Research map / Reinforcement learning / RL from feedback
GRPO stability and variants
Analyze and fix group-relative policy optimization issues such as noise, collapse, and trust regions.
106 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2020–2026) · an approach in RL from feedback, keyword cluster “reasoning · grpo · llm”
State History Tried 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.
1 papers in 2023–24 and 103 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 1.1% to 16.6%.
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
- 2020 · When Counterpoint Meets Chinese Folk Melodies (NeurIPS 2020 · 5 citations)
- 2024 · SurgicAI: A Fine-grained Platform for Data Collection and Benchmarking in Surgical Policy Learning (NeurIPS 2024)
- 2025 · CPPO: Accelerating the Training of Group Relative Policy Optimization-Based Reasoning Models (NeurIPS 2025 · 3 citations)
- 2026 · OmniSapiens: A Foundation Model for Social Behavior Processing via Heterogeneity-Aware Relative Policy Optimization (ICML 2026 · 2 citations)
Earliest papers here
- When Counterpoint Meets Chinese Folk Melodies (NeurIPS 2020 · 5 citations)
- AvgOut: A Simple Output-Probability Measure to Eliminate Dull Responses (AAAI 2020 · 4 citations)
- SurgicAI: A Fine-grained Platform for Data Collection and Benchmarking in Surgical Policy Learning (NeurIPS 2024)
What has been tried
13 papers here report 152 measured changes (a component added, removed or swapped, compared with a baseline): helped 68%, no clear effect 11%, hurt 18%, mixed 3%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| prune completions by absolute advantage | GSM8K test subset · accuracy | 55.72% -> 77.67% | Helped | CPPO: Accelerating the Training of Group Relative Policy Optimization-Based Reasoning Models (NeurIPS 2025 · 3 citations) |
| Add offline question variants with global shared advantage | MathVista · accuracy | +1.1% | Helped | R1-ShareVL: Incentivizing Reasoning Capabilities of Multimodal Large Language Models via Share-GRPO (NeurIPS 2025 · 3 citations) |
| Replace original-response initialization with Gaussian-noise initialization | AdvBench · Harmful Rate (%) | 5.58 -> 4.04 | Helped | Inference-time Alignment in Continuous Space (NeurIPS 2025 · 1 citation) |
| Set exponential transformation to t=10 | Anthropic Helpfulness, Best-of-4 · inference-time win rate | Helped | InfAlign: Inference-aware language model alignment (ICML 2025 · 0 citations) | |
| Increase the number of reference rewards used to estimate quantile rewards | Magpie-Air off-policy · online test reward | Helped | Quantile Reward Policy Optimization: Alignment with Pointwise Regression and Exact Partition Functions (NeurIPS 2025 · 0 citations) | |
| add the current-policy variance regularizer | AIME24 · accuracy | 0.00 -> 20.72 | Helped | GVPO: Group Variance Policy Optimization for Large Language Model Post-Training (NeurIPS 2025 · 0 citations) |
| Remove the hard-negative weighting by using DisCO rather than DisCO-b | AIME datasets and other evaluated settings · performance | Helped | DisCO: Reinforcing Large Reasoning Models with Discriminative Constrained Optimization (NeurIPS 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 33 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- DAPO-Math-17K · 5 of 33 papers
- MATH · 4 of 33 papers
- GSM8K · 3 of 33 papers
- DAPO-Math · 3 of 33 papers
- DeepScaleR · 3 of 33 papers
Baselines
- GRPO · 16 of 33 papers
- DAPO · 6 of 33 papers
- GSPO · 4 of 33 papers
- Dr. GRPO · 3 of 33 papers
- REINFORCE++ · 3 of 33 papers
- DPO · 2 of 33 papers
- SAPO · 2 of 33 papers
- KL-Cov · 2 of 33 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 104 papers; from OpenAlex and the papers' affiliations).
- Microsoft · 3 papers
- Google · 3 papers
- Amazon · 3 papers
- Tsinghua University · 3 papers
- Korea Advanced Institute of Science and Technology · 3 papers
- Columbia University · 2 papers
- Tel Aviv University · 2 papers
- Massachusetts Institute of Technology · 2 papers
- The University of Hong Kong · 2 papers
- Hong Kong University of Science and Technology · 2 papers
Use it
Download BibTeX of 22 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
- RL for LLM reasoning (187 papers)
- Scaling RL with verifiable rewards (119 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).