Research map / Vision-language and multimodal / Multimodal LLMs

From Seeing to Thinking: Decoupling Perception and Reasoning Improves Post-Training of Vision-Language Models

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual 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 a three-stage post-training pipeline (visual perception via RLVR, textual reasoning, visual reasoning via GRPO) that outperforms merged training on Qwen2.5-VL-7B and Qwen3-VL-8B and yields shorter reasoning traces.

Problem

VLM post-training for visual reasoning under-emphasizes visual perception; failures often originate from perception errors that longer reasoning does not correct.

Approach

Decouple VLM post-training into staged capability-specific training (perception data built from DOCCI captions with caption-vs-image filtering, then text math, then visual math), trained with GRPO-style RLVR, and frame the staging as a capability-dimension curriculum orthogonal to difficulty curricula.

Key results

• Staged Qwen3-VL-8B reaches 65.77 overall average vs 64.67 for merged training and 62.19 for base, with 1.46-point gain in math accuracy and 20.8% shorter validation responses (445 vs 562 tokens).
• Staged Qwen2.5-VL-7B improves WeMath from 30.86 (base) to 38.29 and reaches 59.75 overall average vs 56.68 for base.
• Replacing RLVR with caption-based SFT in Stage 1 drops Qwen2.5-VL-7B WeMath accuracy from 38.29 to 30.10 (about 8 points).

Surprising findings

• Reversing stage order to perception-last (3→2→1) degrades Qwen2.5-VL-7B visual math average from 42.3 to 37.7, below the base model's 37.0 only in part; perception-first and text-first orders perform comparably.
• Reasoning-only post-training reduces MMStar by 1.6 points on Qwen2.5-VL-7B, a 'perceptual tax' that perception data mitigates.

Limitations

• Experiments limited to 7-8B scale; no validation at 32B+.
• Perception data pipeline depends on availability of fine-grained captions.
• Three-stage decomposition may not capture finest capability granularity.

Experimental setup

Datasets: DOCCI ORZ-Math-13k CLEVR-Math GeoQA170K Math PUMA DocVQA ArxivQA

Benchmarks: MathVista MINI MathVision MINI MathVerse Vision Intensive WeMath A-OKVQA RealWorldQA MMStar POPE MathVerse Vision Only DynaMath CV-Bench V*Bench

Metrics: accuracy average response length

Baselines: Merged training GThinker-7B MMR1-7B OpenVLThinker-7B R1-OneVision-RL-7B WeThink-7B OneThinker-8B

Builds on: GRPO DeepSeek-R1 EasyR1 ORZ-Math-13k DOCCI

Closest papers by meaning

• Bring Reason to Vision: Understanding Perception and Reasoning through Model Merging (ICML 2025 · similarity 0.93)
• Learning to See Before Seeing: Demystifying LLM Visual Priors from Language Pre-training (ICLR 2026 · similarity 0.93)
• Caption This, Reason That: VLMs Caught in the Middle (NeurIPS 2025 · similarity 0.92)
• SPARC: Separating Perception And Reasoning Circuits for Test-time Scaling of VLMs (ICML 2026 · similarity 0.92)
• Through the Lens of Contrast: Self-Improving Visual Reasoning in VLMs (ICLR 2026 · similarity 0.92)
• When to Think and When to Look: Uncertainty-Guided Lookback (CVPR 2026 · similarity 0.92)