Why Fine-Tuning Encourages Hallucinations and How to Fix It
Paper • 2604.15574 • Published • 26
toread
-
-
Tuna-2: Pixel Embeddings Beat Vision Encoders for Multimodal Understanding and Generation
Paper • 2604.24763 • Published • 71 -
Programming with Data: Test-Driven Data Engineering for Self-Improving LLMs from Raw Corpora
Paper • 2604.24819 • Published • 92 -
GLM-5V-Turbo: Toward a Native Foundation Model for Multimodal Agents
Paper • 2604.26752 • Published • 116 -
Large Language Models Explore by Latent Distilling
Paper • 2604.24927 • Published • 74 -
Accelerating RL Post-Training Rollouts via System-Integrated Speculative Decoding
Paper • 2604.26779 • Published • 16 -
Gated DeltaNet-2: Decoupling Erase and Write in Linear Attention
Paper • 2605.22791 • Published • 34 -
Unsupervised Process Reward Models
Paper • 2605.10158 • Published • 28 -
Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps
Paper • 2605.16928 • Published • 100 -
From Seeing to Thinking: Decoupling Perception and Reasoning Improves Post-Training of Vision-Language Models
Paper • 2605.20177 • Published • 10 -
Rethinking Muon Beyond Pretraining: Spectral Failures and High-Pass Remedies for VLA and RLVR
Paper • 2605.19282 • Published • 9 -
Channel-wise Vector Quantization
Paper • 2605.26089 • Published • 15 -
Negligible in Size, Significant in Effect: On Scale Vectors in Large Language Models
Paper • 2605.26895 • Published • 24 -
Task-Focused Memorization for Multimodal Agents
Paper • 2605.31075 • Published • 41 -
LongTraceRL: Learning Long-Context Reasoning from Search Agent Trajectories with Rubric Rewards
Paper • 2605.31584 • Published • 43 -
Not All Disagreement Is Learnable: Token Teachability in On-Policy Distillation
Paper • 2605.26844 • Published • 26 -
ESPO: Early-Stopping Proximal Policy Optimization
Paper • 2605.29860 • Published • 21 -
NITP: Next Implicit Token Prediction for LLM Pre-training
Paper • 2605.24956 • Published • 37 -
Self-Distilled Policy Gradient
Paper • 2606.04036 • Published • 28 -
MemTrain: Self-Supervised Context Memory Training
Paper • 2606.03197 • Published • 19 -
Latent Reasoning with Normalizing Flows
Paper • 2606.06447 • Published • 8 -
ThoughtFold: Folding Reasoning Chains via Introspective Preference Learning
Paper • 2606.03503 • Published • 25 -
Unified Neural Scaling Laws
Paper • 2605.26248 • Published • 8 -
Your UnEmbedding Matrix is Secretly a Feature Lens for Text Embeddings
Paper • 2606.07502 • Published • 100 -
On the Geometry of On-Policy Distillation
Paper • 2606.07082 • Published • 75 -
Dynamic Linear Attention
Paper • 2606.10650 • Published • 10 -
How Does Reasoning Flow? Tracing Attention-Induced Information Flow for Targeted RL in LLMs
Paper • 2606.10646 • Published • 9 -
Rethinking the Divergence Regularization in LLM RL
Paper • 2606.09821 • Published • 34 -
Redesign Mixture-of-Experts Routers with Manifold Power Iteration
Paper • 2606.12397 • Published • 91 -
MiniMax Sparse Attention
Paper • 2606.13392 • Published • 166 -
Nemotron 3 Ultra: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning
Paper • 2606.15007 • Published • 20 -
Data Journalist Agent: Transforming Data into Verifiable Multimodal Stories
Paper • 2606.11176 • Published • 133 -
Breaking Entropy Bounds: Accelerating RL Training via MTP with Rejection Sampling
Paper • 2606.12370 • Published • 22 -
Ling and Ring 2.6 Technical Report: Efficient and Instant Agentic Intelligence at Trillion-Parameter Scale
Paper • 2606.15079 • Published • 88 -
ExpRL: Exploratory RL for LLM Mid-Training
Paper • 2606.17024 • Published • 7 -
Variable-Width Transformers
Paper • 2606.18246 • Published • 16 -
Zone of Proximal Policy Optimization: Teacher in Prompts, Not Gradients
Paper • 2606.18216 • Published • 65 -
Rethinking the Role of Efficient Attention in Hybrid Architectures
Paper • 2606.15378 • Published • 20 -
Manifold Bandits: Bayesian Curriculum Learning over the Latent Geometry of Large Language Models
Paper • 2606.19750 • Published • 4 -
Beyond Reward Engineering: A Data Recipe for Long-Context Reinforcement Learning
Paper • 2606.18831 • Published • 8 -
The Hitchhiker's Guide to Agentic AI: From Foundations to Systems
Paper • 2606.24937 • Published • 22 -
Holistic Data Scheduler for LLM Pre-training via Multi-Objective Reinforcement Learning
Paper • 2606.24133 • Published • 12 -
Formalizing Latent Thoughts: Four Axioms of Thought Representation in LLMs
Paper • 2606.27378 • Published • 61 -
MultiHashFormer: Hash-based Generative Language Models
Paper • 2606.28057 • Published • 23 -
One-Step Gradient Delay is Not a Barrier for Large-Scale Asynchronous Pipeline Parallel LLM Pretraining
Paper • 2606.30634 • Published • 25 -
Discretizing Reward Models
Paper • 2606.21795 • Published • 17 -
Morphing into Hybrid Attention Models
Paper • 2606.30562 • Published • 52 -
The State-Prediction Separation Hypothesis
Paper • 2607.01218 • Published • 12 -
Program-as-Weights: A Programming Paradigm for Fuzzy Functions
Paper • 2607.02512 • Published • 309 -
ResearchStudio-Idea: An Evidence-Grounded Research-Ideation Skill Suite from ML Conference Outcomes
Paper • 2607.04439 • Published • 63 -
ResearchStudio-Reel: Automate the Last Mile of Research from Paper to Poster, Video, and Blog
Paper • 2607.04438 • Published • 64 -
Do All Visual Tokens Matter Equally? Object-Evidence Preserving Token Merging for Vision-Language Retrieval
Paper • 2607.04605 • Published • 24 -
MANCE: Manifold Aware Concept Erasure
Paper • 2607.03973 • Published • 49 -
LLM-as-a-Verifier: A General-Purpose Verification Framework
Paper • 2607.05391 • Published • 18 -
Hierarchical Sparse Attention Done Right: Toward Infinite Context Modeling
Paper • 2607.02980 • Published • 84 -
OmniOpt: Taxonomy, Geometry, and Benchmarking of Modern Optimizers
Paper • 2607.04033 • Published • 77 -
TurnOPD: Making On-Policy Distillation Turn-Aware for Efficient Long-Horizon Agent Training
Paper • 2607.05804 • Published • 20 -
Sparse Delta Memory: Scaling the State of Linear RNNs through Sparsity
Paper • 2607.07386 • Published • 13 -
Jet-Long: Efficient Long-Context Extension with Dynamic Bifocal RoPE
Paper • 2607.07740 • Published • 25 -
Trust Region Policy Distillation
Paper • 2607.04751 • Published • 36 -
Scalable Visual Pretraining for Language Intelligence
Paper • 2607.09657 • Published • 58 -
Weak-to-Strong Generalization via Direct On-Policy Distillation
Paper • 2607.05394 • Published • 150 -
Token-Level Off-Policy Learning for Faithful Generation Under Distribution Shift
Paper • 2607.17524 • Published • 7 -
DataFlow-Harness: A Grounded Code-Agent Platform for Constructing Editable LLM Data Pipelines
Paper • 2607.16617 • Published • 145 -
SLPO: Scaling Latent Reasoning via a Surrogate Policy
Paper • 2607.19691 • Published • 6 -
Diagnosing and Calibrating Tool-Call Boundary Drift in Multi-Teacher On-Policy Distillation
Paper • 2607.07050 • Published • 5 -
LongStraw: Long-Context RL Beyond 2M Tokens under a Fixed GPU Budget
Paper • 2607.14952 • Published • 213 -
Molt: A Scalable PyTorch-Native Training Framework for Agentic Reinforcement Learning
Paper • 2607.21653 • Published • 32 -
Three-Body Scattering for Generative Modeling
Paper • 2607.18198 • Published • 16 -
The Physics of Multi-Turn Long-Horizon Planning: From Pre-training to Post-training via Single- and Multi-Teacher On-Policy Agentic Distillation
Paper • 2607.24720 • Published • 26 -
Kimi K3: Open Frontier Intelligence
Paper • 2607.24653 • Published • 515 -
Memory for Large Language Models
Paper • 2607.25380 • Published • 16 -
Let's Scale Step by Step: Compute-Efficient Hyperparameter Transfer for Large-Scale Mixture-of-Experts
Paper • 2608.20061 • Published • 46 -
Does On-Policy Distillation Really Distill? From Noisy Teacher to Self-Improvement
Paper • 2608.31046 • Published • 143 -
SHAPE of Chain-of-Thought in Math Reasoning
Paper • 2608.28600 • Published • 31 -
On the Design of Qwen3.8-Next Architecture: Evaluation, Efficiency, and Training Stability
Paper • 2608.30320 • Published • 53 -
Verification-Aware Training for Speculative Decoding
Paper • 2608.30135 • Published • 9 -
Normalized Low-Rank Adaptation
Paper • 2608.31036 • Published • 50 -
Language Models Can Control Their Own Attention
Paper • 2609.02737 • Published • 64 -
Post-Training Language Models for Gold-Medal Performance in Coding Competitions
Paper • 2609.02849 • Published • 9 -
Rethinking On-Policy Distillation of Large Language Models II: One Training Example
Paper • 2609.04172 • Published • 81 -
Locked at the Entrance, Open Inside: Where RLVR Narrows the Solution Space
Paper • 2608.29188 • Published • 10