acceptodds
Under review as a conference paper at ICLR 2027

Faith-GVT: Evaluating Faithfulness in Grounded Visual-Textual Chain-of-Thought

Abstract

Grounded Visual-Textual Chain-of-Thought (GVT-CoT) makes multimodal reasoning more inspectable by exposing both textual rationales and explicit visual evidence regions. Yet existing faithfulness studies in vision–language models (VLMs) largely examine whether textual reasoning reflects influential cues, leaving underexplored whether textual and regional evidence claims are jointly faithful to model behavior and mutually consistent. We address this gap with Faith-GVT, a unified framework that evaluates GVT-CoT along three complementary dimensions: textual faithfulness, visual faithfulness, and text–visual consistency. Using controlled interventions across image, text, and region channels alongside predicted-region masking, we test whether reported textual and regional evidence responds to behaviorally influential information. Across 51,463 examples from five datasets and four VLMs, we find systematic dissociations among textual disclosure, reported localization, and behavioral sensitivity to interventions. Textual faithfulness averages 83.16% under text interventions but only 34.66% under misleading-region interventions, revealing a substantial gap between text and region influence disclosure. The localized bounding box is not uniformly necessary for preserving the model's answer and can diverge from the textual rationale, particularly in fine-grained document reasoning. Matched comparisons with text-only CoT further show that the effect of explicit grounding on textual faithfulness is model- and intervention-dependent. These results show that grounding introduces additional evidence claims whose faithfulness cannot be assessed by text-only methods alone: textual disclosure, reported localization, behavioral sensitivity, and text–visual consistency capture distinct aspects of multimodal reasoning.

Primary area:
interpretability and explainable AI
Keywords:
Multimodal Reasoning, Chain-of-Thought Faithfulness, Vision-Language Models, Visual Grounding, Multimodal Faithfulness

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_n8H3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces FAITH-GVT, evaluating textual faithfulness, visual faithfulness, and cross-modal consistency in explicitly grounded reasoning through image, text, and region interventions, alongside predicted-region masking. Experiments cover 51,463 samples from five datasets and four VLMs. The results reveal clear dissociations among textual explanations, region localization, and behavioral sensitivity, but some metric definitions and intervention designs do not adequately support their interpretation as measures of faithfulness.

Critical strengths
  • Jointly examines textual explanations and explicit regional evidence claims across natural-image and document tasks, offering greater diagnostic value than evaluating answers or explanation fluency alone.

  • Provides large-scale experiments with flip rates, target-match rates, spatial changes, and reasoning-format comparisons on common answer-flipped samples, supplying substantial data for further analysis.

Critical weaknesses
  • FconsF_\mathrm{cons} equates “the explanation is no longer supported after masking” with consistency, without first verifying support in the original image and within the predicted box. An explanation that was visually unsupported from the outset could receive a high score, while redundant evidence outside the box would be treated as inconsistency. The metric therefore does not reliably measure text–region compatibility.

  • The masking experiment changes both the image and the response requirements: C.5 prohibits inference and permits abstention, unlike the baseline prompt. FmaskF_\mathrm{mask} also directly rewards self-reported unanswerability. Without controls using the original image under the same prompt and area-matched off-evidence masks, evidence removal cannot be separated from abstention tendencies or mask-area effects.

  • The random-region condition is not a clean control: C.1/C.2 add textual guidance to “use the highlighted region as a visual cue” only for random regions, not misleading regions. Random-box sizes are also matched to reference boxes rather than misleading boxes. Differences between these conditions therefore cannot be attributed solely to semantic targeting.

  • Answer flips do not establish that the model used the specified intervention content. On BoundingDocs, misleading-region target-match rates are only 7.44% for Qwen and 8.73% for InternVL, suggesting that many flips may reflect generic disruption rather than reliance on the intended alternative evidence. Interpreting low content-disclosure rates as undisclosed visual influence requires analysis on target-matched subsets and additional perturbation controls.

Questions for authors

Please refer to Weaknesses.

Reviewer_s1eFOpenReview ↗
Rating
3(+2)
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The study introduces an framework for evaluating grounded visual-textual chain-of-thought (CoT) based on: textual faithfulness, visual faithfulness, and textual visual consistency. It combines 3 kind of interventions: images, misleading textual hints, misleading and random region highlighting, and predicted-region masking. Its contributions comprise a unified evaluation protocol, a benchmark containing 51,463 examples derived from five existing datasets, and an empirical evaluation of four LLMs.

Validation includes comparisons between explanation formats with and without explicit grounding, geometric localization measurements, and automated semantic judgments.

Critical strengths
  1. FAITH-GVT examines textual disclosure, evidence localization, masking sensitivity, and text-visual consistency within the same grounded output. This reveals consequential discrepancies: a model may acknowledge influential information while localizing it poorly, or preserve its answer after its reported evidence region is removed. These complementary diagnostics give the framework practical value for assessing whether visible explanations provide reliable evidence about model behavior.

  2. The extensive experiments covers 5 datasets and four LLMs, spanning natural images and documents.

Critical weaknesses
  1. The study is lacking comparison between the proposed method with SOTA methods. This leaves the claimed advance over existing evaluation frameworks unclear and limit my assessment of its novelty.

  2. Equations (10)–(11) treat an explanation as inconsistent when its visual claims remain supported outside the predicted box. However, duplicated evidence could satisfy this condition even when the explanation and box are both valid. an explanation unsupported anywhere could receive one because it is also unsupported outside the box. Thus, the equation checks whether support remains outside the box; it does not independently establish that valid support exists inside it.

Questions for authors
  1. Before applying Equations (10)–(11), do you verify that the original rationale is visually supported by the original image and the predicted region? How do you handle duplicated evidence outside that region? Without these safeguards, the metric appears capable of rewarding an unsupported rationale or penalizing a valid explanation whose evidence appears in multiple locations.

  2. Could you provide comparison between the proposed methods to state-of-the-art methods to proof the novelty of your method?

Reviewer_Cv3TOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FAITH-GVT, a unified framework for evaluating the faithfulness of Grounded Visual-Textual Chain-of-Thought (GVT-CoT) along three dimensions: Textual Faithfulness, Visual Faithfulness, and Text–Visual Consistency. Through Image, Text, and Region Interventions, as well as Predicted-Region Masking, the framework analyzes the relationships among the model's textual reasoning, visual localization, and prediction behavior. The experiments cover five datasets, 51,463 samples, and four VLMs. The results show significant differences in Faithfulness across different interventions, and the visual regions reported by the models are not always consistent with their prediction behavior and textual explanations.

Critical strengths
  1. This paper studies the faithfulness of textual explanations and visual evidence in multimodal reasoning, addressing a meaningful research problem.

  2. FAITH-GVT evaluates faithfulness from three complementary dimensions and combines multiple intervention methods, providing a relatively systematic evaluation framework.

  3. The experiments cover five datasets and four VLMs, including comparisons between Text-only CoT and GVT-CoT, providing relatively broad experimental coverage.

Critical weaknesses
  1. This paper evaluates Textual Faithfulness on Answer-Flipped Cases and requires the CoT to mention and use the intervention evidence. However, even if an intervention causes the model's answer to change, this does not necessarily mean that the changed answer is based on the specific content being intervened on. The model may change its attention or evidence selection due to the intervention, but ultimately answer based on other unmodified visual information. In such cases, even if the CoT faithfully reflects the evidence actually used by the model, it may still be judged as Unfaithful because it does not mention the intervention content. Therefore, the current evaluation may not sufficiently distinguish the behavioral influence of an intervention from the model's reliance on specific evidence.

  2. Text Intervention directly provides linguistic cues, whereas Region Intervention requires additional visual recognition, localization, and cross-modal understanding capabilities. Therefore, the differences in Textual Faithfulness between the two may not only stem from Faithfulness itself, but may also be influenced by basic perception and expression capabilities. Although this paper provides some comparisons on the same datasets, it has not sufficiently ruled out the influence of task difficulty across different modalities on the results.

  3. This paper measures Consistency by masking the predicted Bounding Box and checking whether the original CoT can still be visually supported. However, when multiple equivalent or redundant pieces of evidence exist in the image, the CoT may still be supported by other regions after masking, even if the predicted region correctly supports the textual explanation. Therefore, a lower Consistency score does not necessarily indicate inconsistency between the textual explanation and visual localization; it may simply indicate that the predicted region is not the only valid source of evidence. Although this paper acknowledges the limitation of evidence redundancy, it has not sufficiently quantified its impact on the Consistency results.

Questions for authors
  1. When an intervention causes the answer to change, but the model actually relies on other unmodified visual evidence, why must the CoT mention and use the intervened content? How does the current method determine whether the change in the model's answer was influenced by factors other than the intervened content?

  2. How can the differences in Faithfulness across different interventions be distinguished from differences in modality understanding and task difficulty?

  3. If multiple equivalent evidence regions exist, should it still be considered an inconsistency when the CoT remains supported after masking the predicted region? Could we further quantify the impact of evidence redundancy on the Consistency results?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.