Journey Before Destination: On the importance of Visual Faithfulness in Slow Thinking

2025-12-13

AI Summary

Existing reasoning-augmented vision-language models (VLMs) generate chain-of-thought (CoT) outputs suffering from low visual fidelity: intermediate perceptual steps either diverge from image evidence (“hallucinatory reasoning”) or remain faithful yet yield incorrect final predictions—defects invisible to conventional end-answer accuracy evaluation. Method: We introduce “visual fidelity” as a novel, independent evaluation dimension and propose a training-free, reference-free framework that decomposes CoT into perceptual and reasoning steps and self-evaluates each step. We further design a lightweight zero-shot self-reflection mechanism enabling localized regeneration of perceptual steps. Contribution/Results: Evaluated across multiple VLMs and perception-intensive benchmarks, our method significantly reduces non-faithful perception rates while preserving end-answer accuracy, thereby enhancing the reliability and interpretability of multimodal reasoning.

Technology Category

• Computer Vision: Large Vision Models
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Safety and Robustness

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work

Abstract

Reasoning-augmented vision language models (VLMs) generate explicit chains of thought that promise greater capability and transparency but also introduce new failure modes: models may reach correct answers via visually unfaithful intermediate steps, or reason faithfully yet fail on the final prediction. Standard evaluations that only measure final-answer accuracy cannot distinguish these behaviors. We introduce the visual faithfulness of reasoning chains as a distinct evaluation dimension, focusing on whether the perception steps of a reasoning chain are grounded in the image. We propose a training- and reference-free framework that decomposes chains into perception versus reasoning steps and uses off-the-shelf VLM judges for step-level faithfulness, additionally verifying this approach through a human meta-evaluation. Building on this metric, we present a lightweight self-reflection procedure that detects and locally regenerates unfaithful perception steps without any training. Across multiple reasoning-trained VLMs and perception-heavy benchmarks, our method reduces Unfaithful Perception Rate while preserving final-answer accuracy, improving the reliability of multimodal reasoning.

Problem

• Evaluates visual faithfulness in reasoning chains of VLMs
• Distinguishes perception from reasoning steps without training
• Improves reliability by detecting and regenerating unfaithful steps

Innovation

• Proposes training-free framework for evaluating visual reasoning faithfulness
• Decomposes reasoning chains into perception versus reasoning steps
• Uses lightweight self-reflection to regenerate unfaithful perception steps