Paper 2509.25848
More Thought, Less Accuracy? On the Dual Nature of Reasoning in Vision-Language Models
- Published
- Sep 2025
- Research lab
- Independent
- Citations
- 26
- GitHub
- 389 stars
01 In brief
Summary
This paper investigates the dual nature of reasoning in Vision-Language Models (VLMs).
The authors find that while reasoning improves logical inference, longer reasoning lengths can impair perceptual grounding, leading to errors on basic visual questions.
They attribute this to 'visual forgetting,' where models increasingly ignore visual input during extended reasoning.
To address this, they propose Vision-Anchored Policy Optimization (VAPO), which inserts visual claims as anchors during reasoning and adds a perception reward to encourage visually grounded trajectories.
Their model, VAPO-Thinker-7B, achieves state-of-the-art results on multiple benchmarks, improving average accuracy by 2-4% over strong baselines.
02 From the paper
Abstract
Reasoning has emerged as a pivotal capability in Large Language Models (LLMs). Through Reinforcement Learning (RL), typically Group Relative Policy Optimization (GRPO), these models are able to solve complex tasks such as mathematics and code generation. Building on these advances, recent research has sought to extend reasoning to Vision-Language Models (VLMs), yielding promising results across diverse visual tasks. Despite this progress, our study uncovers the dual nature of multimodal reasoning: while it substantially enhances logical inference and facilitates performance on challenging problems, it may gradually impair perceptual grounding, leading to recognition failures on otherwise basic visual questions. Through further analysis, we attribute this phenomenon to visual forgetting, wherein prolonged reasoning causes the model to increasingly disregard visual input. To address this, we propose Vision-Anchored Policy Optimization (VAPO), a simple yet effective method that explicitly steers the reasoning process toward visually grounded trajectories. Our result model, VAPO-Thinker-7B, significantly strengthens the model's reliance on visual information and achieves new state-of-the-art results on a wide range of established benchmarks. Project page: https://xytian1008.github.io/VAPO/