| Filter, Then Reweight: Rethinking Optimization Granularity in On-Policy Distillation | ✓ Link | 68.49 | FiRe-OPD | 2026-06-01 |
| DiagEvo: Diagnosis-Guided Self-Evolution via Hierarchical Error Memory | | 64.8 | DiagEvo (Qwen3-8B, 235B-A22B diagnostician) | 2026-09-01 |
| VASPO: Self-Distillation with Privileged Value Estimation | ✓ Link | 64.75 | VASPO (Qwen3-4B-Instruct-2507) | 2026-10-14 |
| Off-the-Shelf LLMs as Process Scorers: Training-Free Alternative to PRMs for Mathematical Reasoning | | 63.6 | CGS (Qwen2.5-7B → Qwen2.5-72B) | 2026-06-01 |
| J-Zero: Unified Challenger--Solver--Judge Co-Evolution from Zero Data | ✓ Link | 61.76 | J-Zero-8B | 2026-08-27 |
| DARC: Decoupled Asymmetric Reasoning Curriculum for LLM Evolution | ✓ Link | 61.4 | DARC (Qwen3-8B-Base) | 2026-01-20 |
| Guided Self-Evolving LLMs with Minimal Human Supervision | | 60.3 | Qwen3-8B-Base (R-Few 5%) | 2025-12-02 |
| Questioning the Questions: Sustaining Self-Evolution in Reasoning Models | ✓ Link | 58.82 | R-Quest (Qwen3-4B-Base) | 2026-10-03 |
| SoloQ: Calibration-Free Quantization for Diffusion Language Models | | 54.86 | SoloQ-C (Fast-dLLM v2-7B) | 2026-10-05 |
| FlowBalance: Verifier-Grounded Self-Improvement from On-Policy Reasoning Experience | ✓ Link | 53.68 | Qwen3-8B + FlowBalance | 2026-09-03 |
| KV4-RL: Unlocking Faster Asynchronous Rollouts for RLVR with INT4 KV Cache Quantization | | 51.8 | KV4-RL (Qwen3-8B-Base) | 2026-10-10 |
| JustRL: Scaling a 1.5B LLM with a Simple RL Recipe | ✓ Link | 51.47 | JustRL-DeepSeek 1.5B | 2025-12-18 |
| SAGE: Mitigating Long-Horizon Reasoning Biases via Topological Guidance | ✓ Link | 51.34 | Qwen3.5-35B w/SAGE | 2026-09-24 |
| TPPO: Tracking-Value PPO For Language Model Reasoning | ✓ Link | 51.0 | TPPO (Qwen3-4B) | 2026-10-12 |
| TGRL: Temperature-Grouped Reinforcement Learning for Efficient Exploration in LLMs | ✓ Link | 50.6 | TGRL (Qwen3-32B) (extra data) | 2026-09-27 |
| Anti-Self-Distillation for Reasoning RL via Pointwise Mutual Information | ✓ Link | 49.7 | Qwen3-30B-A3B + AntiSD | 2026-05-12 |
| Self-Hinting Language Models Enhance Reinforcement Learning | ✓ Link | 49.2 | SAGE (Qwen3-4B-Instruct) | 2026-02-03 |
| Rethinking Critic Learning in PPO: Understanding and Mitigating Value Flattening | ✓ Link | 47.93 | SP3O Qwen3-4B-Base | 2026-09-16 |
| Native Parallel Reasoner: Reasoning in Parallelism via Self-Distilled Reinforcement Learning | ✓ Link | 47.1 | NPR (Qwen3-4B) | 2025-12-08 |
| iGRPO: Self-Feedback-Driven LLM Reasoning | | 47.06 | DeepSeek-R1-Distill-Qwen-14B + iGRPO | 2026-02-09 |
| Unsupervised Process Reward Models | | 46.7 | uPRM | 2026-05-11 |
| ResRL: Boosting LLM Reasoning via Negative Sample Projection Residual Reinforcement Learning | ✓ Link | 46.0 | ResRL (Qwen3-8B) | 2026-05-01 |
| The Mirage of Optimizing Training Policies: Monotonic Inference Policies as the Real Objective for LLM Reinforcement Learning | | 45.96 | Qwen3-4B (MIPU) | 2026-06-28 |
| ReSPO: Reshaped Sequence Policy Optimization for Gradient Starvation in Off-Policy Learning | ✓ Link | 45.7 | ReSPO (Qwen3-30B-A3B-Base, N=16) | 2026-09-28 |
| Selection Is Not Learning: Choosing Model Futures | | 45.22 | EvoHeir (Qwen3-8B-Base) | 2026-10-15 |
| Heterogeneous Agent Collaborative Reinforcement Learning | | 45.2 | Qwen3-4B-Instruct (HACPO) | 2026-03-03 |
| Can LLMs Learn to Reason Robustly under Noisy Supervision? | ✓ Link | 44.9 | Qwen3-4B-Base (with OLR) | 2026-04-05 |
| GRAIL: Gradient-Reweighted Advantages for Reinforcement Learning with Verifiable Rewards | ✓ Link | 44.73 | GRAIL (Qwen3-8B) | 2026-06-03 |
| Towards Full Pipeline FP8 Reinforcement Learning for LLMs | | 44.6 | Qwen3-8B-Base + Calibrated Clipping (Tensorwise FP8) | 2026-09-19 |
| Figure It Out: Improving the Frontier of Reasoning with Active Visual Thinking | ✓ Link | 44.49 | FigR | 2025-12-30 |
| Rethinking Cross-Tokenizer On-Policy Distillation: From Alignment Coverage to Supervision Reliability | | 44.11 | Strict top-16 (Qwen3-235B-A22B-Instruct-2507 -> Granite-4.1-8B) | 2026-10-06 |
| Reasoning Palette: Modulating Reasoning via Latent Contextualization for Controllable Exploration for (V)LMs | | 43.77 | Qwen3-8B-Base with Reasoning Palette (RLOO, Linear Decay) | 2025-12-19 |
| ThreadWeaver: Adaptive Threading for Efficient Parallel Reasoning in Language Models | ✓ Link | 43.7 | ThreadWeaver (Qwen3-8B) | 2025-11-24 |
| Semifactual Credit-Augmented Policy Optimization | ✓ Link | 43.38 | SCAPO (Qwen3-4B-Base) | 2026-09-30 |
| Recursive Think-Answer Process for LLMs and VLMs | | 42.3 | R1-Distill-Qwen-7B-R-TAP | 2026-03-02 |
| Your Group-Relative Advantage Is Biased | | 41.9 | GSPO+HA-DW (Qwen-3-8B-Base) | 2026-01-13 |
| Unlocking Token Rewards via Training-Free Reward Attribution | | 41.9 | Qwen2.5-Math-7B + GRPO (P2T token reward) (extra data) | 2026-06-03 |
| PretrainZero: Reinforcement Active Pretraining | | 41.2 | PretrainZero (Qwen3-30B-A3B-MoE-Base) | 2025-12-03 |
| Enabling Token-wise Importance Sampling in Off-Policy RLVR with Proximal KL Regularization | ✓ Link | 40.44 | Proximal KL (Qwen3-4B-Base, off-policy, staleness 4, lr 5e-6) | 2026-10-13 |
| From Imitation to Discrimination: Toward A Generalized Curriculum Advantage Mechanism Enhancing Cross-Domain Reasoning Tasks | | 40.1 | REINFORCE++ (with CAPO) (Qwen2.5-7B-Math) | 2025-12-02 |
| SPARK: Stepwise Process-Aware Rewards for Reference-Free Reinforcement Learning | | 40.1 | PRM-CoT (Process-Aware) (Qwen2.5-Math-7B) | 2025-12-02 |
| Learning to Hint for Reinforcement Learning | ✓ Link | 40.1 | HiLL (Qwen2.5-7B-Instruct) | 2026-04-01 |
| Small Generalizable Prompt Predictive Models Can Steer Efficient RL Post-Training of Large Reasoning Models | | 39.7 | DSR-7B with GPS | 2026-02-02 |
| TraPO: A Semi-Supervised Reinforcement Learning Framework for Boosting LLM Reasoning | ✓ Link | 39.3 | Qwen2.5-Math-7B with TraPO (4K labels, 12K unlabeled) | 2025-12-15 |
| The Master Key Hypothesis: Unlocking Cross-Model Capability Transfer via Linear Subspace Alignment | ✓ Link | 39.2 | Qwen3 (14B) + UNLOCK from 4B | 2026-04-07 |
| Multiplex Thinking: Reasoning via Token-wise Branch-and-Merge | ✓ Link | 38.6 | Multiplex Thinking (DeepSeek-R1-Distill-Qwen-7B) | 2026-01-13 |
| Coupled Variational Reinforcement Learning for Language Model General Reasoning | ✓ Link | 38.1 | Qwen3-14B with CoVRL | 2025-12-14 |
| SPOT: Sparse Probing and Outcome Calibration for On-Policy Distillation | ✓ Link | 37.73 | SPOT Qwen3-4B-Base | 2026-08-05 |
| Recovering Reliable Supervision via Critique-Guided On-Policy Distillation | ✓ Link | 37.44 | CriticOPD (Qwen3-1.7B-Non-Thinking) | 2026-10-14 |
| Latent-MOPD: Latent Multi-Teacher On-Policy Distillation | ✓ Link | 36.3 | Latent-MOPD (last 3) | 2026-10-01 |
| Beyond Policy Rollouts: Embedding-Space Directional Exploration with Anchored Noise for LLM RLVR | | 36.03 | EDEN (Qwen2.5-Math-7B) | 2026-10-12 |
| Boosting Reinforcement Learning with Verifiable Rewards via Randomly Selected Few-Shot Guidance | ✓ Link | 35.45 | FEST-GRPO | 2026-05-14 |
| Guiding LLM Post-training Data Engineering with Model Internals from Sparse Autoencoders | | 35.4 | Qwen2.5-Math-7B (SaeRL, GRPO) | 2026-05-26 |
| Rethinking Training-Inference Mismatch in LLM Reinforcement Learning: Where It Arises and How to Correct It | ✓ Link | 34.31 | Qwen3-30B-A3B + CIS | 2026-09-26 |
| On-policy Distillation with Verifiable Reward | ✓ Link | 33.2 | OPDVR (Qwen3-4B) | 2026-08-25 |
| RISE: Recursive Improvement via Self-Extrapolating Policy Distillation | | 32.6 | RISE (weight,Qwen3-8B) | 2026-09-04 |
| Where the Model Changes Its Mind: Hindsight-Divergence Localization for Efficient Reinforcement Learning with Verifiable Rewards | | 32.49 | HDL (Qwen3-8B) | 2026-09-29 |
| Normalized Low-Rank Adaptation | ✓ Link | 32 | DS-R1-Distill-Qwen1.5B NoRA | 2026-08-31 |
| Low-rank Optimization Trajectories Modeling for LLM RLVR Acceleration | ✓ Link | 30.5 | Qwen2.5-14B-Instruct with NExt | 2026-04-13 |
| Rethinking RL for LLM Reasoning: It's Sparse Policy Selection, Not Capability Learning | ✓ Link | 30.5 | ReasonMaxxer (Qwen3-4B) | 2026-05-07 |
| From P(y|x) to P(y): Investigating Reinforcement Learning in Pre-train Space | ✓ Link | 30.32 | DSRL (Qwen3-8B) | 2026-04-15 |
| Spectral-Driven Adaptive-Rank Optimizer for Large Language Model Reinforcement Learning with Verifiable Rewards | ✓ Link | 29.4 | Ares (Qwen3-8B-Base) | 2026-10-12 |
| Adaptive Ability Decomposing for Unlocking Large Reasoning Model Effective Reinforcement Learning | | 28.9 | A2D (Qwen2.5-7B-Instruct) | 2026-01-31 |
| Learn What's Left, Not What's Mastered: Saturation Aware Advantage Reweighting for Multi-Reward Policy Optimization | | 24.8 | Qwen2.5-7B-Instruct with SA-MRPO (3 objectives) (extra data) | 2026-08-17 |
| RelayLLM: Efficient Reasoning via Collaborative Decoding | ✓ Link | 23.53 | Qwen3-0.6B (RelayLLM Difficulty-Aware-Reward) | 2026-01-08 |
| FlashREINFORCE: Critic-Free Single-Rollout Asynchronous RL for Agentic Language Models | ✓ Link | 22.8 | FlashREINFORCE (Qwen2.5-Math-1.5B) | 2026-09-14 |
| Rationale-Guided Policy Optimization: Learning to Reason with Adaptive Rationale Scaffolding | ✓ Link | 22.1 | RGPO | 2026-10-05 |
| Self-Play Search Distillation for Large Language Model Reasoning | | 21.7 | SPSD-Qwen3-4B-Base (OPSD) | 2026-09-25 |
| OVD: On-policy Verbal Distillation | | 18.0 | OVD (Qwen2.5-Math-1.5B, 2 random examples, 800 steps) | 2026-01-29 |
| LastOPD: Taming Collapse in Latent On-Policy Distillation | ✓ Link | 17.74 | LastOPD (Qwen3-4B -> Qwen3-1.7B-Base) | 2026-09-23 |