Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Addressing Overthinking in Large Vision-Language Models via Gated Perception-Reasoning Optimization

Xingjian Diao, Zheyuan Liu, Chunhui Zhang, Weiyi Wu, Keyi Kong, Lin Shi, Kaize Ding, Soroush Vosoughi, Jiang Gui

TL;DR

This work tackles overthinking in large vision-language models by revealing that perceptual grounding failures, not just deliberation depth, limit performance. It introduces Gated Perception-Reasoning Optimization (GPRO), a meta-reasoning controller that token-by-token routes computation among a Fast path, a Slow Perception path, and a Slow Reasoning path, guided by failure-attribution supervision derived from ~790k samples and trained with multi-objective PPO. The approach yields a new efficiency-accuracy frontier across five benchmarks, achieving higher accuracy while generating substantially shorter responses and remaining competitive with proprietary giants at smaller scales. Perceptual grounding emerges as a dominant bottleneck, and calibrating the controller to uncertainty enables surgical deployment of slower paths only when needed, with robust improvements across model sizes and tasks.

Abstract

Large Vision-Language Models (LVLMs) have exhibited strong reasoning capabilities through chain-of-thought mechanisms that generate step-by-step rationales. However, such slow-thinking approaches often lead to overthinking, where models produce excessively verbose responses even for simple queries, resulting in test-time inefficiency and even degraded accuracy. Prior work has attempted to mitigate this issue via adaptive reasoning strategies, but these methods largely overlook a fundamental bottleneck: visual perception failures. We argue that stable reasoning critically depends on low-level visual grounding, and that reasoning errors often originate from imperfect perception rather than insufficient deliberation. To address this limitation, we propose Gated Perception-Reasoning Optimization (GPRO), a meta-reasoning controller that dynamically routes computation among three decision paths at each generation step: a lightweight fast path, a slow perception path for re-examining visual inputs, and a slow reasoning path for internal self-reflection. To learn this distinction, we derive large-scale failure attribution supervision from approximately 790k samples, using teacher models to distinguish perceptual hallucinations from reasoning errors. We then train the controller with multi-objective reinforcement learning to optimize the trade-off between task accuracy and computational cost under uncertainty. Experiments on five benchmarks demonstrate that GPRO substantially improves both accuracy and efficiency, outperforming recent slow-thinking methods while generating significantly shorter responses.

Addressing Overthinking in Large Vision-Language Models via Gated Perception-Reasoning Optimization

TL;DR

This work tackles overthinking in large vision-language models by revealing that perceptual grounding failures, not just deliberation depth, limit performance. It introduces Gated Perception-Reasoning Optimization (GPRO), a meta-reasoning controller that token-by-token routes computation among a Fast path, a Slow Perception path, and a Slow Reasoning path, guided by failure-attribution supervision derived from ~790k samples and trained with multi-objective PPO. The approach yields a new efficiency-accuracy frontier across five benchmarks, achieving higher accuracy while generating substantially shorter responses and remaining competitive with proprietary giants at smaller scales. Perceptual grounding emerges as a dominant bottleneck, and calibrating the controller to uncertainty enables surgical deployment of slower paths only when needed, with robust improvements across model sizes and tasks.

Abstract

Large Vision-Language Models (LVLMs) have exhibited strong reasoning capabilities through chain-of-thought mechanisms that generate step-by-step rationales. However, such slow-thinking approaches often lead to overthinking, where models produce excessively verbose responses even for simple queries, resulting in test-time inefficiency and even degraded accuracy. Prior work has attempted to mitigate this issue via adaptive reasoning strategies, but these methods largely overlook a fundamental bottleneck: visual perception failures. We argue that stable reasoning critically depends on low-level visual grounding, and that reasoning errors often originate from imperfect perception rather than insufficient deliberation. To address this limitation, we propose Gated Perception-Reasoning Optimization (GPRO), a meta-reasoning controller that dynamically routes computation among three decision paths at each generation step: a lightweight fast path, a slow perception path for re-examining visual inputs, and a slow reasoning path for internal self-reflection. To learn this distinction, we derive large-scale failure attribution supervision from approximately 790k samples, using teacher models to distinguish perceptual hallucinations from reasoning errors. We then train the controller with multi-objective reinforcement learning to optimize the trade-off between task accuracy and computational cost under uncertainty. Experiments on five benchmarks demonstrate that GPRO substantially improves both accuracy and efficiency, outperforming recent slow-thinking methods while generating significantly shorter responses.
Paper Structure (29 sections, 6 equations, 4 figures, 3 tables)

This paper contains 29 sections, 6 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Reasoning in Vision-Language Models
  4. Efficiency and Adaptive Computation
  5. Reinforcement Learning for Language Models
  6. Method
  7. Gated Perception-Reasoning Architecture
  8. Meta-Reasoning Controller
  9. Computational Paths
  10. Multi-Objective Training
  11. Training Data Construction
  12. Experiments
  13. Experimental Setup
  14. Baselines.
  15. Benchmarks.
  16. ...and 14 more sections

Figures (4)

  • Figure 1: Error attribution of incorrect predictions from Qwen2.5-VL bai2025qwen2 across model scales. Results are based on about 790k samples from ViRL39k vl-rethinker, MathV360K shi2024math, and Mulberry yao2024mulberry, and show that many errors stem from visual perception rather than reasoning.
  • Figure 2: GPRO architecture overview. The meta-reasoning controller receives text hidden states, uncertainty scores, and global image features to route among Fast Path (FFN), Slow Perception Path (Cross-Attention), and Slow Reasoning Path (Self-Reflection).
  • Figure 3: Case Study 1: Animal size ordering. The baseline produces verbose step-by-step comparisons, while GPRO generates a concise direct answer.
  • Figure 4: Case Study 2: Spatial reasoning with cubes. The baseline generates 833 tokens of detailed analysis, while GPRO produces an efficient solution.