Research map / Reinforcement learning / RL from feedback
Scaling RL with verifiable rewards
Study scaling, exploration, and learning dynamics of RL with verifiable rewards.
119 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 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 117 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 0.0% to 18.9%.
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 · Learning to Generalize from Sparse and Underspecified Rewards (ICML 2019 · 69 citations)
- 2020 · Closed Loop Neural-Symbolic Learning via Integrating Neural Perception, Grammar Parsing, and Symbolic Reasoning (ICML 2020 · 11 citations)
- 2025 · Learning to Reason under Off-Policy Guidance (NeurIPS 2025 · 6 citations)
- 2026 · MedGR2: Breaking the Data Barrier for Medical Reasoning via Generative Reward Learning (AAAI 2026 · 1 citation)
Earliest papers here
- Learning to Generalize from Sparse and Underspecified Rewards (ICML 2019 · 69 citations)
- Closed Loop Neural-Symbolic Learning via Integrating Neural Perception, Grammar Parsing, and Symbolic Reasoning (ICML 2020 · 11 citations)
- Learning to Reason under Off-Policy Guidance (NeurIPS 2025 · 6 citations)
What has been tried
16 papers here report 211 measured changes (a component added, removed or swapped, compared with a baseline): helped 78%, no clear effect 3%, hurt 17%, mixed 2%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| add policy shaping to Mixed-Policy RL | In-distribution math benchmarks · average score | 44.4 -> 47.8 | Helped | Learning to Reason under Off-Policy Guidance (NeurIPS 2025 · 6 citations) |
| Set GAE lambda to 1.0 rather than 0.95 | training reward · reward | Helped | Open-Reasoner-Zero: An Open Source Approach to Scaling Up Reinforcement Learning on the Base Model (NeurIPS 2025 · 5 citations) | |
| replace response-length normalization with Max_Tokens | ScreenSpot · accuracy (Avg.) | 82.3 -> 83.2 | Helped | GUI-G1: Understanding R1-Zero-Like Training for Visual Grounding in GUI Agents (NeurIPS 2025 · 4 citations) |
| Replace summation-form with min-form credit assignment for PRM rewards | Qwen2.5-Math series, training · training stability | sum-form ... collapse at step 25; min-form methods remain stable over 200 steps | Helped | Stop Summation: Min-Form Credit Assignment Is All Process Reward Model Needs for Reasoning (NeurIPS 2025 · 3 citations) |
| prefix exposure (25/50/75% of expert trace) instead of full trace | Math 1k, six benchmarks · avg accuracy | +6.1 / +3.7 / +5.1 over full trace; all below SMEPO | Mixed | Hide to Guide: Learning via Semantic Masking (NeurIPS 2026 · 0 citations) |
| Add learned-verifier selection when choosing training rollouts | MBPP · BoN accuracy | 46.7 -> 49.0 | Helped | Multi-Turn Code Generation Through Single-Step Rewards (ICML 2025 · 0 citations) |
| increase optimization horizon above 1 | Camera Dropbox · reward-hacking behavior | Hurt | MONA: Myopic Optimization with Non-myopic Approval Can Mitigate Multi-step Reward Hacking (ICML 2025 · 0 citations) | |
| Remove the ExP-SFT term (β=0) | MATH training set (1 epoch) · accuracy | 68.7% -> 51.6%; 58.9% -> 44.3% | Hurt | ExPO: Unlocking Hard Reasoning with Self-Explanation-Guided Reinforcement Learning (NeurIPS 2025 · 0 citations) |
| Train on all later prefixes rather than only the current prefix | GSM8K validation · pass@1 | 80.5 -> 82.3 | Helped | $Q\sharp$: Provably Optimal Distributional RL for LLM Post-Training (NeurIPS 2025 · 0 citations) |
| train the UPM with AdaLN while freezing LLaDA | GSM8K · accuracy | 47.27 -> 69.24 | Helped | Reinforcing the Diffusion Chain of Lateral Thought with Diffusion Language Models (NeurIPS 2025 · 0 citations) |
| Add verifiable dense rewards to the thinking-policy objective | Mathematical reasoning tasks · anytime accuracy | Helped | Optimizing Anytime Reasoning via Budget Relative Policy Optimization (NeurIPS 2025 · 0 citations) | |
| Increase the JEPO sample size n | MATH · training performance at a fixed KL budget | Helped | Beyond Verifiable Rewards: Scaling Reinforcement Learning in Language Models to Unverifiable Data (NeurIPS 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 44 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 · 9 of 44 papers
- GSM8K · 7 of 44 papers
- NuminaMath · 3 of 44 papers
- OpenR1-Math-220k · 3 of 44 papers
- DAPO-MATH-17k · 3 of 44 papers
- AIME24 · 2 of 44 papers
- MATH500 · 2 of 44 papers
- NuminaMath 1.5 · 2 of 44 papers
Baselines
- GRPO · 11 of 44 papers
- SFT · 5 of 44 papers
- STaR · 2 of 44 papers
- RISE · 2 of 44 papers
- SFT+RL · 2 of 44 papers
- DAPO · 2 of 44 papers
- LUFFY · 2 of 44 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 117 papers; from OpenAlex and the papers' affiliations).
- Google DeepMind · 3 papers
- Meta · 3 papers
- Massachusetts Institute of Technology · 2 papers
- Cornell University · 2 papers
- Chinese Academy of Sciences · 2 papers
- University of Chinese Academy of Sciences · 2 papers
- Shanghai AI Laboratory · 2 papers
- Chinese academy of science · 2 papers
- Megvii · 2 papers
- StepFun · 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)
- 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).