Research map / Vision-language and multimodal / Multimodal LLMs

More Thought, Less Accuracy? On the Dual Nature of Reasoning in Vision-Language Models

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.

Shows that longer reasoning in VLMs causes visual forgetting and perception errors, and proposes VAPO, an RL method that inserts visual-claim anchors into reasoning and adds a perception reward, yielding VAPO-Thinker-7B with state-of-the-art results.

Problem

Multimodal reasoning in VLMs can degrade perceptual grounding as reasoning gets longer, causing recognition failures on basic visual questions (visual forgetting).

Approach

Insert visual-claim anchors (GPT-5-generated true/false statements about the image) at random points in the reasoning trace, have the model judge each claim, and add a late-emphasis weighted perception reward to GRPO, gated on answer correctness.

Key results

• VAPO-Thinker-7B reaches 51.1 average on math benchmarks vs 49.1 for Vision-R1-7B, and 63.1 average on general benchmarks vs 59.9 for prior best.
• Visual replay alleviates degradation, outperforming vanilla reasoning by around 1.5% in the analysis.
• On Qwen3-VL-8B, VAPO improves average accuracy from 64.9 (GRPO) to 68.8.

Surprising findings

• Over 50% of failures under full reasoning are perception errors, and about a third of those are recoverable via early decision.
• Augmenting GRPO with the same visual claims as extra training examples yields only marginal gains (49.0 to 49.3), suggesting the gain comes from the reward mechanism rather than data augmentation.

Limitations

• Quality of GPT-5-generated visual claims bounds performance; claim generation relies on a closed-source model.
• Introduces several new hyperparameters (K, beta, gamma) that need tuning and vary by task type.
• Evaluated only on single-image tasks.

Experimental setup

Datasets: ViRL39K MathVerse MathVista MathVision LogicVista WeMath Geometry3k MMMU MMStar HallusionBench MMVet V*

Benchmarks: MathVerse MathVista MathVision LogicVista WeMath Geometry3k MMMU MMStar HallusionBench MMVet V* RealWorldQA

Metrics: accuracy

Baselines: GPT-5-Thinking Gemini-2.5-Pro InternVL2.5-8B R1-OneVision-7B VLAA-Thinker-7B Vision-R1-7B DAPO GRPO

Builds on: GRPO DAPO Qwen2.5-VL Vision-R1 ViRL39K

Closest papers by meaning

• Reason-RFT: Reinforcement Fine-Tuning for Visual Reasoning of Vision Language Models (NeurIPS 2025 · similarity 0.93)
• Perception-Aware Policy Optimization for Multimodal Reasoning (ICLR 2026 · similarity 0.92)
• MindOmni: Unleashing Reasoning Generation in Vision Language Models with RGPO (NeurIPS 2025 · similarity 0.92)
• Semi-off-Policy Reinforcement Learning for Vision-Language Slow-Thinking Reasoning (NeurIPS 2025 · similarity 0.92)
• Activating Visual Context and Commonsense Reasoning Through Masked Prediction in VLMs (AAAI 2026 · similarity 0.92)
• Think or Not? Selective Reasoning via Reinforcement Learning for Vision-Language Models (NeurIPS 2025 · similarity 0.92)