Fading Focus: Mitigating Visual Attention Degradation in Large Vision-Language Models

W1. The primary observation and motivation, which is that hallucination increases as sequence length grows (i.e., long-term bias), has been well-documented in prior works [1-5]. While the visualization in Figures 1-4 differs from previous studies, the core idea regarding long-term bias remains the same, and is not novel. Additionally, similar attention-based approaches have been explored in the past, as referenced in [2].

W2. The improvements shown in Table 3 are minimal. The authors should have compared their method with more existing techniques to better establish the superiority of IKOD. Given the modest gains, multiple runs should have been conducted to validate the consistency of the improvements over the baseline.

W3. The paper lacks a comparison with related work such as [3], which also involves attention modification to mitigate object hallucination in LVLMs. Although [3] does not directly target long-term bias, it raises an important point that higher attention does not always correlate with higher information content. However, in the method proposed here, tokens with lower image attention weights are selected as anchor tokens for aggregation. This potential conflict between approaches needs to be addressed. A discussion comparing these two methods is necessary to clarify their differences and prevent any confusion

W4. Several figures, such as Figure 1 and Figure 4, are not self-explanatory. For instance, in Figure 1, the meaning of the x-axis and y-axis is unclear, as is the significance of the “top-20%” and “bottom-20%” tokens (the text refers to these as “first-20%” and “last-20%”). In Figure 4, the method for measuring hallucination is not well-explained.

W5. While VILA and SVIT models are used in Figure 2, they are absent from the main experiments. It would strengthen the paper if InstructBLIP were included in Figure 2, and if the VILA and SVIT models were used as backbones in the experimental section for consistency.

W6. It appears that the method is sensitive to the alpha value depending on the model used. Specifically, in the case of InstructBLIP, the impact of the proposed method seems minimal regardless of the alpha value, and performance may even degrade at higher alpha levels. An explanation for this behavior is needed. Additionally, a deeper analysis or statistics on which anchor tokens were selected during the KV merging process would be helpful to provide further insight into this issue.

[1] Analyzing and Mitigating Object Hallucination in Large Vision-Language Models, ICLR 2024.
[2] Mitigating hallucination in large multi-modal models via robust instruction tuning, ICLR 2024.
[3] Don't Miss the Forest for the Trees: Attentional Vision Calibration for Large Vision Language Models, Arxiv 2024.
[4] Multi-Modal Hallucination Control by Visual Information Grounding, CVPR 2024.
[5] VLind-Bench: Measuring Language Priors in Large Vision-Language Models, Arxiv 2024.

The inference process becomes relatively cumbersome.

1. Figures are hard to understand.

• Fig1: what axes represent layer or head? What does the color represent?
• Fig2 is not referred to in the main text.
• Fig5: the legends are confusing, does the orange and blue tokens also apply to KV cache?

2. How significant is the image attention diminishing? With longer sequences, the average attention score will also decrease. How can we identify the actual cause?
3. The conclusion that weak image attention leads to hallucination may not be valid. Fig.4 only shows that hallucinated tokens are prone to appear later in the sequence. For example, in Fig.4(a), there are also tokens with low image attention scores at the beginning of the sequence, but they do not cause much hallucination.
4. What are the lengths of image tokens and text tokens used in the experiments? In multi-modal models, there are usually more image tokens than text tokens. If that's the case, will an increase in generated text tokens significantly affect image attention?
5. The design choice of KV merging needs further clarification. Why anchor tokens need to be chosen? If the motivation is to decrease the length, is it possible to use a uniform merging?

1. The main contribution which is the observation that LVLMs tend to hallucinate more with the length of a generated sequence is not novel and has been discussed in previous works e.g. M3ID [1] - which is also missing in this work.
2. The formal definitions are not very clear and in some parts incomplete. Therefore it makes it hard to understand IKOD in detail and appreciate the technical contribution of this paper.
3. In general the paper has multiple flaws in presentation and formulations. There are multiple typos and sentences without a meaning. It lacks clarity in several aspects as detailed in Questions section.
4. (minor) inconsistent references - multiple papers have their proceedings versions instead of arxiv. It is typically preferable to cite official proceedings. E.g. CHAIR:
   Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. Object hallucination in image captioning. In Empirical Methods in Natural Language Processing (EMNLP), 2018

[1] Favero, Alessandro, et al. "Multi-modal hallucination control by visual information grounding." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024.

This method requires an additional forward pass during inference, which will lead to extra inference computational cost.