acceptodds
Withdrawn from ICLR 2027

What Matters, What Misleads: Visual Token Attribution and Hallucination Mitigation in LVLMs

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large Vision-Language Models (LVLMs) can produce sophisticated multimodal predictions, yet it remains unclear which visual evidence actually drives those predictions. Revealing this evidence is fundamental to understanding LVLM behavior and enabling targeted interventions on the visual information that contributes to hallucinations. However, existing attribution methods based on attention, gradients, or perturbations can overlook computational pathways, provide unreliable estimates, or require costly and context-dependent interventions, motivating a direct approach to measuring the contribution of localized visual-token subsets. In this work, we introduce VIRAH, a training-free framework that retains a local neighborhood of visual tokens at a chosen vision-encoder depth, removes other visual tokens, and propagates the retained representations through the remaining model to measure the evidence supported by selected regions. The resulting output distribution serves as a self-probe of the visual evidence supported by selected regions. We evaluate VIRAH for visual attribution, where the probability of an affirmative answer to a target concept query provides the regional attribution score. We further demonstrate that this self-probing capability can be directly leveraged to develop a training-free hallucination mitigation method, using regional object-identification entropy to identify unreliable visual tokens for removal during captioning. Across various LVLMs (Qwen, InternVL, and LLaVA) and multiple benchmarks, VIRAH consistently outperforms existing attribution methods. Specifically, on Qwen2.5-VL-7B, it improves InsertionDeletion by 46.08% and AUPR by 11.94% over the strongest attribution baseline on MS-COCO. For hallucination mitigation on MS-COCO, this simple training-free intervention achieves performance competitive with or surpassing recent methods, reducing CHAIR by up to 1.90% and CHAIR by up to 2.20% while improving Recall by up to 4.04%.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Computer Vision, Large Vision Language Models, Attribution Methods, Mechanistic Interpretability

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_UVwXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses two related problems: identifying visual regions associated with hallucinations and mitigating hallucinations in vision-language models (VLMs). The proposed method, VIRAH, selectively retains small groups of visual tokens and probes the model to assess the information contained in each region. It then identifies regions associated with high predictive entropy and removes their corresponding visual tokens before generating the final caption, aiming to reduce hallucinations.

Critical strengths

Intuitive and easy-to-implement methodology: The proposed approach to identifying potentially hallucination-prone visual regions is intuitive and relatively easy to implement. The paper also acknowledges some of the limitations of its method.

Integration of existing techniques: Although the individual techniques offer relatively incremental novelty, their integration into a unified framework for visual attribution and hallucination mitigation provides an interesting direction for further research.

Critical weaknesses

Lack of theoretical justification for hallucination mitigation: Although the proposed method may demonstrate empirical improvements, removing high-entropy visual tokens does not necessarily reduce hallucinations. In fact, removing visual information could make the model rely more heavily on contextual or linguistic priors, potentially reinforcing the very hallucinations the method aims to mitigate. The paper does not establish a rigorous connection between regional predictive entropy, hallucination risk, and the effectiveness of token removal.

Limited practical applicability and computational efficiency: The proposed approach requires repeatedly probing different image regions for each new input, making it potentially computationally expensive. Moreover, the method does not determine whether hallucinations would occur before applying the intervention, instead removing high-entropy regions regardless of whether they would actually cause hallucinations. These limitations raise concerns about its practicality for system-level hallucination detection and mitigation. The paper also does not provide end-to-end latency measurements for the complete mitigation pipeline under realistic deployment conditions.

Insufficient justification for the necessity of regional probing: Since the method does not directly identify hallucinated claims but instead removes regions associated with high predictive entropy, it remains unclear whether the expensive regional probing procedure is necessary for the observed improvements. The authors should investigate whether simpler interventions could achieve comparable results at a lower computational cost.

Missing ablation studies validating the central hypothesis: The experimental evaluation lacks sufficient ablation studies to establish the relationship between regional predictive entropy and hallucination risk. In particular, the authors should compare removing high-entropy versus low-entropy regions, evaluate entropy-based selection against alternative uncertainty measures, and investigate whether regional entropy actually correlates with the likelihood of hallucination. Such experiments are essential to validate the proposed uncertainty score and determine whether the observed improvements arise from the claimed mechanism rather than from token removal more generally.

Questions for authors

See above

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

This paper addresses visual attribution and hallucination mitigation in large vision-language models. The authors propose VIRAH, a training-free framework that selectively retains local visual-token neighborhoods at different vision-encoder depths and measures their contributions through the model's output distribution. For visual attribution, VIRAH uses target-specific affirmative-answer probabilities to identify relevant image regions. For hallucination mitigation, VIRAH estimates regional predictive uncertainty and removes visual tokens associated with high-entropy regions before generating captions. Experiments across multiple LVLM backbones and benchmarks demonstrate improved attribution faithfulness and localization, along with reduced object hallucinations while preserving visual content.

Critical strengths
  1. VIRAH provides a common visual-token intervention mechanism for identifying relevant visual evidence and mitigating hallucinations. This connection between model interpretability and practical hallucination reduction is meaningful and well motivated.
  2. The framework operates directly on intermediate visual representations without requiring additional training, external detectors, or modifications to model parameters. Its layer-controlled token intervention also enables the analysis of visual evidence at different representation depths.
Critical weaknesses
  1. Although VIRAH uses high regional predictive entropy to identify potentially unreliable visual evidence, the relationship between entropy and hallucination occurrence is not directly quantified. Further analysis would strengthen the justification for entropy-guided token removal.
  2. A controlled comparison against random removal, low-entropy removal, and other token-selection strategies under matched removal ratios would better isolate the contribution of entropy-guided selection.
  3. Although the paper reports attribution latency in Table 8, the end-to-end inference overhead of the hallucination mitigation pipeline, including uncertainty estimation, clustering, and caption generation, is not explicitly reported.
Questions for authors
  1. VIRAH assumes that regions with high predictive entropy provide unreliable visual evidence and contribute to hallucinations. What empirical evidence supports the relationship between regional predictive entropy and hallucination occurrence?
  2. How does the proposed entropy-based token selection compare with random, low-entropy, or alternative token removal strategies under matched removal ratios?
  3. Although the paper analyzes the computational cost of visual attribution, the complete inference overhead during hallucination mitigation remains unclear. It would be helpful to provide end-to-end latency measurements, including regional uncertainty estimation, clustering, token intervention, and final caption generation.
Reviewer_m5H4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes an intervention method for the visual encoder of LVLMs that extends to visual attribution and hallucination mitigation methods. Specifically, the method selects an intervention layer of the visual encoder, and retains the visual tokens only for a local region of concern. For attribution, the region can be scored for saliency with the probability of LVLM's answer to a yes-or-no question. For hallucination mitigation, tokens are clustered by the entropy in LVLM's answer and uncertain tokens are removed. Experiments show advantage of the method on benchmarks of both tasks.

Critical strengths
  • The paper proposes an interesting intervention method that leverages structural property of the Transformer to input local images.
  • The general direction is reminiscent of the cited early interpretability method RISE, which is worth revisiting for VLM.
  • The method derives strategies that leads to reasonable performances on both visual attribution and hallucination mitigation tasks.
Critical weaknesses
  • The attribution task VIRAH works on seems to be limited to yes-or-no questions, whereas existing methods support the explaination of arbitrary autoregressive generation. This would render the application of VIRAH very limited, as visual attribution is expected to explain general model behaviors.
  • The proposed method inserts a subset of visual tokens to the remaining layers of the visual encoder and then the LLM. While this is structurally allowed by the Transformer, it is hard to say how much of a distribution shift this is causing, especially with smaller neighborhood size and without training. It is worth examining if general visual capability is impaired by this operation. For example, a result on a general LVLM benchmark with the hallucination mitigation paradigm could be helpful to provide this evidence.
  • As the operation incurs repeated calls of the visual encoder, it is worth examining the inference cost on hallucination mitigation tasks, particularly compared with other training-free methods. Currently the paper only includes a comparison on the attribution task.
  • The ablation study for intervention layer fluctuate considerably, especially for hallucination tasks, making it hard to identify a clear trend.
Questions for authors
  • What is the sample size for the ablation study of intervention layer?
  • Is it possible to extend the method to attribution tasks explored by EAGLE and other existing work?

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

Related papers

Discussion (0)

Sign in to comment.