Research map / Vision-language and multimodal / Multimodal LLMs
Open Vision Reasoner: Transferring Linguistic Cognitive Behavior for Visual Reasoning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Transfers linguistic cognitive behaviors to a Qwen2.5-VL-7B MLLM via large language-only cold-start SFT followed by ~900-step multimodal PPO, producing Open-Vision-Reasoner (OVR) with state-of-the-art 7B visual math reasoning.
Problem
How linguistic cognitive behaviors (backtracking, subgoal setting, verification, backward chaining) transfer to multimodal LLMs and how they emerge and scale for visual reasoning, given that prior multimodal RL work relied on learned reward models or small-scale training.
Approach
A two-stage pipeline: a massive text-only cold start distilled from DeepSeek-R1 on the LLM module of Qwen2.5-VL-7B, then large-scale PPO with binary rule-based match rewards on mixed text and multimodal data under the Open-Reasoner-Zero setup, plus a systematic analysis of visual cognitive behavior emergence and transfer rates.
Key results
- OVR-7B reaches 95.3% on MATH500, 63.5% on AIME 2024 and 52.1% on AIME 2025 among open 7B models in the language comparison.
- OVR-7B achieves 51.8% on MathVision and 54.6% on MathVerse, the first post-trained Qwen2.5-VL-7B-based model to exceed 50% on MathVision.
- Backtracking transfer rate from language to vision grows from 2.5% to 17.3% across training, while verification shows near-zero transfer.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add linguistic cold-start fine-tuning followed by multimodal RL | Qwen2.5-VL-7B · MathVision | score: 25.5 -> 51.8 | Helped |
| Add linguistic cold-start fine-tuning followed by multimodal RL | Qwen2.5-VL-7B · MathVerse vision-only | score: 41.1 -> 54.6 | Helped |
| Add linguistic cold-start fine-tuning followed by multimodal RL | Qwen2.5-VL-7B · MATH500 | score: 67.4 -> 95.3 | Helped |
| Add linguistic cold-start fine-tuning followed by multimodal RL | Qwen2.5-VL-7B · AIME 2024 | score: 6.7 -> 63.5 | Helped |
| Add linguistic cold-start fine-tuning followed by multimodal RL | Qwen2.5-VL-7B · MMMU-Pro | score: +7.2% over prior SOTA methods | Helped |
| Add linguistic cold-start fine-tuning (ablation) | Qwen2.5-VL · MMBench en | score: 85.3 -> 86.1 | Helped |
| Add linguistic cold-start fine-tuning (ablation) | Qwen2.5-VL · MMBench cn | score: 84.6 -> 82.1 | Hurt |
| Add linguistic cold-start fine-tuning (ablation) | Qwen2.5-VL · BLINK | score: 53.7 -> 51.5 | Hurt |
| Add linguistic cold-start fine-tuning (ablation) | Qwen2.5-VL · MMStar | score: 62.1 -> 62.4 | Helped |
| Add linguistic cold-start fine-tuning (ablation) | Qwen2.5-VL · HallusionBench | score: 49.0 -> 55.0 | Hurt |
| Add linguistic cold-start fine-tuning (ablation) | Qwen2.5-VL · POPE | score: 86.1 -> 82.5 | Hurt |
| Add multimodal RL after cold start (ablation) | Qwen2.5-VL + Cold Start · BLINK | score: 51.5 -> 54.1 | Helped |
Surprising findings
- Cold start causes perception degradation (e.g., MMBench cn 84.6 to 82.1 and RealWorldQA 69.3 to 63.1), which RL partially recovers.
- Visual reflection behaviors appear in large numbers from the very start of cold-start, and RL first suppresses then amplifies behavior prevalence.
Limitations
- RL on perception-focused tasks such as OCR and counting shows limited scalability, with response length stagnating while reward rises.
- Verification behavior transfers near zero, and the explanations for transfer patterns are untested hypotheses.
- Final model is a uniform average of intermediate checkpoints, and the cold start degrades some perception benchmarks.
Experimental setup
Datasets: AIME MATH Numina-Math Tulu3 MATH OpenR1-Math-220k Geometry3k GeoQA IconQA ChartQA ScienceQA PuzzleVQA AlgoPuzzleVQA
Benchmarks: AIME 2024 AIME 2025 MATH500 GPQA Diamond MMLU MMLU-Pro MathVista MathVision MathVerse DynaMath WeMath LogicVista
Metrics: accuracy behavior transfer rate
Baselines: Qwen2.5-VL-7B Open-Reasoner-Zero-7B DeepSeek-R1-Distill-Qwen-7B ReVisual-R1 VL-Rethinker-7B MM-Eureka-Qwen-7B OpenVLThinker-7B VLAA-Thinker-Qwen2.5-7B
Builds on: Open-Reasoner-Zero PPO GAE RL with a cold start Qwen2.5-VL
Closest papers by meaning
- Reason-RFT: Reinforcement Fine-Tuning for Visual Reasoning of Vision Language Models (NeurIPS 2025 · similarity 0.93)
- OpenVLThinker: Complex Vision-Language Reasoning via Iterative SFT-RL Cycles (NeurIPS 2025 · similarity 0.92)
- VisuRiddles: Fine-grained Perception is a Primary Bottleneck for Multimodal Large Language Models in Abstract Visual Reasoning (ICLR 2026 · similarity 0.92)
- MMFineReason: Closing the Multimodal Reasoning Gap via Open Data-Centric Methods (NeurIPS 2026 · similarity 0.92)
- Machine Mental Imagery: Empower Multimodal Reasoning with Latent Visual Tokens (CVPR 2026 · similarity 0.91)
- More Thought, Less Accuracy? On the Dual Nature of Reasoning in Vision-Language Models (ICLR 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.