Research map / Vision-language and multimodal / Multimodal LLMs

Reasoning-Aligned Perception Decoupling for Scalable Multi-modal Reasoning

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 0 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.

Proposes RAPID, which decouples an MLLM (perception: query-relevant caption plus tentative solution) from a frozen text-only LLM reasoner, and trains the MLLM with VPO, a GRPO-style RL objective whose reward is the correctness of the external reasoner's answer, enabling plug-and-play reasoner swapping at inference.

Problem

MLLMs lag on multi-modal math and science reasoning because their internal LLMs are outdated, and swapping in a stronger LLM requires costly vision-language realignment. Caption-then-reason pipelines exist but their captions are not optimized for final answer correctness.

Approach

Decouple perception from reasoning so a frozen external text-only LLM does the reasoning, and optimize the MLLM's query-relevant captions with RL using the reasoner's answer correctness as reward (VPO), with a caption-presence penalty to prevent reward hacking by solving directly.

Key results

• Qwen2.5-VL-7B with RAPID (Qwen3-8B reasoner) reaches 53.2 average accuracy across seven benchmarks, up from 42.0 for the original MLLM (+11.2).
• Qwen2.5-VL-72B with RAPID (GPT-OSS-120B reasoner) reaches 58.0 average, reported as the best among compared models.
• Ablation on 7B: decoupling adds 5.5 points, GRPO adds 3.0, VPO adds 1.7 without caption penalty and 1.0 more with it.

Surprising findings

• Using the stronger Qwen3-8B as the training-time reward reasoner gave worse results than the weaker R1-7B, attributed to it rewarding shorter captions.
• Decoupling with the same MLLM as both captioner and reasoner still improved over end-to-end use (46.1 vs 44.7 average).

Limitations

• Domain-specific design for math and science reasoning; extension to other domains like spatial reasoning needs architectural adaptation.
• Sequential VPO training causes reasoning forgetting in the MLLM, especially at 3B scale, requiring an extra GRPO stage.
• The caption-presence check has a 26% false positive rate on a 400-sample audit, though the authors argue it has limited downstream impact.

Experimental setup

Datasets: ViRL39K

Benchmarks: MathVista MathVision MathVerse MMMU WeMath DynaMath LogicVista

Metrics: accuracy worst-case accuracy (DynaMath) strict accuracy (WeMath)

Baselines: Claude-3.7-Sonnet Gemini-2.0-Flash GPT-4o InternVL3 Gemma-3 Ovis2 MM-Eureka VL-Rethinker QVQ-72B-Preview ReVisual-R1-7B ECSO OmniCaptioner

Builds on: GRPO Qwen2.5-VL DeepSeek-R1-Distill verl ViRL39K

Closest papers by meaning

• From Seeing to Thinking: Decoupling Perception and Reasoning Improves Post-Training of Vision-Language Models (ICML 2026 · similarity 0.92)
• More Thought, Less Accuracy? On the Dual Nature of Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.92)
• Vision-aligned Latent Reasoning for Multi-modal Large Language Model (ICML 2026 · similarity 0.92)
• Bad Seeing or Bad Thinking? Rewarding Perception for Multimodal Reasoning (ICML 2026 · similarity 0.91)
• Perceptual-Evidence Anchored Reinforced Learning for Multimodal Reasoning (CVPR 2026 · similarity 0.91)
• Perception-R1: Advancing Multimodal Reasoning Capabilities of MLLMs via Visual Perception Reward (ICLR 2026 · similarity 0.91)