Research map / Vision-language and multimodal / Multimodal LLMs

Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 1 citations (OpenAlex)

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.

Builds a 200K multimodal CoT cold-start dataset via modality bridging (Qwen2.5-VL-72B + DeepSeek-R1), then applies SFT followed by GRPO with a progressive thinking-length suppression schedule (PTST) and hard format-result reward to train Vision-R1 for multimodal math reasoning.

Problem

Direct RL on MLLMs fails to elicit complex reflective reasoning, and cold-start SFT-then-RL produces overthinking that hinders optimization on multimodal math tasks.

Approach

Modality bridging converts images to text descriptions (via a Pseudo-CoT-guided MLLM pass) so text-only DeepSeek-R1 can write human-like CoT for multimodal VQA, and PTST progressively relaxes a response-length cap during GRPO (4K then 8K) so the model first learns correct short reasoning then extends it.

Key results

• Vision-R1-7B reaches 73.5% on MathVista, 0.4 points below OpenAI O1 (73.9%), and improves the Qwen2.5-VL-7B base by about 5.7 points average over math benchmarks.
• Vision-R1-72B reports 78.2% on MathVista, and Vision-R1-32B reports 76.4%.
• In ablation, PTST with a 4K-then-8K schedule reaches 55.4% average versus 47.7% for fixed 16K RL (Vision-R1-Long) and 50.7% for RL-only (Vision-R1-Zero).

Surprising findings

• Cold-start SFT without CoT annotations followed by PTST drops average accuracy to 39.8%, well below RL-only training.
• Training with fixed 16K response length severely underperforms, indicating that unconstrained longer reasoning hurts accuracy.

Limitations

• Scaling to 32B and 72B relies on additional RL data (~20K), so gains are not from the 10K setup alone.
• Evaluation is centered on math benchmarks; general-capability gains are assessed only for the Llama-3.2-11B cold-start variant.

Experimental setup

Datasets: LLaVA-CoT (100K) Mulberry (260K) We-Math MathVision PolyMATH SceMQA Geometry3K MAmmoTH-VL MMIQ

Benchmarks: MathVista MathVerse MM-Math DynaMath MMStar ChartQA MME HallBench

Metrics: accuracy average accuracy

Baselines: Qwen2.5-VL-7B Qwen2.5-VL-32B Qwen2.5-VL-72B InternVL2.5-78B LLaVA-CoT-11B Mulberry-7B OpenAI O1 GPT-4o

Builds on: DeepSeek-R1-Zero GRPO Qwen2.5-VL LLaVA-CoT Mulberry

Closest papers by meaning

• Revisual-R1: Advancing Multimodal Reasoning From Optimized Cold Start to Staged Reinforcement Learning (ICLR 2026 · similarity 0.94)
• Perception-R1: Advancing Multimodal Reasoning Capabilities of MLLMs via Visual Perception Reward (ICLR 2026 · similarity 0.93)
• Incentivizing Versatile Video Reasoning in MLLMs via Data-Efficient Reinforcement Learning (CVPR 2026 · similarity 0.91)
• Stable and Efficient Single-Rollout RL for Multimodal Reasoning (CVPR 2026 · similarity 0.91)
• Learn to Think: Improving Multimodal Reasoning through Vision-Aware Self-Improvement Training (ICML 2026 · similarity 0.90)
• Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.90)