Research map / Vision-language and multimodal / Multimodal LLMs
Vero: Open Reinforcement Learning Recipes for Visual Reasoning
Gabriel Sarch, Linrong Cai, Qunzhong Wang, Haoyang Wu, Danqi Chen, Zhuang Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use RL instead of SFT on Vero-600K (ablation) | Qwen2.5-VL-7B-Instruct · VeroEval overall | avg score: 52.8 -> 57.2 | Improved |
| train SFT on Vero-600K instead of FineVision (ablation) | Qwen2.5-VL-7B-Instruct · VeroEval overall | avg score: 46.2 -> 52.8 | Improved |
| use task-routed reward instead of math_verify (ablation) | Qwen2.5-VL-7B-Instruct · VeroEval overall | avg score: 51.8 -> 57.2 | Improved |
| use task-routed reward instead of math_verify (ablation) | Qwen2.5-VL-7B-Instruct · VeroEval Captioning & IF | category avg: 34.3 -> 70.6 | Improved |
| use GSPO instead of GRPO (ablation) | Qwen2.5-VL-7B-Instruct · VeroEval 5 categories | avg score: 54.3 -> 54.7 | Improved |
| use GSPO instead of DAPO for entropy stability (ablation) | Qwen2.5-VL-7B-Instruct · training entropy | entropy: 0.22 -> 0.58 | Improved |
| use uniform task-category weighting instead of difficulty weighting (ablation) | Qwen3-VL-8B-Instruct · VeroEval benchmark avg gain | score change: +5.2 -> +5.8 | Improved |
| include Knowledge & Recognition category in mixture (ablation) | Qwen3-VL-8B-Instruct · VeroEval benchmark avg gain | score change: +4.9 -> +5.8 | Improved |
| add question filtering (ablation) | Qwen2.5-VL-7B-Instruct · Spatial & Action | category avg: 56.3 -> 58.2 | Improved |
| add question filtering (ablation) | Qwen2.5-VL-7B-Instruct · Grounding, Counting & Search | category avg: 54.7 -> 54.1 | Hurt |
| train on diverse task mixture instead of single category (Chart & OCR) (ablation) | Qwen2.5-VL-7B-Instruct · Grounding, Counting & Search | score change: -3.2 -> positive | Improved |
| select data with Vero filtering instead of random subset (ablation) | Qwen2.5-VL-7B-Instruct · STEM | avg benchmark gain: -0.2 -> +4.6 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- OneThinker-8B — OneThinker: All-in-one Reasoning Model for Image and Video (CVPR 2026) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Reason-RFT: Reinforcement Fine-Tuning for Visual Reasoning of Vision Language Models (NeurIPS 2025 · similarity 0.92)
- No Labels, No Problem: Training Visual Reasoners with Multimodal Verifiers (ICLR 2026 · similarity 0.92)
- Open Vision Reasoner: Transferring Linguistic Cognitive Behavior for Visual Reasoning (NeurIPS 2025 · similarity 0.92)
- More Thought, Less Accuracy? On the Dual Nature of Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.92)
- RewardMap: Tackling Sparse Rewards in Fine-grained Visual Reasoning via Multi-Stage Reinforcement Learning (ICLR 2026 · similarity 0.91)
- From Seeing to Thinking: Decoupling Perception and Reasoning Improves Post-Training of Vision-Language Models (ICML 2026 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
