Research map / Vision-language and multimodal / Multimodal LLMs
Seeing is Solving: Unlocking Efficient Multimodal RL via View Alignment
Qinsi Wang, Jing Shi, Kun Wan, Handong Zhao, Hancheng Ye, Zishan Shao, Jinghan Ke, Yudong Liu, Daniel Miranda, Purvak Lapsiya, Yiran Chen, Wentian Zhao
Read the paperCheck your idea against 137,217 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.
Proposes view-alignment-based efficient RL fine-tuning for vision-language models to reduce the training cost of multimodal RLFT.
Problem
RL fine-tuning of VLMs for multimodal reasoning is prohibitively expensive, and existing methods port LLM RLFT techniques without accounting for dynamic text-vision alignment.
Approach
Exploits the dynamic text-vision alignment property of multimodal models to make multimodal RL fine-tuning more efficient, rather than directly porting LLM RLFT techniques. Full method details are not available from the abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add PVA-based difficulty-aware curriculum data processing (ablation) | vanilla GRPO · six reasoning benchmarks, Qwen2.5-VL-7B | average accuracy: 42.7 -> 44.7 | Improved |
| add RVA-based CoT reward on top of data processing (ablation) | FOCUS-RL (Data) · six reasoning benchmarks, Qwen2.5-VL-7B | average accuracy: 44.7 -> 45.3 | Improved |
| use PVA curriculum instead of GainRL difficulty estimate | GRPO + GainRL · six reasoning benchmarks, Qwen2.5-VL-7B | average accuracy: 43.8 -> 44.7 | Improved |
| use PVA curriculum instead of AdaRFT | GRPO + AdaRFT · six reasoning benchmarks, Qwen2.5-VL-7B | average accuracy: 43.1 -> 44.7 | Improved |
| add FOCUS-RL (Data+Policy) to vanilla GRPO | vanilla GRPO · LogicVista, Qwen2.5-VL-7B | accuracy: 44.2 -> 49.1 | Improved |
| add FOCUS-RL to vanilla GRPO | vanilla GRPO · six reasoning benchmarks, Qwen2.5-VL-3B | average accuracy: 33.2 -> 35.3 | Improved |
| add FOCUS-RL to vanilla GRPO | vanilla GRPO · six reasoning benchmarks, InternVL3-2B | average accuracy: 30.9 -> 33.6 | Improved |
| add FOCUS-RL to vanilla GRPO | vanilla GRPO · six reasoning benchmarks, InternVL3-8B | average accuracy: 40.7 -> 42.4 | Improved |
| add FOCUS-RL to GRPO training for speed | vanilla GRPO · MathVista, Qwen2.5-VL-7B | iterations to peak accuracy: 2.3x speedup | Improved |
| add FOCUS-RL to GRPO training for speed | vanilla GRPO · MathVista, InternVL3-8B | iterations to peak accuracy: 4x speedup | Improved |
| add FOCUS-RL (curriculum from easy to hard) to GRPO | vanilla GRPO · Qwen2.5-VL-7B training | completion length stability | 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.
- OpenVLThinker-7B — OpenVLThinker: Complex Vision-Language Reasoning via Iterative SFT-RL Cycles (NeurIPS 2025) · 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
- On Robustness and Chain-of-Thought Consistency of RL-Finetuned VLMs (ICML 2026 · similarity 0.93)
- Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models (ICLR 2026 · similarity 0.92)
- Perception-R1: Advancing Multimodal Reasoning Capabilities of MLLMs via Visual Perception Reward (ICLR 2026 · similarity 0.92)
- Unveiling the Compositional Ability Gap in Vision-Language Reasoning Model (NeurIPS 2025 · similarity 0.92)
- Reason-RFT: Reinforcement Fine-Tuning for Visual Reasoning of Vision Language Models (NeurIPS 2025 · similarity 0.92)
- Beyond SFT-to-RL: Pre-alignment via Black-box On-policy Distillation for Multimodal RL (NeurIPS 2026 · similarity 0.92)
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.
