Research map / Large language models / Reasoning and chain of thought
RL with verifiable rewards for reasoning
Train reasoning LLMs with reinforcement learning and analyze mathematical reasoning gains.
270 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2025–2026) · an approach in Reasoning and chain of thought, keyword cluster “test time · scaling · rl”
State History Tried Baselines Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
0 papers in 2023–24 and 270 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 0.0% to 19.1%.
Open questions the data points to
- Papers report Qwen2.5-7B on MATH500 (pass@1) anywhere from 34.5 to 60.5: comparisons against it depend on whose number you take.
- Papers report Qwen2.5-32B on AMC (pass@1) anywhere from 32.6 to 52.5: comparisons against it depend on whose number you take.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2025 · DAPO: An Open-Source LLM Reinforcement Learning System at Scale (NeurIPS 2025 · 19 citations)
- 2026 · ReTool: Reinforcement Learning for Strategic Tool Use in LLMs (ICLR 2026 · 3 citations)
Earliest papers here
- DAPO: An Open-Source LLM Reinforcement Learning System at Scale (NeurIPS 2025 · 19 citations)
- ReSearch: Learning to Reason with Search for LLMs via Reinforcement Learning (NeurIPS 2025 · 10 citations)
- VL-Rethinker: Incentivizing Self-Reflection of Vision-Language Models with Reinforcement Learning (NeurIPS 2025 · 7 citations)
What has been tried
46 papers here report 516 measured changes (a component added, removed or swapped, compared with a baseline): helped 65%, no clear effect 7%, hurt 25%, mixed 4%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add Overlong Filtering | AIME24 avg@32 · accuracy | 30 -> 36 | Helped | DAPO: An Open-Source LLM Reinforcement Learning System at Scale (NeurIPS 2025 · 19 citations) |
| Add Forced Rethinking during training | MathVision test · Pass@1 accuracy | 29.8 -> 32.3 | Helped | VL-Rethinker: Incentivizing Self-Reflection of Vision-Language Models with Reinforcement Learning (NeurIPS 2025 · 7 citations) |
| Use more RL training steps (150 to 450) | Omni-MATH-Train · pass@1 | 26.1 -> 42.5 | Helped | Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model? (NeurIPS 2025 · 3 citations) |
| use random rather than fixed prompt masking for likelihood estimation | GSM8K RL training · correctness reward | Helped | d1: Scaling Reasoning in Diffusion Large Language Models via Reinforcement Learning (NeurIPS 2025 · 3 citations) | |
| Add policy gradient loss | MATH500 · accuracy | 39.8 -> 71.8 | Helped | Reinforcement Learning for Reasoning in Large Language Models with One Training Example (NeurIPS 2025 · 3 citations) |
| add the decoupled PPO objective | AIME24 · accuracy | 23.3 -> 42.2 | Helped | AREAL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning (NeurIPS 2025 · 3 citations) |
| Train the generator and verifier jointly instead of fixing either component | Algorithmic reasoning test data · generator accuracy | Helped | RL Tango: Reinforcing Generator and Verifier Together for Language Reasoning (NeurIPS 2025 · 2 citations) | |
| Train with NSR only, penalizing incorrect responses without reinforcing correct ones | MATH, AIME 2025, and AMC23 · Pass@k | Helped | The Surprising Effectiveness of Negative Reinforcement in LLM Reasoning (NeurIPS 2025 · 2 citations) | |
| remove Self-Reflection SFT | Average (MathVista, MathVerse, MathVision, MMMU-Pro, Physics) · accuracy | 53.5 -> 51.2 | Hurt | SRPO: Enhancing Multimodal LLM Reasoning via Reflection-Aware Reinforcement Learning (NeurIPS 2025 · 2 citations) |
| Use task-specific reward instead of binary accuracy reward | Text games, overall · accuracy | 57.01% -> 63.24% | Helped | RLVR-World: Training World Models with Reinforcement Learning (NeurIPS 2025 · 2 citations) |
| Set target zero-variance percentage to 0% rather than 25% | Six-benchmark average · accuracy (%) | 48.5 -> 48.1 | No clear effect | Act Only When It Pays: Efficient Reinforcement Learning for LLM Reasoning via Selective Rollouts (NeurIPS 2025 · 2 citations) |
| Increase the number of auxiliary Monte Carlo rollouts K | MATH · accuracy | Helped | VinePPO: Refining Credit Assignment in RL Training of LLMs (ICML 2025 · 1 citation) |
Baseline numbers that disagree
The same baseline on the same benchmark and metric, reported more than 5% apart by different papers here. Training budgets, data splits and reimplementations differ, so state which number you compare with.
| Baseline | Benchmark | Metric | Reported | Papers |
|---|---|---|---|---|
| Qwen2.5-7B | MATH500 | pass@1 | 34.5 – 60.5 | 2 papers, e.g. NeurIPS 2025 and NeurIPS 2025 |
| Qwen2.5-32B | AMC | pass@1 | 32.6 – 52.5 | 2 papers, e.g. NeurIPS 2025 and NeurIPS 2025 |
| Qwen2.5-32B | MATH-500 | pass@1 | 55.8 – 68.6 | 2 papers, e.g. NeurIPS 2025 and NeurIPS 2025 |
What reviewers will expect
Datasets and baselines named in the 270 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- MATH · 50 of 270 papers
- GSM8K · 38 of 270 papers
- MATH500 · 38 of 270 papers
- AIME 2024 · 24 of 270 papers
- DeepScaleR · 21 of 270 papers
- DAPO-Math-17K · 21 of 270 papers
- AIME24 · 19 of 270 papers
- AIME 2025 · 18 of 270 papers
Baselines
- GRPO · 75 of 270 papers
- DAPO · 26 of 270 papers
- SFT · 25 of 270 papers
- GPT-4o · 22 of 270 papers
- Qwen2.5-7B-Instruct · 13 of 270 papers
- Base model · 13 of 270 papers
- RLOO · 11 of 270 papers
- Qwen3-8B · 11 of 270 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 270 papers; from OpenAlex and the papers' affiliations).
- Tsinghua University · 19 papers
- Nanyang Technological University · 11 papers
- Hong Kong University of Science and Technology · 8 papers
- Tencent · 8 papers
- Peking University · 8 papers
- Shanghai Jiao Tong University · 7 papers
- Zhejiang University · 7 papers
- The Chinese University of Hong Kong · 7 papers
- The University of Hong Kong · 6 papers
- Microsoft · 6 papers
Use it
Download BibTeX of 20 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Efficient and adaptive thinking (231 papers)
- Test-time compute scaling (205 papers)
- Routing, pruning and distillation for reasoning (135 papers)
- Process reward models (39 papers)
Elsewhere in the topic
- Long CoT dynamics and efficiency (136 papers)
- Chain-of-thought training and analysis (127 papers)
- Efficient LLM serving (126 papers)
- Search and self-correction in reasoning (123 papers)
- Speculative decoding (92 papers)
- KV cache compression (45 papers)
- Theory of autoregressive reasoning (28 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).