Research map / Vision-language and multimodal / Multimodal LLMs

Revisual-R1: Advancing Multimodal Reasoning From Optimized Cold Start to Staged Reinforcement Learning

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning 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.

Proposes ReVisual-R1, a three-stage recipe (text-only cold start, multimodal GRPO with Prioritized Advantage Distillation, then text-only RL) that yields open 3B/7B multimodal reasoning models with strong math benchmark results.

Problem

Multimodal LLMs struggle to activate complex long-chain reasoning when text-style RL (GRPO/R1-style) is applied directly, with weak cold-start data and gradient stagnation from zero-advantage groups.

Approach

Uses a text-only, high-difficulty cold start (283K text samples) instead of multimodal SFT, a Prioritized Advantage Distillation (PAD) filter-and-resample step on top of GRPO to counter gradient stagnation, and a final text-only RL stage to polish multimodal reasoning.

Key results

• ReVisual-R1-7B averages 53.1% across the main table, reaching 89.2% on MATH500 and 53.3% on AIME24, the top open-source 7B result in nine of ten benchmarks.
• Ablation: CS+MRL+TRL achieves 49.6 average versus 47.1 for cold start alone and 45.5 for the reversed CS+TRL+MRL order.
• PAD improves the CS+MRL average from 45.1 (GRPO baseline) to 47.7, above DAPO at 46.1.

Surprising findings

• Text-only cold-start data yields larger gains on multimodal benchmarks than multimodal cold-start datasets such as Vision-R1 and R1-One-Vision.
• Mixed multimodal and text RL performs worse than the staged MRL-then-TRL ordering.

Limitations

• No theoretical foundation for why text-centric cold start and staged RL work.
• Scalability evidence limited to 3B and 7B Qwen2.5-VL models; MoE and other model families untested.
• Data-type and ratio interplay between multimodal and text data not systematically explored.

Experimental setup

Datasets: GRAMMAR Vision-R1 R1-One-Vision DeepMath OpenR1-Math Big-Math-RL OpenThoughts MAVIS FigureQA OpenMathR

Benchmarks: MathVerse MathVision MathVista DynaMath WeMath LogicVista AIME24 AIME25 GPQA MATH500 MMMU MMMU-Pro

Metrics: pass@1 accuracy average@32 accuracy

Baselines: VLAA-Thinker-7B OpenVLThinker-7B MM-Eureka-Qwen-7B MMR1-Math-v0 ThinkLite-7B-VL VL-Rethinker-7B FAST-3B VLAA-Thinker-3B doubao-1.5-vision-pro-32k GPT-4o Claude-3.7-Sonnet Gemini-2.0-Flash

Builds on: GRPO DeepSeek-R1 DAPO

Closest papers by meaning

• Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models (ICLR 2026 · similarity 0.94)
• Stable and Efficient Single-Rollout RL for Multimodal Reasoning (CVPR 2026 · similarity 0.92)
• Diving into Self-Evolving Training for Multimodal Reasoning (ICML 2025 · similarity 0.91)
• Perception-R1: Advancing Multimodal Reasoning Capabilities of MLLMs via Visual Perception Reward (ICLR 2026 · similarity 0.91)
• Lavida-R1: Advancing Reasoning for Unified Multimodal Diffusion Language Models (ICML 2026 · similarity 0.91)
• SRPO: Enhancing Multimodal LLM Reasoning via Reflection-Aware Reinforcement Learning (NeurIPS 2025 · similarity 0.91)