Research map / Vision-language and multimodal / Multimodal LLMs
Reinforcement learning for visual reasoning
Trains vision-language models with reinforcement learning and data augmentation to reason visually.
163 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2024–2026) · an approach in Multimodal LLMs, keyword cluster “vlms · visual reasoning · chain”
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 162 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 0.5% to 7.5%.
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
- 2024 · Fine-Tuning Large Vision-Language Models as Decision-Making Agents via Reinforcement Learning (NeurIPS 2024 · 21 citations)
- 2025 · JarvisIR: Elevating Autonomous Driving Perception with Intelligent Image Restoration (CVPR 2025 · 24 citations)
- 2026 · MedEyes: Learning Dynamic Visual Focus for Medical Progressive Diagnosis (AAAI 2026 · 7 citations)
Earliest papers here
- Fine-Tuning Large Vision-Language Models as Decision-Making Agents via Reinforcement Learning (NeurIPS 2024 · 21 citations)
- JarvisIR: Elevating Autonomous Driving Perception with Intelligent Image Restoration (CVPR 2025 · 24 citations)
- Critic-V: VLM Critics Help Catch VLM Errors in Multimodal Reasoning (CVPR 2025 · 6 citations)
What has been tried
31 papers here report 378 measured changes (a component added, removed or swapped, compared with a baseline): helped 65%, no clear effect 6%, hurt 26%, 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 entropy regularization | CleanBench-Real validation · Reward | 0.50 -> 0.67 | Helped | JarvisIR: Elevating Autonomous Driving Perception with Intelligent Image Restoration (CVPR 2025 · 24 citations) |
| add CoT reasoning | gym_cards / NumberLine · episode success rate (%) | 26.9 → 89.4 | Helped | Fine-Tuning Large Vision-Language Models as Decision-Making Agents via Reinforcement Learning (NeurIPS 2024 · 21 citations) |
| Replace Critic-V with Self-Refine (no DPO-trained critic) | MathVista · score (%) | 61.4 -> 63.4 | Helped | Critic-V: VLM Critics Help Catch VLM Errors in Multimodal Reasoning (CVPR 2025 · 6 citations) |
| Add RL training after warm-start instruction tuning | MVBench-test · accuracy | 59.0 -> 67.8 | Helped | Pixel Reasoner: Incentivizing Pixel Space Reasoning via Curiosity-Driven Reinforcement Learning (NeurIPS 2025 · 6 citations) |
| Add GRPO refinement | SAT-2 · accuracy | 58.83 { (±1.53)} -> 62.93 { (±1.50)}; (–4.10%) | Helped | Grounded Reinforcement Learning for Visual Reasoning (NeurIPS 2025 · 5 citations) |
| Replace fixed thresholds with tiered segmentation-accuracy thresholds | ReasonSeg-zero-shot test · gIoU | 58.6 -> 60.2 | Helped | SAM-R1: Leveraging SAM for Reward Feedback in Multimodal Segmentation via Reinforcement Learning (NeurIPS 2025 · 5 citations) |
| Add visual-perception actions to vanilla reward and search actions | Three benchmarks · accuracy | 47.2 -> 49.3 | Helped | VRAG-RL: Empower Vision-Perception-Based RAG for Visually Rich Information Understanding via Iterative Reasoning with Reinforcement Learning (NeurIPS 2025 · 4 citations) |
| Add CoT-SFT reasoning activation before RL | Spatial Transformation · accuracy | 42.13 vs. 74.61 | Helped | Reason-RFT: Reinforcement Fine-Tuning for Visual Reasoning of Vision Language Models (NeurIPS 2025 · 4 citations) |
| Combine unsolved samples with samples requiring more than five iterations | 8 visual benchmarks average · Avg. score | 62.04 -> 63.89 | Helped | SoTA with Less: MCTS-Guided Sample Selection for Data-Efficient Visual Reasoning Self-Improvement (NeurIPS 2025 · 4 citations) |
| Increase independent judgments from K=1 to K=7 | VL-RewardBench · Macro accuracy | 60.3% to 62.7% | Helped | VL-RewardBench: A Challenging Benchmark for Vision-Language Generative Reward Models (CVPR 2025 · 3 citations) |
| Add data sampling | MathVista, MathVision, MathVerse · accuracy | 69.9 -> 73.8; 27.2 -> 30.6; 48.4 -> 50.6 | Helped | Fast-Slow Thinking GRPO for Large Vision-Language Model Reasoning (NeurIPS 2025 · 3 citations) |
| Remove the vision error codebook | LLaVA-150k subset · MME | 1533 -> 1505 | Hurt | Continual SFT Matches Multimodal RLHF with Negative Supervision (CVPR 2025 · 2 citations) |
What reviewers will expect
Datasets and baselines named in the 162 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- Geometry3K · 21 of 162 papers
- ViRL39K · 14 of 162 papers
- MathVista · 14 of 162 papers
- MathVerse · 13 of 162 papers
- ChartQA · 12 of 162 papers
- MathVision · 10 of 162 papers
- WeMath · 8 of 162 papers
- RefCOCOg · 7 of 162 papers
Baselines
- GPT-4o · 42 of 162 papers
- Qwen2.5-VL-7B · 36 of 162 papers
- GRPO · 19 of 162 papers
- Qwen2.5-VL-3B · 15 of 162 papers
- Gemini-2.5-Pro · 14 of 162 papers
- DAPO · 12 of 162 papers
- OpenVLThinker-7B · 12 of 162 papers
- Vision-R1-7B · 12 of 162 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 163 papers; from OpenAlex and the papers' affiliations).
- The Chinese University of Hong Kong · 7 papers
- Peking University · 7 papers
- Alibaba Group · 6 papers
- Tsinghua University · 6 papers
- Nanjing University · 5 papers
- Nanyang Technological University · 5 papers
- University of Illinois, Urbana Champaign · 5 papers
- University of Maryland, College Park · 5 papers
- Microsoft · 5 papers
- Hong Kong University of Science and Technology · 5 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
- VLM capability analysis (264 papers)
- Chain-of-thought multimodal reasoning (332 papers)
- Multimodal LLM benchmarks (252 papers)
- Spatial reasoning in VLMs (207 papers)
Elsewhere in the topic
- Instruction-tuned multimodal assistants (199 papers)
- Benchmarks and analysis of MLLMs (181 papers)
- Interleaved image-text generation with MLLMs (153 papers)
- Unified any-modality language models (152 papers)
- Attention-based visual token pruning (117 papers)
- Decoding-time hallucination mitigation (91 papers)
- Training-free token reduction for MLLMs (90 papers)
- Visual instruction tuning (68 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).