Research map / Reinforcement learning
RL from feedback: research map
758 papers on RL from feedback in Reinforcement learning, from ICML, NeurIPS, ICLR, CVPR, AAAI, ICCV and ECCV, and the journals TPAMI, IJCV and JMLR (2016–2026), grouped into 2 clusters and 7 approaches. The busiest year so far is 2026.
What it is
Training models with reinforcement learning from human preferences or automatic checks rather than a hand-written reward. This is how chat models are aligned (RLHF) and, more recently, how reasoning models are trained with verifiable rewards, for example GRPO on math problems with checkable answers.
Terms you will see on this page:
- RLHF
- reinforcement learning from human feedback, usually via a reward model trained on preferences.
- Verifiable reward
- a reward computed automatically, such as whether a math answer is correct.
- GRPO
- an RL algorithm that compares a group of sampled answers to the same prompt instead of using a critic.
Start reading with the most cited papers here:
- Deep Reinforcement Learning from Human Preferences (NeurIPS 2017 · 500 citations)
- Interactive Learning from Policy-Dependent Human Feedback (ICML 2017 · 104 citations)
Within Reinforcement learning, its share grew from 5.8% in 2023–24 to 26.8% in 2025–26 (91 → 620 papers at ICML, NeurIPS, ICLR, CVPR and AAAI, the venues with data for all four years).
Explore RL from feedback in the interactive map
Working on something in this topic? Describe your idea in scime atlas to see which approach it falls under, the closest papers by meaning and how crowded the spot has become.
What changed recently
Approaches whose share of the topic moved most between 2023–24 and 2025–26 (ICML, NeurIPS, ICLR, CVPR and AAAI):
- Scaling RL with verifiable rewards rising: 0% → 19% of the topic (0 → 117 papers)
- GRPO stability and variants rising: 1% → 17% of the topic (1 → 103 papers)
- RL for LLM reasoning rising: 14% → 28% of the topic (13 → 171 papers)
- Theory of RLHF falling: 44% → 10% of the topic (40 → 61 papers)
- Reward modeling versus direct optimization falling: 22% → 11% of the topic (20 → 71 papers)
Papers from 2025–2026 that are already cited most:
- Flow-GRPO: Training Flow Matching Models via Online RL (NeurIPS 2025 · 36 citations)
- SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training (ICML 2025 · 6 citations)
- Learning to Reason under Off-Policy Guidance (NeurIPS 2025 · 6 citations)
- Open-Reasoner-Zero: An Open Source Approach to Scaling Up Reinforcement Learning on the Base Model (NeurIPS 2025 · 5 citations)
Approaches and key papers
RL for LLM reasoning · 524 papers
Keywords: reasoning, grpo, llm
Approaches in this cluster:
- RL for LLM reasoning (188 papers)
Apply reinforcement learning to improve multi-step reasoning in language models. - Scaling RL with verifiable rewards (121 papers)
Study scaling, exploration, and learning dynamics of RL with verifiable rewards. - GRPO stability and variants (108 papers)
Analyze and fix group-relative policy optimization issues such as noise, collapse, and trust regions. - RL fine-tuning of generative models (107 papers)
Fine-tune diffusion and multimodal generative models with policy-gradient and reward objectives.
Most cited and most cited since 2024:
- Sequence Tutor: Conservative Fine-Tuning of Sequence Generation Models with KL-control (ICML 2017 · 73 citations)
- Learning to Generalize from Sparse and Underspecified Rewards (ICML 2019 · 69 citations)
- Flow-GRPO: Training Flow Matching Models via Online RL (NeurIPS 2025 · 36 citations)
- Auto MC-Reward: Automated Dense Reward Design with Large Language Models for Minecraft (CVPR 2024 · 25 citations)
Learning from human preferences · 234 papers
Keywords: human, rlhf, preferences
Approaches in this cluster:
- Theory of RLHF (112 papers)
Analyze sample efficiency and reward-model behavior in reinforcement learning from human feedback. - Reward modeling versus direct optimization (95 papers)
Compare reward-model RLHF with DPO-style direct policy optimization and refine preference objectives. - Preference-based RL reward learning (27 papers)
Learn rewards from pairwise preferences with improved queries, exploration, and reward-agnostic methods.
Most cited and most cited since 2024:
- Deep Reinforcement Learning from Human Preferences (NeurIPS 2017 · 500 citations)
- Interactive Learning from Policy-Dependent Human Feedback (ICML 2017 · 104 citations)
- Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model (CVPR 2024 · 49 citations)
- Towards Automated RISC-V Microarchitecture Design with Reinforcement Learning (AAAI 2024 · 13 citations)
Related topics in Reinforcement learning
- MDPs and planning (726)
- Visual and representation RL (2,264)
- Policy gradient and actor-critic (1,570)
- Imitation learning (450)
- Offline RL (646)
