acceptodds
Under review as a conference paper at ICLR 2027

When More Seeing Does Not Help: Understanding and Mitigating Hallucinations in Think-with-Images Models

Abstract

Think-with-Images has emerged as a promising paradigm for multimodal reasoning. By acquiring visual evidence through tool calls (e.g., cropping and zooming), these models are widely expected to ground their reasoning in the image and mitigate visual hallucinations. In this paper, we revisit this paradigm and find that more tool invocations fail to consistently reduce hallucinations. Our analyses uncover a pronounced visual recency bias: the model over-relies on the recent visual observation while neglecting salient evidence from preceding tool calls, leading to hallucinations. Motivated by these findings, we propose Rebalancing Attention over Visual Evidence (RAVE), a lightweight, training-free strategy that recycles excessive attention from visual tokens in the latest tool output and reallocates it to historical visual contexts. Extensive experiments across six models and fifteen benchmarks demonstrate that RAVE not only mitigates hallucinations but also enhances multimodal perception and reasoning capabilities. These results show that reliable multimodal reasoning hinges not on invoking more tools, but on how to effectively integrate them along the reasoning chain. The code will be released.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Visual Hallucinations, Think-with-Images, Multimodal Chain-of-Thought, Multimodal Large Language Models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_YpEJOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Rebalancing Attention over Visual Evidence (RAVE), a lightweight, training-free strategy to mitigate hallucinations in think-with-image models. The motivating experiments identifies that visual recency bias is a key factor for hallucinations in think-with-image models. Based on this obversation, the authors proposed RAVE, a training-free strategy that re-distributes the misallocated visual attention to facilitate better integration of semantic information. Experiment results shows RAVE mitigates the model hallucinations in all evaluated benchmarks and baselines. The paper also includes experiments for other general perception and reasoning benchmarks, also ablation studies.

Critical strengths

I really enjoyed reading this work. The paper is well written and easy to follow. The observations are informative and interesting. Proposed method is effective and well examined using comprehensive experiments.

Critical weaknesses
  • This paper only focuses on the hallucinations rather than overall visual capabitilieis, I think tool-calls has broader impacts beyond hallucinations. A more general study will be more interesting and appealing.
  • There is no confidence intervals in the main experiment results. Since some performance gaps are minor, it important to report such stats.
Questions for authors
  • In section 2.1, the author mentioned to apply temperature coefficients T >1 to amplify the attention assigned to all visual tokens. I am wondering why this is the case.
  • How non image manipulation tools will affect here? Can RAVE extend to such cases?
Reviewer_UKsVOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies the problem of hallucinations in thinking-with-images paradigm, and argues that more tool-calls do not necessarily reduce hallucinations. They find that VLMs overly rely on recent visual observation neglecting the important historical evidence from prior tool-calls. They propose a rebalancing method, termed as RAVE which is a training-free strategy to redistribute excessive attention from visual tokens in the latest tool-output to the historical context. They also find that insufficient semantic coverage of tokens is an important contributor to visual hallucinations. Therefore, RAVE reallocates the tokens based on semantic diversity and attention sufficiency, i.e., a historical token receives higher redistribution score if it currently receives little attention (sufficiency) and is semantically different from already selected tokens (semantic diversity). Across several benchmarks and VLM families, RAVE performs consistently better.

Critical strengths

S1. The paper studies an important problem, and motivates it with thorough analyses (Figs. 2 and 3) highlighting the under-utilization of historical context in reasoning.

S2. The proposed method, RAVE, out-performs prior mitigation techniques across several benchmarks and model families indicating that the method is effective in reducing visual hallucinations (Tab. 1), and also improves performance across visual and mathematical reasoning tasks (Tab. 2).

Critical weaknesses

W1. One of the contributions of this work is analyzing hallucinations in think-with-images paradigm and attributing the hallucination to visual recency bias in tool-calls, i.e., more attention is allocated to the latest tool-call and historical tool-outputs are given less and less attention. However, this may be conflated with position bias in transformers, e.g., recent tokens get allocated more attention regardless of whether they originate from the tool [1,2]. A useful experiment would be to compare attention distribution under three settings (i) multiple images provided with the original image at the beginning, (ii) same images provided sequentially through tool-calls, (iii) same images provided at the beginning in different orders. If same phenomenon, as this paper diagnoses, is observed across all three settings, then it is a general positional bias in transformers rather than something that is specific to tool-calls or thinking-with-images paradigm. Further, comparing the magnitude of bias across these settings would also clarify if tool-calls introduce or amplify this behavior.

[1] Tian, X., Zou, S., Yang, Z., & Zhang, J. (2025, June). Identifying and mitigating position bias of multi-image vision-language models. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 10599-10609). IEEE.

[2] Xing, Y., Li, Y., Laptev, I., & Lu, S. (2024). Mitigating object hallucination via concentric causal attention. Advances in neural information processing systems, 37, 92012-92035.

W2. It is a little unclear why RAVE works. Does it work because it identifies semantically complementary evidence, or simply because it discourages concentrated attention distribution? In addition to semantic diversity ablation in Tab. 3, an additional experiment with uniform redistribution should help clarify this.

Questions for authors

Q1. If I understood it correctly, there is a mismatch between Theorem 1 (E.3 in appendices) and Algorithm 1 (RAVE). Particularly, the theorem shows 2-approximation guarantee for semantic coverage when selecting K historical tokens with greedy farthest-point sampling, but Algorithm 1 (in line 10), iterates over the entire historical context (H). If K=∣H∣K = \left|H\right|, then the guarantee is trivially true, and potentially unnecessary. Is there a typo in Algorithm 1's line 10? If not, could the authors explain how would the approximation guarantee for intermediate prefixes relate to final attention redistribution over entire token history?

Q2. Some of the results in Tabs. 1 and 2 are modest. Are these numbers reported with single runs? If so, could the authors provide confidence intervals/standard deviation to assess the statistical significance?

2 comments

I thank the authors for their additional clarification experiments, discussions and comments. All of the points raised earlier have been addressed. I remain in favor of accepting this work.

Dear Reviewer UKsV,

We sincerely thank you for your positive feedback, and we are glad to know that our additional experiments and discussions have addressed all of your concerns. Your insightful comments and suggestions are invaluable in improving the quality and clarity of our paper.

Thank you once again for the time and effort you devoted to reviewing our paper.

Best regards,

Authors

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

This paper investigates visual hallucinations in Think-with-Images models and finds that increasing visual tool calls does not consistently reduce hallucinations. Through attention analysis, the authors identify a visual recency bias, where models overemphasize recent visual observations while neglecting historical visual evidence. To address this issue, the authors propose RAVE, a training-free method that redistributes attention toward historical visual evidence. Experiments across six models and fifteen benchmarks demonstrate improvement.

Critical strengths
  1. The paper identifies the relationship between visual tool calls and hallucinations, revealing that additional visual observations do not necessarily improve reasoning reliability. It further investigates visual recency bias as a potential cause, providing useful insights into the challenges of integrating visual evidence across multiple tool calls.
Critical weaknesses
  1. SoFA (CVPR 2025) has shown that multi-image models exhibit positional attention biases, with some models favoring later images while underutilizing earlier or middle ones. It also proposes a training-free attention adjustment method. The authors should clarify how the proposed Visual Recency Bias relates to and differs from existing visual attention biases in multimodal reasoning.
  2. Several recent studies have investigated the reliability of visual evidence during TWI reasoning. The authors should consider comparing RAVE with relevant training-free methods, such as Active-Look and RTWI, or at least include representative TWI methods from the past two years for a more comprehensive evaluation.
Questions for authors
  1. RAVE mainly addresses cases where relevant visual evidence has been acquired but is not effectively utilized. However, I think errors in TWI reasoning generally fall into two categories: (1) correct localization but incorrect reasoning, and (2) incorrect initial localization followed by repeated inspection of the wrong region. Would attention redistribution still help in the second case, where relevant visual evidence may never have been acquired? A related case study would be helpful.

Overall, the paper is substantial and experimentally solid. I would consider raising my rating if the authors can clarify the novelty of their findings and the advantages of RAVE over existing TWI methods.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.