Source-linked AI summary
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
LVLMs can overthink simple queries because slow reasoning adds verbosity and inefficiency, while existing adaptive methods overlook visual perception failures. GPRO uses failure-attribution supervision and a meta-reasoning controller to route computation among fast, perception, and reasoning paths, improving accuracy while reducing response length across five benchmarks.
Problem
Existing adaptive reasoning methods overlook visual perception failures, although overthinking can make LVLM responses verbose, inefficient, and sometimes less accurate.
Method
GPRO uses approximately 790k teacher-attributed samples and a token-level controller that routes generation through fast, visual re-analysis, or reasoning-refinement paths.
Results
Across five benchmarks, GPRO improves accuracy while substantially reducing response lengths.
Takeaways & Limitations
Adaptive multimodal reasoning can allocate computation to perception or reasoning instead of applying intensive deliberation uniformly.
Takeaways & Limitations
Failure-attribution supervision depends on GPT-based teacher judgments, which may be ambiguous when perception and reasoning errors are intertwined.
Abstract
from arXiv · showhide
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.
1 Introduction
LVLMs benefit from step-by-step reasoning, but uniform deliberation can cause overthinking and errors rooted in visual perception. GPRO addresses this by learning to route computation between fast generation, visual re-analysis, and reasoning refinement.
- Motivation: Chain-of-thought reasoning improves LVLM problem solving by decomposing complex tasks into intermediate steps.The approach has been applied across arithmetic, scientific, and visually grounded reasoning tasks.
- Motivation: Visual perception failures are a substantial source of LVLM errors and may not be corrected by additional deliberation.The analysis motivates accounting for perceptual uncertainty alongside reasoning adaptation.
- Motivation: Uniform intensive computation causes verbose rationales, test-time inefficiency, and sometimes errors on simple questions.Overthinking includes redundant visual descriptions and unnecessary reasoning chains for direct recognition or counting tasks.
- Approach: GPRO learns failure distinctions from approximately 790k samples labeled by stronger teacher models as perceptual hallucinations or reasoning failures.The supervision provides targeted signals for deciding when additional perceptual or reasoning computation is warranted.
- Approach: GPRO routes token-level computation through a lightweight fast path, a visual re-analysis path, or a self-reflective reasoning path.The meta-reasoning controller evaluates internal state at each generation step to allocate computation where needed.
- Results: Experiments on five benchmarks show strong accuracy with substantially reduced response lengths.The paper presents this as an effective approach for adaptive multimodal reasoning.
2 Related Work
Prior vision-language reasoning methods improve complex-task performance through self-correction, search, adaptive computation, and reinforcement learning, but often incur substantial overhead. GPRO extends adaptive reasoning with visual-perception signals so its controller can trigger visual re-analysis as well as reasoning refinement.
- Reasoning in Vision-Language Models: Chain-of-thought and iterative self-correction enable multimodal models to decompose problems and revise reasoning paths.Tree of Thoughts additionally explores alternative paths and backtracking when initial approaches fail.
- Reasoning in Vision-Language Models: These reasoning methods can push performance on complex tasks but often incur substantial computational overhead.The related-work discussion identifies computational cost as a recurring limitation.
- Adaptive Computation: Adaptive computation methods route simpler inputs through lighter processing while reserving greater capacity for complex cases.Mixture-of-Experts, early exits, and model cascades are presented as examples of selective computation.
- Reinforcement Learning: Reinforcement learning is used to train adaptive computation policies that balance objectives such as accuracy and efficiency.Prior multimodal systems apply RL through formalization, rule-based training, or iterative self-improvement.
- GPRO Positioning: GPRO adds visual perception signals to adaptive reasoning, enabling dedicated visual re-analysis when perceptual uncertainty is high.Prior adaptive reasoning methods discussed here focus solely on reasoning depth.
3 Method
GPRO adds token-level routing to LVLMs, selecting fast generation, visual re-analysis, or internal reasoning based on the model’s state and uncertainty. It is trained with failure-attribution data and a multi-objective reward balancing accuracy, computational cost, and calibration.
- 3 Method: GPRO replaces selected decoder feed-forward layers with a Gated Perception-Reasoning module that dynamically allocates computation.Each module contains a meta-reasoning controller and three computational paths.
- 3.1 Gated Perception-Reasoning Architecture: The controller uses decoder hidden states, predictive entropy, and global image features to select fast, perception, or reasoning computation for each token.The resulting action selects one of three paths at each generation step.
- 3.1.2 Computational Paths: The fast path preserves efficient base-model generation for straightforward cases, while the perception path uses cross-attention to re-examine potentially misinterpreted visual details.The reasoning path instead supports internal self-reflection and correction without generating additional output tokens.
- 3.1.2 Computational Paths: The reasoning path passes the current hidden state and recent context to a meta-reasoning module, enabling correction of logical errors before subsequent generation.This path is engaged when logical uncertainty is high.
- 3.2 Multi-Objective Training: GPRO trains its controller with PPO-based reinforcement learning using task, cost, and calibration rewards whose weighting coefficients trade accuracy against efficiency.The cost reward penalizes expensive slow paths, while calibration encourages uncertainty before errors and confidence before correct outputs.
- 3.2 Multi-Objective Training: Training data labels visual perception failures versus reasoning errors from approximately 790K multimodal samples, then oversamples difficult cases into model-specific curricula.Teacher-model attribution analyzes the question, image, incorrect answer, and ground truth; curricula contain approximately 40K samples for 3B and 30K for 7B models.
4 Experiments
Experiments evaluate GPRO across five multimodal reasoning benchmarks, measuring accuracy and response length alongside ablations and qualitative cases. GPRO improves the efficiency–accuracy trade-off, with visual re-analysis and calibrated routing especially important for avoiding overthinking.
- Experimental Setup: GPRO is evaluated on five benchmarks using accuracy and average response length to assess both performance and efficiency.The benchmarks cover visual mathematical reasoning and integrated multimodal capabilities.
- Main Results: 1.8% higher accuracy and 51.5% fewer tokens than Qwen2.5-VL-7B on MathVerse demonstrate GPRO’s improved efficiency–accuracy trade-off.Response length decreases from 388.9 to 188.4 tokens.
- Main Results: GPRO-7B reaches 31.2% on MathVision versus 29.9% for R1-OneVision while reducing inference cost by approximately 3.5×.The comparison illustrates the cost of unconditional long-context reasoning.
- Main Results: GPRO-7B achieves 31.2% on MathVision versus 30.4% for GPT-4o and matches Qwen-VL-Max at 74.2% on MathVista.These results show competitiveness with substantially larger proprietary models on selected benchmarks.
- Main Results: GPRO-3B outperforms FAST-3B on four of five benchmarks and improves over Qwen2.5-VL-3B by 9.6% on MathVerse.The reported gains extend across both 3B and 7B settings.
- Ablation Study: Removing Slow Perception reduces accuracy by 3.4% on MathVision and 4.4% on MathVerse, exceeding the 1.7% reduction associated with removing Slow Reasoning.The ablation results identify visual grounding as the larger contributor in these tests.
- Ablation Study: The calibration reward causes a 2.3%–2.5% performance drop when removed, while GPRO-7B assigns 73% of tokens to Fast Path, 17% to Slow Perception, and 10% to Slow Reasoning.The controller therefore uses slow computation selectively rather than as a default mode.
- Qualitative Analysis: Qualitative cases show GPRO avoiding verbose reasoning for simple visual ordering while combining perception and reasoning paths for cube-based spatial reasoning.The cube case uses Slow Perception to interpret the 3D structure and Slow Reasoning to verify counting logic.
5 Conclusion
The paper introduces GPRO, a gated adaptive-computation framework that attributes failures to perception or reasoning and routes token generation accordingly. Across five benchmarks, it improves accuracy while reducing response length.
- Conclusion: GPRO uses failure-attribution supervision to distinguish perceptual errors from reasoning errors and guide token-level routing among three computation paths.The paths support fast execution, visual re-examination, and reasoning refinement.
- Conclusion: Across five challenging benchmarks, GPRO improves accuracy while consistently reducing response length.The conclusion presents this as evidence for adaptive multimodal reasoning.
6 Limitations
GPRO remains effective on vision–language reasoning benchmarks, but its supervision and routing design impose important limitations. The method depends on GPT-based attribution and uses discrete routing over fixed computation paths.
- GPT-based failure attribution may produce ambiguous labels when perception and reasoning errors are closely intertwined.The supervision quality depends on the reliability of the teacher model’s judgments.
- GPRO’s discrete routing scheme limits interventions to a fixed set of computation paths.The authors identify finer-grained or continuous control as a direction for future work.
Ethical Considerations
The study uses publicly available datasets and academically licensed models, and reports no ethical concerns to the authors’ knowledge.
- All experiments use publicly available datasets and models licensed for academic research.
- The authors report no ethical concerns to the best of their knowledge.
A Benchmarks •
The evaluation spans six multimodal benchmarks covering mathematical, visual, dynamic, and integrated reasoning. Together, they test visual grounding, multi-step reasoning, robustness across variants, and broad multimodal capabilities.
- MathVision evaluates multimodal mathematical reasoning across 16 disciplines and five difficulty levels using visual scenarios such as diagrams and graphs.Its tasks require precise visual grounding followed by multi-step mathematical reasoning.
- MathVerse varies textual and visual information in diagram-based mathematics to expose failures from incorrect visual grounding.
- MathVista covers geometry, algebra, logic, and scientific reasoning across diverse visual inputs from 31 multimodal datasets.
- DynaMath generates concrete problem variants by altering visual and numerical conditions to test generalization beyond isolated instances.
- MM-Vet evaluates integrated multimodal capabilities including visual recognition, spatial reasoning, OCR, mathematics, and commonsense understanding.
B Baselines •
The baselines include proprietary and open-weight vision–language models, multimodal reasoning systems, and adaptive fast–slow methods. They represent approaches based on stronger backbones, explicit reasoning, reinforcement learning, and adaptive reasoning depth.
- GPT-4o, Claude-3.5 Sonnet, and Qwen-VL-Max provide competitive proprietary multimodal baselines.
- Qwen2-VL and Qwen2.5-VL are open-weight models emphasizing stronger visual encoding, grounding, instruction following, and reasoning stability.
- Mulberry uses Collective Monte Carlo Tree Search to construct multimodal reasoning data with explicit intermediate states.
- Virgo applies long-form chain-of-thought uniformly across inputs without distinguishing perceptual uncertainty from reasoning errors.
- Curr-ReFT, LMM-R1, MM-R1, Vision-R1, R1-OneVision, and OpenVLThinker use reinforcement learning or structured reasoning to improve multimodal capabilities.Their designs include curriculum learning, rule-based rewards, cross-modal decomposition, modality bridging, formalization, planning, reflection, and verification.
- FAST adapts reasoning depth using question difficulty and image complexity, targeting shorter reasoning on easy problems and deeper reasoning on hard ones.
C Examples of Failure Attribution
The paper distinguishes visual perception failures from reasoning errors in its failure-attribution pipeline. This distinction links each failure type to a different corrective response.
- Visual perception failures involve misreading, misidentifying, or incorrectly grounding key visual evidence.
- Reasoning errors occur when the model captures relevant visual information but makes mistakes during inference or calculation.
- Perception failures call for re-examining the image, whereas reasoning errors require a different correction focused on subsequent reasoning.
D Extended Case Study Figures
The extended figures enlarge the main paper’s case studies and failure-attribution examples for closer inspection. They illustrate visual and reasoning failures alongside concise-versus-verbose solution behavior.
- Enlarged versions of the main-paper case-study figures improve readability and inspection of visual details and model outputs.
- Figures 5 and 6 show examples of Qwen visual perception failures with corresponding GPT-based error attribution.
- Figures 7 and 8 show examples of Qwen reasoning failures with corresponding GPT-based error attribution.
- In animal size ordering, the baseline gives verbose step-by-step comparisons while GPRO produces a concise direct answer.
- In spatial reasoning with cubes, the baseline generates 833 tokens of detailed analysis while GPRO produces an efficient solution.