| Loop the Loopies! | | 80.5 | Loopie-20B-A2B (Thinking) | 2026-07-17 |
| JustRL: Scaling a 1.5B LLM with a Simple RL Recipe | ✓ Link | 76.59 | JustRL-Nemotron 1.5B | 2025-12-18 |
| ZGCM-1: A Fully Open and Extremely Efficient Foundation Model for Math and Agentic Search | ✓ Link | 76.06 | ZGCM-1 | 2026-09-11 |
| LLaDA2.0: Scaling Up Diffusion Language Models to 100B | ✓ Link | 74.07 | LLaDA2.0-flash | 2025-12-10 |
| VASPO: Self-Distillation with Privileged Value Estimation | ✓ Link | 73.12 | VASPO (Qwen3-4B-Instruct-2507) | 2026-10-14 |
| Stop When Reasoning Converges: Semantic-Preserving Early Exit for Reasoning Models | ✓ Link | 72.3 | Qwen3-30B-A3B-Thinking with PUMA | 2026-05-17 |
| Self-Hinting Language Models Enhance Reinforcement Learning | ✓ Link | 71.2 | SAGE (Qwen3-4B-Instruct) | 2026-02-03 |
| Filter, Then Reweight: Rethinking Optimization Granularity in On-Policy Distillation | ✓ Link | 70.47 | FiRe-OPD | 2026-06-01 |
| TGRL: Temperature-Grouped Reinforcement Learning for Efficient Exploration in LLMs | ✓ Link | 69.9 | TGRL (Qwen3-32B) (extra data) | 2026-09-27 |
| DoubtLess: Training-Free Contrastive Decoding for Efficient Reasoning | ✓ Link | 69.0 | DoubtLess (Qwen3-32B steered by Qwen3-0.6B) | 2026-10-13 |
| SAGE: Mitigating Long-Horizon Reasoning Biases via Topological Guidance | ✓ Link | 68.26 | Qwen3.5-35B w/SAGE | 2026-09-24 |
| Agentic Chain-of-Thought Steering for Efficient and Controllable LLM Reasoning | ✓ Link | 67.4 | ACTS_{π_SFT} (Qwen3-8B) | 2026-06-02 |
| Think Longer to Explore Deeper: Learn to Explore In-Context via Length-Incentivized Reinforcement Learning | ✓ Link | 67.0 | Qwen3-4B (GSPO + LIE) | 2026-02-12 |
| FlowBalance: Verifier-Grounded Self-Improvement from On-Policy Reasoning Experience | ✓ Link | 66.85 | Qwen3-8B + FlowBalance | 2026-09-03 |
| LEAD: Length-Efficient Adaptive and Dynamic Reasoning for Large Language Models | ✓ Link | 65.14 | LEAD (DeepSeek-R1-Distill-Qwen-7B) | 2026-05-10 |
| Memory-Efficient Looped Transformer: Decoupling Compute from Memory in Looped Language Models | | 64.7 | MELT-1.6B | 2026-05-08 |
| OmniOPD: Logit-Free On-Policy Distillation via Speculative Verification | | 64.58 | OmniOPD (Qwen3-4B student, Gemini-2.5-flash teacher) | 2026-05-31 |
| RISE: Recursive Improvement via Self-Extrapolating Policy Distillation | | 64.4 | RISE (weight,Qwen3-8B,3-seedmean) | 2026-09-04 |
| Selection Is Not Learning: Choosing Model Futures | | 64.14 | EvoHeir (Gemma-4-E4B-it) | 2026-10-15 |
| Native Parallel Reasoner: Reasoning in Parallelism via Self-Distilled Reinforcement Learning | ✓ Link | 63.7 | NPR (Qwen3-4B-Instruct-2507) | 2025-12-08 |
| ThreadWeaver: Adaptive Threading for Efficient Parallel Reasoning in Language Models | ✓ Link | 63.5 | ThreadWeaver (Qwen3-8B) | 2025-11-24 |
| Where the Model Changes Its Mind: Hindsight-Divergence Localization for Efficient Reinforcement Learning with Verifiable Rewards | | 62.95 | HDL (Llama3.1-8B) | 2026-09-29 |
| On the Off-Policy Teacher in On-Policy Distillation | | 62.4 | SCOUT | 2026-09-29 |
| TPPO: Tracking-Value PPO For Language Model Reasoning | ✓ Link | 62.4 | TPPO (Qwen3-4B) | 2026-10-12 |
| Recovering Reliable Supervision via Critique-Guided On-Policy Distillation | ✓ Link | 62.33 | CriticOPD (Olmo-3-7B-Instruct-SFT) | 2026-10-14 |
| Small Generalizable Prompt Predictive Models Can Steer Efficient RL Post-Training of Large Reasoning Models | | 62.1 | DSR-7B with GPS | 2026-02-02 |
| MetaAgent-X : Breaking the Ceiling of Automatic Multi-Agent Systems via End-to-End Reinforcement Learning | | 61.0 | MetaAgent-X RL (Qwen3 8B) | 2026-05-14 |
| Graph-Based Chain-of-Thought Pruning for Reducing Redundant Reflections in Reasoning LLMs | | 59.85 | DeepSeek-R1-Distill-Qwen-7B (Graph-based CoT Pruning) | 2026-04-07 |
| Group Adaptive Clipping Policy Optimization | ✓ Link | 59.4 | DeepSeek-R1-Distill-Qwen-1.5B + GAPO | 2026-08-31 |
| Towards Full Pipeline FP8 Reinforcement Learning for LLMs | | 58.3 | Qwen3-8B-Base + Calibrated Clipping (Rowwise FP8) | 2026-09-19 |
| Semifactual Credit-Augmented Policy Optimization | ✓ Link | 56.82 | SCAPO (Qwen3-4B-Base) | 2026-09-30 |
| KV4-RL: Unlocking Faster Asynchronous Rollouts for RLVR with INT4 KV Cache Quantization | | 56.8 | KV4-RL (Qwen3-8B-Base) | 2026-10-10 |
| Rethinking Critic Learning in PPO: Understanding and Mitigating Value Flattening | ✓ Link | 54.5 | SP3O Qwen3-8B-Base | 2026-09-16 |
| Beyond Imitation: Filtering On-Policy Distillation by Reasoning Progress | | 54.31 | R²-OPD (Qwen3-1.7B) | 2026-08-19 |
| Recursive Think-Answer Process for LLMs and VLMs | | 53.8 | R1-Distill-Qwen-7B-R-TAP | 2026-03-02 |
| DiagEvo: Diagnosis-Guided Self-Evolution via Hierarchical Error Memory | | 53.8 | DiagEvo (Qwen3-8B, 235B-A22B diagnostician) | 2026-09-01 |
| Expert-Space Exploration in MoE Reinforcement Learning | ✓ Link | 53.5 | Qwen3-30B-A3B + ESRL | 2026-09-11 |
| Eliciting Weak-to-Strong Generalization with On-Policy Reverse Distillation | ✓ Link | 53.26 | Qwen3-8B + OPRD | 2026-09-08 |
| J-Zero: Unified Challenger--Solver--Judge Co-Evolution from Zero Data | ✓ Link | 53.12 | J-Zero-8B | 2026-08-27 |
| ReSPO: Reshaped Sequence Policy Optimization for Gradient Starvation in Off-Policy Learning | ✓ Link | 51.2 | ReSPO (Qwen3-30B-A3B-Base, N=16) | 2026-09-28 |
| ASI-Evolve: AI Accelerates AI | ✓ Link | 50.96 | Algorithm1 (ASI-Evolve) | 2026-03-31 |
| Your Group-Relative Advantage Is Biased | | 50.0 | DAPO+HA-DW (Qwen-3-8B-Base) | 2026-01-13 |
| GRAIL: Gradient-Reweighted Advantages for Reinforcement Learning with Verifiable Rewards | ✓ Link | 49.17 | GRAIL (Qwen3-8B) | 2026-06-03 |
| Can LLMs Learn to Reason Robustly under Noisy Supervision? | ✓ Link | 48.9 | Qwen3-4B-Base (with OLR) | 2026-04-05 |
| Rethinking Training-Inference Mismatch in LLM Reinforcement Learning: Where It Arises and How to Correct It | ✓ Link | 48.62 | Qwen3-30B-A3B + CIS | 2026-09-26 |
| MAXS: Meta-Adaptive Exploration with LLM Agents | ✓ Link | 48.47 | MAXS (MiMo-VL-7B-SFT) | 2026-01-14 |
| Learning to Hint for Reinforcement Learning | ✓ Link | 48.1 | HiLL (Qwen2.5-7B-Instruct) | 2026-04-01 |
| DiRL: An Efficient Post-Training Framework for Diffusion Language Models | ✓ Link | 47.3 | DiRL-8B-Instruct | 2025-12-23 |
| Boosting Reinforcement Learning with Verifiable Rewards via Randomly Selected Few-Shot Guidance | ✓ Link | 47.0 | FEST-GRPO | 2026-05-14 |
| On-policy Distillation with Verifiable Reward | ✓ Link | 47.0 | OPDVR (Qwen3-4B) | 2026-08-25 |
| When Reasoning Meets Its Laws | ✓ Link | 46.8 | DeepSeek-R1-8B (SFT-Compo) | 2025-12-19 |
| Questioning the Questions: Sustaining Self-Evolution in Reasoning Models | ✓ Link | 46.67 | R-Quest (Qwen3-4B-Base) | 2026-10-03 |
| Reasoning Palette: Modulating Reasoning via Latent Contextualization for Controllable Exploration for (V)LMs | | 46.61 | Qwen3-8B-Base with Reasoning Palette (RLOO, Linear Decay) | 2025-12-19 |
| LLaDA MoE v2: Scaling Mixture-of-Experts Diffusion Language Models | | 46.44 | LLaDA MoE v2 Instruct | 2026-08-04 |
| Guided Self-Evolving LLMs with Minimal Human Supervision | | 46.4 | Qwen3-8B-Base (R-Few 5%) | 2025-12-02 |
| Enabling Token-wise Importance Sampling in Off-Policy RLVR with Proximal KL Regularization | ✓ Link | 45.99 | Proximal KL (Qwen3-4B-Base, off-policy, staleness 16, lr 5e-6) | 2026-10-13 |
| Solving Physics Olympiad via Reinforcement Learning on Physics Simulators | ✓ Link | 44.53 | Qwen2.5-32B + RL (synthetic) | 2026-04-13 |
| CERA-MoA: Co-Evolving Routing Mechanisms with Continually Learning LLM Agents | ✓ Link | 44.1 | CERA-MoA (Heterogeneous pool) | 2026-09-16 |
| Spectral-Driven Adaptive-Rank Optimizer for Large Language Model Reinforcement Learning with Verifiable Rewards | ✓ Link | 43.4 | Ares (Qwen3-8B-Base) | 2026-10-12 |
| SPARK: Stepwise Process-Aware Rewards for Reference-Free Reinforcement Learning | | 42.7 | PRM-CoT (Process-Aware) (Qwen2.5-Math-7B) | 2025-12-02 |
| From P(y|x) to P(y): Investigating Reinforcement Learning in Pre-train Space | ✓ Link | 41.82 | DSRL (Qwen3-8B) | 2026-04-15 |
| Multiplex Thinking: Reasoning via Token-wise Branch-and-Merge | ✓ Link | 41.7 | Multiplex Thinking (DeepSeek-R1-Distill-Qwen-7B) | 2026-01-13 |
| Rethinking RL for LLM Reasoning: It's Sparse Policy Selection, Not Capability Learning | ✓ Link | 41.1000 | ReasonMaxxer (Qwen2.5-7B) | 2026-05-07 |
| Stratagem: Learning Transferable Reasoning via Trajectory-Modulated Game Self-Play | ✓ Link | 39.9 | Stratagem (All Games) | 2026-04-20 |
| From Imitation to Discrimination: Toward A Generalized Curriculum Advantage Mechanism Enhancing Cross-Domain Reasoning Tasks | | 39.7 | GRPO (with CAPO) (Qwen2.5-7B-Math) | 2025-12-02 |
| The Molecular Structure of Thought: Mapping the Topology of Long Chain-of-Thought Reasoning | | 37.48 | Qwen-2.5-7B-Instruct + 20K OSS-Mole-Syn | 2026-01-09 |
| The Master Key Hypothesis: Unlocking Cross-Model Capability Transfer via Linear Subspace Alignment | ✓ Link | 36.3 | Qwen3 (14B) + UNLOCK from 4B | 2026-04-07 |
| FlashREINFORCE: Critic-Free Single-Rollout Asynchronous RL for Agentic Language Models | ✓ Link | 34.5 | FlashREINFORCE (Qwen2.5-Math-1.5B) | 2026-09-14 |
| OVD: On-policy Verbal Distillation | | 31.6 | OVD (Qwen2.5-Math-1.5B, 2 random examples, 800 steps) | 2026-01-29 |
| Are We Measuring Strategy or Phrasing? The Gap Between Surface- and Approach-Level Diversity in LLM Math Reasoning | | 29.7 | DIVER-Judge (Qwen2.5-3B) | 2026-06-29 |
| Every Token Counts: Generalizing 16M Ultra-Long Context in Large Language Models | ✓ Link | 27.85 | HSA-UltraLong-Inst | 2025-11-28 |
| Learn What's Left, Not What's Mastered: Saturation Aware Advantage Reweighting for Multi-Reward Policy Optimization | | 26.1 | Qwen2.5-7B-Instruct with SA-MRPO (3 objectives) (extra data) | 2026-08-17 |
| Rationale-Guided Policy Optimization: Learning to Reason with Adaptive Rationale Scaffolding | ✓ Link | 24.9 | RGPO | 2026-10-05 |
| LastOPD: Taming Collapse in Latent On-Policy Distillation | ✓ Link | 24.26 | LastOPD (Qwen3-4B -> Qwen3-1.7B-Base) | 2026-09-23 |
| Smaller Models are Natural Explorers for Policy-Level Diversity in GRPO | ✓ Link | 21.9 | 4B→14B S2L | 2026-06-02 |
| Mid-Training with Self-Generated Data Improves Reinforcement Learning in Language Models | | 8.66 | Llama 3.2–3B–Instruct (mid-trained, n=64) | 2026-05-08 |