repoGitHubTrust 82 · PrimaryPublished 6d agoLive · 6d ago
hscspring/rl-llm-nlp
Curated, opinionated index of post-R1 LLM × Reinforcement Learning. Many deep-dive blog posts cross-linked to many papers — GRPO, DAPO, DPO, PPO, RLHF, GSPO, CISPO, VAPO, Reward Modeling, MoE RL stability, Verifier-Free RL, Training-Free RL, Agentic RL, DeepSeek-R1 reproduction.
Lineage graph
Paper → model → repo connections mined from source citations (Tier-1 exact match).
Why these links exist
Every edge carries a method, confidence, and the source snippet that justified it — so bad links are debuggable.
- PossiblePossibly related (embedding) · 65%Is One Layer Enough? Training A Single Transformer Layer Can Match Full-Parameter RL Training →
- PossiblePossibly related (embedding) · 65%Which Tokens Matter? Adaptive Token Selection for RLVR with the Relative Surprisal Index →
- PossiblePossibly related (embedding) · 60%Reinforcement Learning without Ground-Truth Solutions can Improve LLMs →
- PossiblePossibly related (embedding) · 60%RL without TD learning →
- PossiblePossibly related (embedding) · 58%Weak-to-Strong Generalization via Direct On-Policy Distillation →
- PossiblePossibly related (embedding) · 54%Multimodal Reward Hacking in Reinforcement Learning →
- PossiblePossibly related (embedding) · 56%SCOPE-RL: Optimizing Reasoning Paths Before and After Success →
- PossiblePossibly related (embedding) · 55%Active Offline-to-Online Reinforcement Learning →
Implements
paperIs One Layer Enough? Training A Single Transformer Layer Can Match Full-Parameter RL TrainingpaperWhich Tokens Matter? Adaptive Token Selection for RLVR with the Relative Surprisal IndexpaperReinforcement Learning without Ground-Truth Solutions can Improve LLMspaperWeak-to-Strong Generalization via Direct On-Policy Distillation
Covers
Implements (incoming)
paperMultimodal Reward Hacking in Reinforcement LearningpaperSCOPE-RL: Optimizing Reasoning Paths Before and After SuccesspaperActive Offline-to-Online Reinforcement LearningpaperDirectional Constraints for Efficient Exploration in Safe Reinforcement LearningpaperVerifier-Based Reinforcement Fine-Tuning of Reasoning Models for Thermal Energy Storage ControlpaperRing-Zero: Scaling Zero RL to a Trillion Parameters for Emergent ReasoningpaperA Learning-Rate-Gated Failure of GRPO in a Small Language and Vision-Language Model Web Agent: A Controlled Null and Its MechanismpaperFrom Critic to Confidence: PPO for Language-Based Quantitative Prediction with Confidence Estimation
Covers (incoming)
Related across the graph
paperSCOPE-RL: Optimizing Reasoning Paths Before and After SuccesspaperVerifier-Based Reinforcement Fine-Tuning of Reasoning Models for Thermal Energy Storage ControlnewsRL without TD learningpaperWhich Tokens Matter? Adaptive Token Selection for RLVR with the Relative Surprisal IndexpaperActive Offline-to-Online Reinforcement LearningpaperFrom Critic to Confidence: PPO for Language-Based Quantitative Prediction with Confidence EstimationpaperWeak-to-Strong Generalization via Direct On-Policy DistillationpaperA Learning-Rate-Gated Failure of GRPO in a Small Language and Vision-Language Model Web Agent: A Controlled Null and Its MechanismnewsThe Little Book of Reinforcement LearningpaperRing-Zero: Scaling Zero RL to a Trillion Parameters for Emergent ReasoningpaperMultimodal Reward Hacking in Reinforcement LearningpaperIs One Layer Enough? Training A Single Transformer Layer Can Match Full-Parameter RL TrainingpaperReinforcement Learning without Ground-Truth Solutions can Improve LLMspaperDirectional Constraints for Efficient Exploration in Safe Reinforcement Learning