Agentic Reasoning for Robust Vision Systems via Increased Test-Time Compute

2025-09-19

AI Summary

Existing vision systems lack training-free robust reasoning capabilities for high-stakes domains such as remote sensing and medical diagnosis. Method: We propose a training-free intelligent visual reasoning framework featuring a novel Think–Critique–Act agent-style reasoning loop. It seamlessly integrates vision-language models with pure vision models to dynamically construct verifiable reasoning chains and scale computational resources at test time—without fine-tuning or additional training. Contribution/Results: By augmenting only the inference process, our method significantly improves system trustworthiness and robustness. On multiple challenging visual reasoning benchmarks, it achieves up to a 40 percentage-point absolute accuracy gain. This work provides the first systematic empirical validation of the critical role of “amplified test-time computation” in enhancing robustness for complex visual decision-making.

Technology Category

• Computer Vision: Visual Reasoning & Symbolic Representations
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning
• Knowledge Representation and Reasoning: Reasoning with Beliefs

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Developing trustworthy intelligent vision systems for high-stakes domains, emph{e.g.}, remote sensing and medical diagnosis, demands broad robustness without costly retraining. We propose extbf{Visual Reasoning Agent (VRA)}, a training-free, agentic reasoning framework that wraps off-the-shelf vision-language models emph{and} pure vision systems in a emph{Think--Critique--Act} loop. While VRA incurs significant additional test-time computation, it achieves up to 40% absolute accuracy gains on challenging visual reasoning benchmarks. Future work will optimize query routing and early stopping to reduce inference overhead while preserving reliability in vision tasks.

Problem

• Enhancing robustness of vision systems without retraining
• Developing trustworthy AI for high-stakes visual domains
• Improving accuracy on challenging visual reasoning tasks

Innovation

• Agentic reasoning framework with Think-Critique-Act loop
• Wraps off-the-shelf vision-language and pure vision models
• Training-free approach requiring increased test-time compute