Research map / Reinforcement learning / RL from feedback
RL for LLM reasoning
Apply reinforcement learning to improve multi-step reasoning in language models.
187 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2019–2026) · an approach in RL from feedback, keyword cluster “reasoning · grpo · llm”
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.
13 papers in 2023–24 and 171 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 14.3% to 27.6%.
Open questions the data points to
- Papers report GRPO on GSM8K test (accuracy) anywhere from 45.7 to 89.07: 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
- 2019 · Neural Machine Translation with Adequacy-Oriented Learning (AAAI 2019)
- 2021 · Adaptive Prior-Dependent Correction Enhanced Reinforcement Learning for Natural Language Generation (AAAI 2021 · 1 citation)
- 2022 · Teacher Forcing Recovers Reward Functions for Text Generation (NeurIPS 2022 · 4 citations)
- 2023 · Reward Design with Language Models (ICLR 2023 · 21 citations)
- 2024 · Auto MC-Reward: Automated Dense Reward Design with Large Language Models for Minecraft (CVPR 2024 · 25 citations)
- 2025 · Training Language Models to Self-Correct via Reinforcement Learning (ICLR 2025 · 7 citations)
- 2026 · UI-R1: Enhancing Efficient Action Prediction of GUI Agents by Reinforcement Learning (AAAI 2026 · 6 citations)
Earliest papers here
- Neural Machine Translation with Adequacy-Oriented Learning (AAAI 2019)
- Adaptive Prior-Dependent Correction Enhanced Reinforcement Learning for Natural Language Generation (AAAI 2021 · 1 citation)
- Teacher Forcing Recovers Reward Functions for Text Generation (NeurIPS 2022 · 4 citations)
What has been tried
46 papers here report 558 measured changes (a component added, removed or swapped, compared with a baseline): helped 67%, no clear effect 8%, hurt 21%, 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 Reward Critic to Reward Designer | Explore Diamond Ore Underground · Succ. | 35.1% to 40.5% | Helped | Auto MC-Reward: Automated Dense Reward Design with Large Language Models for Minecraft (CVPR 2024 · 25 citations) |
| Remove multi-turn training | MATH500 · Δ(t1, t2) | 4.4% -> -2.4% | Hurt | Training Language Models to Self-Correct via Reinforcement Learning (ICLR 2025 · 7 citations) |
| Increase verification iterations from one to three, five, or ten | GeneralPoints-L rule OOD · success rate improvement | +0.48% (1 step); +2.15% (3 steps); +2.99% (5 steps); +5.99% (10 steps) | Helped | SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training (ICML 2025 · 6 citations) |
| add episode-level and step-level relative advantages | ALFWorld and WebShop · success rate (%) | Helped | Group-in-Group Policy Optimization for LLM Agent Training (NeurIPS 2025 · 4 citations) | |
| Increase sample reuse from T=1 to T=8 | Synthetic LLM problems · Gold reward | T=8; T=1 | No clear effect | Preference Fine-Tuning of LLMs Should Leverage Suboptimal, On-Policy Data (ICML 2024 · 4 citations) |
| Use PAV as the outcome re-ranker instead of a standard ORM | MATH, N=128 · accuracy | 4% | Helped | Rewarding Progress: Scaling Automated Process Verifiers for LLM Reasoning (ICLR 2025 · 3 citations) |
| remove numerical reference and balance requirements from the reward weight initializer prompt | Underwater data collection task · weight-search iterations | 0.40±0.49 -> 1.20±0.75 | Hurt | ERFSL: An Efficient Reward Function Searcher via Large Language Models for Custom-Environment Multi-Objective Reinforcement Learning (Student Abstract) (AAAI 2025 · 2 citations) |
| Use module-level MDP rather than turn-level MDP for value estimation | MultiWOZ simulation · Success Rate | Helped | Universal Post-Processing Networks for Joint Optimization of Modules in Task-Oriented Dialogue Systems (AAAI 2025 · 2 citations) | |
| Add probability-mask optimization to SPO-chain (int5) | GSM8K test · accuracy | 55.0% -> 56.7% | Helped | Segment Policy Optimization: Effective Segment-Level Credit Assignment in RL for Large Language Models (NeurIPS 2025 · 2 citations) |
| Use IQL + AWR instead of IQL + REINFORCE | Twenty Questions · return | -20 -> -14.1 | Helped | ArCHer: Training Language Model Agents via Hierarchical Multi-Turn RL (ICML 2024 · 2 citations) |
| Reduce block size M | Anthropic HH · win rate vs KL tradeoff | Hurt | Controlled Decoding from Language Models (ICML 2024 · 2 citations) | |
| Set KL coefficient β to 0 | GSM8K CoT · performance | 50.5 -> 46.6 | Hurt | Training Large Language Models for Reasoning through Reverse Curriculum Reinforcement Learning (ICML 2024 · 2 citations) |
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 |
|---|---|---|---|---|
| GRPO | GSM8K test | accuracy | 45.7 – 89.07 | 2 papers, e.g. NeurIPS 2025 and NeurIPS 2025 |
What reviewers will expect
Datasets and baselines named in the 72 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 · 14 of 72 papers
- GSM8K · 10 of 72 papers
- ALFWorld · 5 of 72 papers
- WebShop · 4 of 72 papers
- MMLU · 4 of 72 papers
- HumanEval · 4 of 72 papers
- NuminaMath · 3 of 72 papers
- OpenR1-Math-220K · 3 of 72 papers
Baselines
- GRPO · 15 of 72 papers
- PPO · 10 of 72 papers
- SFT · 7 of 72 papers
- RLOO · 6 of 72 papers
- DPO · 4 of 72 papers
- DAPO · 4 of 72 papers
- ReAct · 3 of 72 papers
- Dr.GRPO · 3 of 72 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 184 papers; from OpenAlex and the papers' affiliations).
- Google · 8 papers
- Tsinghua University · 6 papers
- Google DeepMind · 6 papers
- Meta · 6 papers
- Nanjing University · 6 papers
- Fudan University · 5 papers
- City University of Hong Kong · 5 papers
- Nanyang Technological University · 5 papers
- Huawei · 5 papers
- Tencent · 5 papers
Use it
Download BibTeX of 23 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
- Scaling RL with verifiable rewards (119 papers)
- GRPO stability and variants (106 papers)
- RL fine-tuning of generative models (94 papers)
Elsewhere in the topic
- Theory of RLHF (110 papers)
- Reward modeling versus direct optimization (95 papers)
- Preference-based RL reward learning (26 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).