploration bonus for deep reinforcement learning methods that is easy to implement and adds minimal overhead to the computation performed. The bonus is the error of a neural network predicting features of the observations given by a fixed randomly initialized neural network. We also introduce a method to flexibly combine intrinsic and extrinsic rewards. We find that the random network distillation (RND) bonus combined with this increased flexibility enables significant progress on several hard exploration Atari games. In particular we establish state of the art performance on Montezuma's Revenge, a game famously difficult for deep reinforcement learning methods. To the best of our knowledge, this is the first method that achieves better than average human performance on this game without using demonstrations or having access to the underlying state of the game, and occasionally completes the first level.

Training long-horizon LLM agents with reinforcement learning is challenging because sparse outcome rewards reveal whether a task succeeds, but not which intermediate actions caused the outcome or how they should be corrected. Recent methods alleviate this issue by generating rewards or textual hints from turn-level action-output signals, or by using feedback-conditioned self-distillation. However, generating feedback at every turn is inefficient when many intermediate turns are already successful or neutral, and applying feedback at a fixed or misaligned turn often fails to supervise the actions that contributed to the failure. To bridge this gap, we propose HINT-SD, a targeted self-distillation framework that uses full-trajectory hindsight to select failure-relevant actions and applies feedback-conditioned distillation only on targeted action spans. Experiments on BFCL v3 and AppWorld show that our method improves over the dense per-turn feedback baseline by up to 18.80 percent while achieving 2.26times lower time per training step, suggesting that selecting where to distill is a key factor for both effective and efficient long-horizon agent training.

Reinforcement learning with verifiable rewards (RLVR) is a powerful recipe for improving language-model reasoning, but it is expensive to repeat on every new strong model because the target model must generate many rollouts during training. As models scale, post-training itself becomes a bottleneck. We study a weak-to-strong alternative: run RL on a smaller model where rollouts are cheaper, then reuse what that RL run learned to improve a stronger target model. Directly distilling the post-RL weak teacher is not enough, because the teacher's final policy mixes useful RL gains with the limitations of the smaller model. We propose Direct On-Policy Distillation (Direct-OPD), which transfers the teacher's RL-induced policy shift instead. Direct-OPD compares the post-RL teacher with its own pre-RL reference and treats their log-ratio as a dense implicit reward for the student. In plain terms, the checkpoint pair tells us which actions RL made the weak model more or less likely to take, and Direct-OPD applies that signal on the stronger student's own on-policy states. This directly reuses the weak model's RL supervision signal without running sparse-reward RL on the target model. Empirically, Direct-OPD consistently leverages weaker teachers to improve stronger target models; notably, it boosts Qwen3-1.7B from 48.3% to 58.3% on AIME 2024 in just 4 hours on 8 A100 GPUs. It outperforms step-matched direct RL and enables the sequential composition of multiple policy shifts. Our results show that RL outcomes can be reused across model scales as implicit reward signals, not merely as final models to imitate.

Large language model training in open-ended domains lacks verifiable rewards, making task preferences difficult to formalize as effective supervision. Contexts can convey such preferences, yet provide little additional supervision once distilled into the student, motivating contexts that evolve with student performance. However, directly using evolving contexts as in-training supervision results in an unstable distillation target and conflicting distributions, requiring mechanisms to stabilize target and downweight conflicts. In this paper, we analyze the effect of contexts through a decomposition of the reverse KL objective, revealing two findings: the student is distilled toward the geometric mean of context-conditioned teachers, and the objective contains a conflict term that measures conflicts among these teachers. Based on this decomposition, we propose Flux-OPD, an OPD paradigm that uses evolving contexts as in-training supervision to capture task preferences in open-ended domains. Flux-OPD treats the differences between context-conditioned and context-free teachers as contextual difference signals, injects them as contextual corrections into the context-free teacher anchor, and weights their correction strength using the conflict term as an indicator. Experiments on open-ended tasks show that Flux-OPD outperforms existing OPD paradigms, highlighting the potential to combine teacher supervision with evolving contexts.

Reinforcement learning with verifiable rewards (RLVR) and on-policy distillation (OPD) have emerged as two dominant methods for post-training reasoning LLMs. Prior work uses OPD's dense token-level supervision to complement the sparse RL reward, fusing the two signals within a single step: either as a weighted-additive combination or a teacher-modulated rescaling of the RL advantage. In this paper, we show that a simple two-stage scheme, OPD-then-RL, consistently outperforms pure OPD, pure RLVR, and all such joint baselines across logic and math reasoning benchmarks. Beyond the empirical results, we further provide a systematic understanding of this through pass@k behavior, learning dynamics, and parameter updates, yielding a consistent explanation: OPD expands the student's coverage of teacher-supported solutions and RL sharpens within that support, while jointly optimizing the two signals causes them to interfere. To provide a practical recipe, we find that the OPD validation score is the key signal for when to switch to RL, and that OPD is a better cold start for RL than SFT. Together, our results establish OPD-then-RL as a simple yet strong way to combine the two methods, turning two entangled signals into complementary stages.

On-policy distillation (OPD) has emerged as a widely used paradigm for post-training large language models, reducing the train--test mismatch of conventional distillation by supervising the student on its own generated trajectories. However, existing OPD objectives remain largely token-local and outcome-agnostic, optimizing teacher--student agreement at each prefix despite reasoning quality being determined at the trajectory level. Reinforcement learning with verifiable rewards (RLVR), particularly Group Relative Policy Optimization (GRPO), provides complementary outcome-level supervision but suffers from sparse rewards and coarse credit assignment. We show that OPD and RLVR exhibit complementary blind spots: teacher signals provide dense local guidance but are weakly aligned with rollout correctness, whereas group-relative rewards capture task success but provide coarse token-level credit and vanish on all-failure groups. We introduce DiffGate, an outcome-gated objective that combines GRPO with selective, bounded teacher guidance. Teacher supervision is applied only to failed trajectories, scaled by group difficulty, and smoothly bounded to prevent extreme teacher--student discrepancies from dominating optimization. The verifier therefore determines which trajectories receive teacher guidance, while the teacher provides dense token-level update directions within those trajectories. Across Qwen3-0.6B and Qwen3-1.7B students, DiffGate improves code avg@8 over matched GRPO by +1.7 and +1.8 points and pass@8 by +1.6 and +5.7 points, respectively. On mathematics, avg@8 remains within 0.5 points of GRPO while pass@8 improves by +1.1 and +3.9 points. Overall, DiffGate improves pass@8 across all four model--domain settings, demonstrating improved solution coverage under our evaluation protocol.

Reinforcement learning with verifiable rewards (RLVR) improves the reasoning capabilities of large language models using automatically verifiable outcome signals, but these signals are typically sparse and at the sequence-level. On-policy self-distillation (OPSD) mitigates this sparsity by querying a privileged teacher at student-visited prefixes and providing dense token-level distributional supervision. Although this dense supervision alleviates signal sparsity, we find that standard OPSD still underexploits the temporal structure of the rollout. It assigns every local divergence the same coefficient, regardless of its position or the divergence sequence in which it occurs. In on-policy autoregressive generation, the same divergence magnitude can follow different discrepancy histories, reflecting different evolutions of the mismatch between the teacher and student. Since the local scalar alone cannot distinguish these temporal contexts, standard OPSD cannot adapt its token-level weights to the realized discrepancy sequence. To address this limitation, we propose Divergence-Adaptive Supervision Horizons (DASH). DASH maps the gap between each local distillation signal and the sequence-level mean to an adaptive propagation gate and then uses these gates to control backward multi-step aggregation. By doing so, DASH adjusts token-level supervision weights according to how local divergences evolve during generation. Experiments on three mathematical reasoning benchmarks across three model scales show that DASH improves over our matched vanilla OPSD reruns on every benchmark at all three scales. DASH reuses the teacher and student distributions that OPSD already computes, so the gains require no additional teacher or student forward pass. Code: https://github.com/DBtxy/DASH-OPSD

Since the advent of knowledge distillation, KL divergence has been the standard loss in distillation. Recently, on-policy distillation (OPD) has emerged as an efficient post-training paradigm for LLMs. As a distillation method, OPD naturally inherits KL divergence as its standard loss. However, in this work, we find that KL divergence may not be necessary for OPD. We show that simply preserving the update direction is sufficient for effective OPD. As long as the update direction is toward the teacher, OPD works. More precisely, it is not the direction of every token, but the direction of a small subset of tokens where the teacher and student disagree strongly. We first show that simply assigning a reward of (+1) to tokens where the teacher probability is higher than the student probability and (-1) where it is lower, which merely encourages updates toward the teacher, reproduces almost the same training mode as OPD with reverse KL. We further show that only the direction of a small subset of tokens with large teacher-student disagreement is critical, and training works as long as their update direction is toward the teacher, even if other tokens are pulled away from the teacher. And as an application of these findings, we introduce Consensus Multi-Teacher On-Policy Distillation (C-MOPD) to improve Multi-Teacher On-Policy Distillation (MOPD). Unlike MOPD, which routes each sample to a single teacher and may cause capability conflicts across domains, C-MOPD lets every sample be supervised by all teachers. Experiments show that C-MOPD consistently outperforms MOPD on both math and code benchmarks. Our code is available at https://github.com/LeapLabTHU/KL-Free-OPD.

Multimodal reinforcement learning with verifiable rewards (RLVR) provides reliable task-level supervision, yet offers limited guidance for localizing errors in visual reasoning. On-policy self-distillation (OPSD) complements this supervision with dense token-level corrections from a privileged teacher equipped with an evidence-centered visual view. However, the resulting corrections may conflate visual evidence with non-visual shifts induced by linguistic priors, response style, or the privileged view itself. We formalize this phenomenon as privileged-correction contamination: a token-level correction is contaminated when its update direction cannot be consistently attributed to the privileged visual evidence. Directly distilling contaminated corrections may therefore reinforce erroneous token-level updates. To address this issue, we propose reinforcement learning with visually attributable on-policy self-distillation, an evidence-consistent paradigm that augments task-level RLVR supervision with privileged token-level corrections grounded in visual evidence. We instantiate this paradigm with Visual-Attribution Gating (VAG), which uses the verifier-derived advantage as the optimization anchor and validates each privileged correction against the contrast between evidence-present and evidence-degraded teacher views. By admitting only corrections directionally supported by this visual contrast, VAG preserves visually grounded guidance while suppressing inconsistent residuals. Experiments across six visual reasoning benchmarks show that VAG consistently outperforms GRPO and OPSD methods, establishing new state-of-the-art average performance.

Reinforcement learning with verifiable rewards (RLVR) has become a standard paradigm for improving large language models (LLMs) on various tasks, yet its sparse outcome rewards lack token-level credit assignment for intermediate steps. To address this, on-policy self-distillation (OPSD) leverages a self-teacher with privileged context to provide additional dense learning signals. However, because the self-teacher is often overconfident and imposes excessive penalties on long reasoning trajectories, OPSD frequently struggles in practice. To mitigate this, we propose self-instructing policy optimization (SIPO) with a contrastive self-teacher to provide dense credit. At each iteration, SIPO samples multiple rollouts per prompt from the current policy, scores them with environment rewards, and constructs two teacher contexts for each rollout by pairing the reference answer with mistakes made within the group. The model then re-evaluates its own responses under both contexts, using the difference between the two teacher log-probabilities as token-level feedback, so that biases shared by both contexts are expected to largely cancel. The resulting objective yields a token-level advantage for every rollout: the reward still sets the main direction of each update while the self-teacher redistributes credit across tokens. Even in groups where every rollout fails and group-relative advantages vanish, SIPO still provides a learning signal. By preserving direct optimization of the task reward while providing dense, token-level feedback, this approach bridges reinforcement learning and on-policy self-distillation. Extensive experiments across multiple reasoning and code-generation benchmarks demonstrate that SIPO outperforms both RLVR and OPSD baselines without an external teacher or additional generation.

On-Policy Self-Distillation (OPSD), which aligns a student model's token-level distribution with that of its self-teacher prompted by privileged context, has emerged as a promising direction to mitigate reward sparsity in sequence-level post-training paradigms. However, OPSD often fails to achieve continuous performance gains during training. In this paper, we find that the model's limited self-teaching capability underlies this phenomenon by generating unreliable guidance on certain samples. To address this problem, we propose Teacher-Evolution Self-Distillation (TESD), which enables the self-teacher to generate rollouts and receive environmental feedback based on task correctness and context independence. First, using these rewards as training objectives, we perform teacher-oriented optimization to continuously enhance the model's self-teaching capability during distillation. Second, by estimating teacher guidance reliability from these rewards, we apply adaptive student-oriented optimization to down-weight samples with unreliable guidance. Extensive experiments across multiple datasets and model scales show that TESD consistently outperforms representative token-level and sequence-level methods, highlighting the necessity of improving supervision quality in OPSD.

Self-distillation policy optimization provides dense token-level supervision for LLM post-training through feedback-conditioned self-teachers. However, directly matching the self-teacher can disrupt successful behaviors on correct rollouts while providing insufficient correction on incorrect ones. We propose Asymmetric Policy Optimization via Self-Distillation (APSD), which constructs correctness-dependent targets by mixing student and self-teacher distributions: interpolation preserves successful behaviors on correct rollouts, while extrapolation beyond the self-teacher strengthens correction on incorrect ones. We analyze APSD through divergence decomposition, reverse-KL barycentric geometry, and gradient analysis, highlighting the importance of gradient flow through student-dependent targets. Experiments under reinforcement learning with verifiable rewards and rich feedback show that APSD achieves strong overall performance against reinforcement-learning and self-distillation baselines and matches SDPO performance with 2.6x fewer training steps on average across Qwen3 model scales.

With the flourishing development of intelligent warehousing systems, the technology of Automated Guided Vehicle (AGV) has experienced rapid growth. Within intelligent warehousing environments, AGV is required to safely and rapidly plan an optimal path in complex and dynamic environments. Most research has studied deep reinforcement learning to address this challenge. However, in the environments with sparse extrinsic rewards, these algorithms often converge slowly, learn inefficiently or fail to reach the target. Random Network Distillation (RND), as an exploration enhancement, can effectively improve the performance of proximal policy optimization, especially enhancing the additional intrinsic rewards of the AGV agent which is in sparse reward environments. Moreover, most of the current research continues to use 2D grid mazes as experimental environments. These environments have insufficient complexity and limited action sets. To solve this limitation, we present simulation environments of AGV path planning with continuous actions and positions for AGVs, so that it can be close to realistic physical scenarios. Based on our experiments and comprehensive analysis of the proposed method, the results demonstrate that our proposed method enables AGV to more rapidly complete path planning tasks with continuous actions in our environments. A video of part of our experiments can be found at https://youtu.be/lwrY9YesGmw.

Post-training of language models (LMs) increasingly relies on the following two stages: (i) knowledge distillation, where the LM is trained to imitate a larger teacher LM, and (ii) reinforcement learning from human feedback (RLHF), where the LM is aligned by optimizing a reward model. In the second RLHF stage, a well-known challenge is reward hacking, where the LM over-optimizes the reward model. Such phenomenon is in line with Goodhart's law and can lead to degraded performance on the true objective. In this paper, we investigate whether a similar phenomenon, that we call teacher hacking, can occur during knowledge distillation. This could arise because the teacher LM is itself an imperfect approximation of the true distribution. To study this, we propose a controlled experimental setup involving: (i) an oracle LM representing the ground-truth distribution, (ii) a teacher LM distilled from the oracle, and (iii) a student LM distilled from the teacher. Our experiments reveal the following insights. When using a fixed offline dataset for distillation, teacher hacking occurs; moreover, we can detect it by observing when the optimization process deviates from polynomial convergence laws. In contrast, employing online data generation techniques effectively mitigates teacher hacking. More precisely, we identify data diversity as the key factor in preventing hacking. Overall, our findings provide a deeper understanding of the benefits and limitations of distillation for building robust and efficient LMs.

Continual learning, enabling models to acquire new skills and knowledge without degrading existing capabilities, remains a fundamental challenge for foundation models. While on-policy reinforcement learning can reduce forgetting, it requires explicit reward functions that are often unavailable. Learning from expert demonstrations, the primary alternative, is dominated by supervised fine-tuning (SFT), which is inherently off-policy. We introduce Self-Distillation Fine-Tuning (SDFT), a simple method that enables on-policy learning directly from demonstrations. SDFT leverages in-context learning by using a demonstration-conditioned model as its own teacher, generating on-policy training signals that preserve prior capabilities while acquiring new skills. Across skill learning and knowledge acquisition tasks, SDFT consistently outperforms SFT, achieving higher new-task accuracy while substantially reducing catastrophic forgetting. In sequential learning experiments, SDFT enables a single model to accumulate multiple skills over time without performance regression, establishing on-policy distillation as a practical path to continual learning from demonstrations.

Recent advances in multimodal reward modeling have been largely driven by a paradigm shift from discriminative to generative approaches. Building on this progress, recent studies have further employed reinforcement learning from verifiable rewards (RLVR) to enhance multimodal reward models (MRMs). Despite their success, RLVR-based training typically relies on labeled multimodal preference data, which are costly and labor-intensive to obtain, making it difficult to scale MRM training. To overcome this limitation, we propose a Multi-Stage Reinforcement Learning (MSRL) approach, which can achieve scalable RL for MRMs with limited multimodal data. MSRL replaces the conventional RLVR-based training paradigm by first learning a generalizable reward reasoning capability from large-scale textual preference data, and then progressively transferring this capability to multimodal tasks through caption-based and fully multimodal reinforcement-learning stages. Furthermore, we introduce a cross-modal knowledge distillation approach to improve preference generalization within MSRL. Extensive experiments demonstrate that MSRL effectively scales the RLVR-based training of generative MRMs and substantially improves their performance across both visual understanding and visual generation tasks (e.g., from 66.6% to 75.9% on VL-RewardBench and from 70.2% to 75.7% on GenAI-Bench), without requiring additional multimodal preference annotations. Our code is available at: https://github.com/wangclnlp/MSRL.

Reinforcement learning has emerged as a powerful tool for improving diffusion-based text-to-image models, but existing methods are largely limited to single-task optimization. Extending RL to multiple tasks is challenging: joint optimization suffers from cross-task interference and imbalance, while cascade RL is cumbersome and prone to catastrophic forgetting. We propose DiffusionOPD, a new multi-task training paradigm for diffusion models based on Online Policy Distillation (OPD). DiffusionOPD first trains task-specific teachers independently, then distills their capabilities into a unified student along the student own rollout trajectories. This decouples single-task exploration from multi-task integration and avoids the optimization burden of solving all tasks jointly from scratch. Theoretically, we lift the OPD framework from discrete tokens to continuous-state Markov processes, deriving a closed-form per-step KL objective that unifies both stochastic SDE and deterministic ODE refinement via mean-matching. We formally and empirically demonstrate that this analytic gradient provides lower variance and better generality compared to conventional PPO-style policy gradients. Extensive experiments show that DiffusionOPD consistently surpasses both multi-reward RL and cascade RL baselines in training efficiency and final performance, while achieving state-of-the-art results on all evaluated benchmarks.

Search-augmented reasoning agents interleave internal reasoning with calls to an external retriever, and their performance relies on the quality of each issued query. However, under outcome-reward reinforcement learning, every search decision in a rollout shares the same trajectory-level reward, leaving individual queries without step-specific credit. Recent process-supervision approaches address this gap by drawing step-level signals from outside the policy, relying either on a much larger teacher model, or on sub-question annotations produced by a stronger external system. In contrast, we propose SD-Search, which derives step-level supervision from the policy itself through on-policy hindsight self-distillation, requiring neither an external teacher nor additional annotations. In SD-Search, a single model plays two roles that differ only in conditioning: a student that sees only the context available at inference time, and a teacher that additionally conditions on a compact hindsight block summarizing the search queries and final outcomes of a group of rollouts sampled from the same question. Since the teacher knows how each rollout unfolded and which ones succeeded, its query distribution implicitly marks which decisions were worth making, and the student is trained to recover this behavior by minimizing the token-level Jensen--Shannon divergence to the teacher at search-query positions. This layers a dense, step-level signal on top of GRPO's coarse trajectory reward. Crucially, this signal is produced by the policy itself within the standard RL training loop, without external model inference, auxiliary annotation pipeline, or additional training stage.

Robot learning requires adaptation methods that improve reliably from limited, mixed-quality interaction data. This is especially challenging in long-horizon, contact-rich tasks, where end-to-end policy finetuning remains inefficient and brittle. World models offer a compelling alternative: by predicting the outcomes of candidate action sequences, they enable online planning through counterfactual reasoning. However, training action-conditioned robotic world models directly in the real world requires diverse data at impractical scale. We introduce Simulation Distillation (SimDist), a framework that uses physics simulators as a scalable source of action-conditioned robot experience. During pretraining, SimDist distills structural priors from the simulator into a world model that enables planning from raw real-world observations. During real-world adaptation, SimDist transfers the encoder, reward model, and value function learned in simulation, and updates only the latent dynamics model using real-world prediction losses. This reduces adaptation to supervised system identification while preserving dense, long-horizon planning signals for online improvement. Across contact-rich manipulation and quadruped locomotion tasks, SimDist rapidly improves with experience, while prior adaptation methods struggle to make progress or degrade during online finetuning. Project website and code: https://sim-dist.github.io

Self-improvement for multimodal large language models (MLLMs) is typically driven by reward-based methods that provide only coarse scalar feedback. Distillation offers a richer alternative through dense token-level supervision, but in the visual domain it usually depends on privileged context constructed using external annotations and tools, or stronger models. We introduce CVPD (Contrastive Counterfactual Visual Process Distillation), which, to the best of our knowledge, is the first fully self-contained framework for dense, on-policy, token-level visual self-distillation for MLLMs. CVPD identifies visual blind spots where zooming into a region changes and sharpens the model's answer distribution, while removing the same region leaves the full-image behavior largely unchanged. Such regions reveal perceptual information that the model can encode but fails to consistently utilize under full-image conditioning. We propose a three-gate Counterfactual Criterion that identifies these regions directly from the model's own responses and converts them into dense contrastive supervision for self-distillation. On Qwen3-VL-8B-Instruct, CVPD outperforms six self-evolving baselines across twelve benchmarks, including methods that rely on external GPT-4o supervision, without a single regression. It achieves gains of +3.60 on OCRBench, +3.38 on MMStar Fine-Grained Perception, and +3.08 on MMStar Logical Reasoning, while maintaining or improving performance on broader multimodal benchmarks.

Multi-turn agents trained with reinforcement learning (RL) receive a single scalar reward per trajectory, which motivates self on-policy distillation (OPD) to supply dense token-level supervision from a self-teacher with privileged task skills, letting a skill-free student internalize them. This recipe, however, is undermined by two findings in agentic tasks: privileged information alone does not always make a teacher reliable, and the benefit of teacher supervision is stage-dependent. We therefore propose RetireOPD (Self-Retiring On-Policy Distillation), which first optimizes a decoupled, skill-conditioned teacher with environment rewards and then trains a skill-free student jointly with RL and OPD. Rather than following a predefined distillation schedule, RetireOPD adopts Adaptive Retirement: the student drops the teacher on its own once their discrepancy stops shrinking and it reaches a target fraction of the teacher's success rate, after which training proceeds with RL alone. Across Qwen2.5 models from 1.5B to 7B, RetireOPD improves ALFWorld success rate over RL baseline by 14.1% to 18.8% and WebShop accuracy by 11.8% to 19.0%, and surpasses its own skill-conditioned teacher in every setting.

Self-Distillation Fine-Tuning (SDFT) enables a language model to act as its own teacher: by conditioning on a demonstration, the model produces an implicit reward via pointwise mutual information, which guides on-policy learning without external supervision. However, SDFT operates at training time: it requires gradient updates and access to expert demonstrations, making it inapplicable at inference. We propose test-time self-distillation, a decoding-time method that extracts a steering signal from the self-distillation framework without any parameter updates, reward models, or training data. Our key insight is that counterfactual contexts, i.e. fixed textual templates that hypothetically prime the model for excellent versus poor reasoning, can substitute for the demonstration. The log-odds ratio of a candidate answer under these two counterfactual conditions defines a new reward signal. We derive the optimal KL-regularized policy under this reward, which takes the form of a Gibbs reweighting of the base distribution. Crucially, this reweighting is global: it cannot be decomposed into independent per-token operations without ignoring future trajectory quality. We therefore approximate the target distribution via beam search. Experiments on mathematical reasoning (MATH500), code generation (HumanEval), and graduate-level science QA (GPQA) across multiple model scales show that test-time self-distillation improves over standard sampling, low temperature, beam search and power sampling baselines on average, demonstrating that the self-distillation principle can be operationalized at inference time.

Many useful language-model tasks cannot be evaluated by exact outcome verification. Rubric-based reinforcement learning (RL) addresses this issue by scoring open-ended responses against explicit criteria. However, because the reward is assigned after the complete response, the training signal does not directly identify which individual decisions contributed to the final score. We propose a two-stage training framework that uses rubrics first as privileged teacher context for dense token-level supervision, then as rewards for further RL. In the first stage, rubric-privileged on-policy distillation (RP-OPD), a student without access to the rubric matches a rubric-aware teacher's next-token distributions at student-generated prefixes. In the second stage, RL directly optimizes the rubric reward and improves beyond the observed distillation plateau. We evaluate the framework on health and science tasks using open-weight models. Across HealthBench, ResearchQA, and RubricHub Science, we compare post-training methods and vary the amount of SFT or RP-OPD training before RL, finding that our two-stage framework achieves the highest scores among the methods evaluated. RP-OPD + RL shows limited signs of reward hacking on RubricHub Science, whereas the SFT + RL baseline increasingly receives high rewards for claims of rubric compliance without providing the required content. These findings support using rubrics to guide on-policy distillation before applying rubric-based RL.

Reward hacking is a form of misalignment in which models overoptimize proxy rewards without genuinely solving the underlying task. Precisely measuring reward hacking occurrence remains challenging because true task rewards are often expensive or impossible to compute. We introduce Countdown-Code, a minimal environment where models can both solve a mathematical reasoning task and manipulate the test harness. This dual-access design creates a clean separation between proxy rewards (test pass/fail) and true rewards (mathematical correctness), enabling accurate measurement of reward-hacking rates. Using this environment, we study reward hacking in open-weight LLMs and find that such behaviors can be unintentionally learned during supervised fine-tuning (SFT) when even a small fraction of reward-hacking trajectories leak into training data. As little as 1\% contamination in distillation SFT data is sufficient for models to internalize reward hacking which resurfaces during subsequent reinforcement learning (RL). We further show that RL amplifies misalignment and drives its generalization beyond the original domain. We open-source our environment and code to facilitate future research on reward hacking in LLMs. Our results reveal a previously underexplored pathway through which reward hacking can emerge and persist in LLMs, underscoring the need for more rigorous validation of synthetic SFT data. Code is available at https://github.com/zohaib-khan5040/Countdown-Code.

Language-model agents are usually trained by reinforcement learning from one reward per episode, and privileged self-distillation enriches it by letting the same policy, given a skill, teach its skill-free self through token probabilities. However, we identify two phenomena that question this channel. Invisible Advantage: a skill in context lifts WebShop success from 42.2% to 56.2%, yet changes the probabilities of fewer than a quarter of the sampled tokens. Much to Align: a skill changes the hidden states of over 80% of response tokens, in a way that linear probes can trace back to the specific skill. To exploit this, we propose Privileged Representation On-policy Self-Distillation (PR-OPD). After a GRPO warm start, the policy writes a hindsight skill for each trajectory, re-reads its own responses with that skill as a stop-gradient teacher, and aligns its projected hidden states to the teacher's at every layer alongside the reward objective, with no external skill library, separate teacher, or inference overhead. On ALFWorld and WebShop with two backbones, PR-OPD achieves the best overall results in every setting, improving over GRPO by up to 4.7 points in ALFWorld success and 14.0 points in WebShop accuracy. Code is available at https://github.com/balibata/PR-OPD.

Diffusion distillation, exemplified by Distribution Matching Distillation (DMD), has shown great promise in few-step generation but often sacrifices quality for sampling speed. While integrating Reinforcement Learning (RL) into distillation offers potential, a naive fusion of these two objectives relies on suboptimal raw sample evaluation. This sample-based scoring creates inherent conflicts with the distillation trajectory and produces unreliable rewards due to the noisy nature of early-stage generation. To overcome these limitations, we propose GDMD, a novel framework that redefines the reward mechanism by prioritizing distillation gradients over raw pixel outputs as the primary signal for optimization. By reinterpreting the DMD gradients as implicit target tensors, our framework enables existing reward models to directly evaluate the quality of distillation updates. This gradient-level guidance functions as an adaptive weighting that synchronizes the RL policy with the distillation objective, effectively neutralizing optimization divergence. Empirical results show that GDMD sets a new SOTA for few-step generation. Specifically, our 4-step models outperform the quality of their multi-step teacher and substantially exceed previous DMDR results in GenEval and human-preference metrics, exhibiting strong scalability potential.

Fine-grained visual reasoning requires multimodal large language models (MLLMs) to identify task-relevant visual evidence and ground their reasoning in local image regions. Existing agentic methods typically rely on reinforcement learning with verifiable rewards or supervised fine-tuning on large-scale annotated reasoning traces, leading to costly exploration, hand-designed verification rules, or heavy dependence on textual supervision. A natural way to avoid such external answer labels is to learn from trajectories sampled by the student itself, which points to On-Policy Distillation (OPD). To understand what OPD can and cannot provide for visual reasoning, we revisit it as negative-free stop-gradient alignment. This perspective shows that, although OPD provides effective token-level correction, its ceiling is constrained by the absence of trajectory-level discrimination. Motivated by these observations, we propose V-Zero, an answer-label-free framework for visual reasoning with contrastive evidence gating. V-Zero uses no annotated textual answer labels; instead, during training it pairs a question-relevant regional crop with a negative visual view to evaluate student-sampled trajectories and gate dense token-level distillation. Experiments on multiple visual reasoning benchmarks show that V-Zero consistently improves fine-grained visual reasoning while preserving strong generalization. Notably, V-Zero is more than 5times faster than previous supervised fine-tuning methods and more than 10times faster than reinforcement learning baselines. Code and dataset will be released at https://github.com/eVI-group-SCU/V-Zero

Reinforcement learning for multi-turn agents suffers from a credit-assignment mismatch: rewards are sparse and trajectory-level, while success often hinges on a few local decisions. Existing online policy distillation (OPD) provides denser token-level supervision, but typically treats heterogeneous agent trajectories as monolithic strings rather than causal interaction units. We present StepOPSD, a post-rollout preference self-distillation framework that takes the agent step as the unit of credit redistribution. StepOPSD decomposes trajectories into action-centered step segments, rescoring them under hindsight-enriched teacher contexts and converting token-level log-probability gaps into sign-preserving advantage shaping with a normalized per-step credit budget before the GRPO update. Across ALFWorld and Search-QA with Qwen3-1.7B and Qwen2.5-3B-Instruct, StepOPSD attains best or second-best results on subsets most sensitive to local causal errors, including first-place performance on ALFWorld Heat (79.1%), PickTwo (95.0%), Search-QA TriviaQA (61.6%), and tied-best performance on HotpotQA (40.4%). The results further reveal a consistent two-knob law: smaller α_clip acts as a broadly stabilizing local trust region, whereas the optimal global mixing strength λ_mix remains task-dependent. These findings suggest that step-aware distillation is most useful when trajectory-level rewards are weakly aligned with the local action that determines downstream success.

On-policy distillation (OPD) provides dense, per-token supervision for language model post-training, but its effectiveness is bottlenecked by teacher quality: external teachers suffer from distribution mismatch, while self-distillation with privileged conditioning is limited by in-context learning capacity. We propose RISE (Recursive Improvement via Self-Extrapolating Policy Distillation), which constructs a synthetic teacher directly from the model's own RLVR training trajectory. By extrapolating the displacement between the current checkpoint and a trailing anchor---in parameter space or output logit space---RISE converts a sparse outcome-induced parameter update into a dense token-level target, without any external model or privileged conditioning. RISE combines RLVR and OPD in a complementary loop: outcome rewards ground the extrapolation toward correct reasoning, while the extrapolated teacher refines token-level decisions. Moreover, since the teacher is refreshed every iteration as the student improves, distillation becomes a recursive improvement mechanism rather than a one-shot compression step. Experiments spanning mathematical reasoning, multi-domain STEM, code generation, and multi-turn agentic tasks show that RISE outperforms RLVR-only training and on-policy self-distillation across all settings.

A small trainable advisor can steer a frozen language-model executor using natural-language advice. In addition to learning from task rewards, the advisor can use feedback from completed interactions to improve its advice. However, a plausible correction need not change execution, yet learning from such corrections can still affect the advisor's future decisions in other contexts. In a shared-parameter model, we prove that such corrections can limit learning if their targets favor useful advice less strongly than those of other corrections. Keeping them less often than the rest improves the model's eventual performance compared to learning from every correction. Motivated by this, our method, Advisor Self-Distillation (AdviSD), pairs outcome-based reinforcement learning with self-distillation from a feedback-conditioned copy of the advisor selectively. Reflection proposes corrections, and the advisor scores the same recorded executor response with and without its issued advice, using the magnitude of the difference to select decisions for supervision. This approach does not require executor likelihoods or additional executor rollouts. Experiments with Qwen3-8B advisors for Gemini and Claude show that AdviSD outperforms advisor-GRPO by 4.2-6.4 percentage points on BFCL-v3 and by 3.9-5.1 score points on EnvScaler. The trained advisors generalize to out-of-domain tasks and transfer across different executor versions and model families. AdviSD also beats matched-count random selection, supporting the value of its selection rule.

Online training enables computer-use agents (CUAs) to improve through interaction with executable environments. However, existing methods primarily rely on sparse outcome rewards, which provide no supervision for intermediate actions. On-policy self-distillation (OPSD) offers token-level learning signals through privileged rescoring, but directly applying it to CUA online training presents two challenges: fixed guidance may become misaligned with the student's current state, and guidance-induced probability shifts may conflict with step-level correctness. We introduce ComputerSD, an online self-distillation method for CUAs that converts real-time feedback from executed GUI transitions into guidance for policy learning. A fine-tuned GUI analyzer produces guidance and a step-level value score after each action; the guidance provides privileged context, while the score regulates the resulting OPSD signals. ComputerSD jointly optimizes token-level OPSD and trajectory-level GRPO in a fully asynchronous training framework. On OSWorld-Verified, ComputerSD outperforms outcome-only GRPO by 1.9 and 4.1 percentage points on the general-purpose Qwen3-VL-8B-Thinking and specialized EvoCUA-8B backbones, respectively. Evaluation in out-of-distribution settings further supports the generalizability of ComputerSD. These results demonstrate the effectiveness of learning from real-time feedback through online self-distillation for CUAs.

On-policy distillation (OPD) trains a target policy on its own rollouts using token-level supervision from a stronger teacher, but typically relies on the teacher being substantially more capable than the target. Recent work relaxes this requirement by transferring the relative policy difference between an expert and a reference policy rather than the expert's absolute distribution. Yet it remains unclear which differences provide useful supervision and which targets can benefit from them. We develop a unified view in which the log-probability ratio between a proxy expert and a proxy reference defines an implicit reward for KL-regularized policy optimization. This formulation induces an on-policy update for the target, recovers conventional OPD as a special case, and enables any-to-any transfer across models that differ in scale, training stage or generation. Our systematic experiments yield three findings. First, a stronger expert is not necessarily a better source of supervision: differences induced by task-specific training, including reinforcement learning and supervised fine-tuning, transfer consistently, whereas those arising from model scale alone transfer weakly or not at all. Second, the expert and reference serve complementary roles: the expert provides the capability to be transferred, while the reference filters out shared behavior and isolates the relevant residual. Third, even a useful difference does not transfer to every target: success requires the induced update to move the target toward the expert and generalize beyond sampled rollouts. Raw policy distance does not predict transferability, whereas an appropriate reference can rescue otherwise unsuccessful transfer. These results characterize when relative policy differences provide effective supervision for OPD.

Reasoning in large language models is typically trained via supervised fine-tuning (SFT) on expert traces, often framed as knowledge distillation, or reinforcement learning (RL) with outcome-based verifiable rewards. However, SFT focuses on imitation rather than optimisation, while outcome-based RL requires a well-defined reward function. We propose an inverse reinforcement learning (IRL) framework that learns (partially) dense token-level reasoning reward models directly from expert demonstrations. We show that this learned reward serves a dual purpose: (1) as a dense training signal that optimises policies to reason more effectively, outperforming SFT baselines on GSM8K (79\% vs. 56\%) and MedReason (74\% vs. 65\%); and (2) as an inference-time assistant that improves performance via reward-guided reranking, yielding gains of up to 12 percentage points on Llama3 architectures. Furthermore, our dense rewards provide interpretable, step-wise diagnostics that can indicate the location of logical errors. This work proposes a process-level reasoning learning framework from data, bridging the gap between imitation and reinforcement learning for reasoning.