Mitigating Object Hallucination via Concentric Causal Attention

Summary

The paper attributes hallucinations in Large Vision-Language Models (LVLMs) to Rotary Positional Encoding (RoPE). It observes that LVLMs inherit a long-term decay issue from RoPE, where the inner-product of two tokens decays relative to their distance. This results in weaker visual-text interactions when the tokens are distant, leading to more frequent hallucinations if relevant visual tokens are far from the current generating token. To address this, the paper proposes a novel method called Concentric Causal Attention (CCA) to mitigate the effects of RoPE. Experiments show that CCA effectively reduces hallucinations and enhances the perception capability of LVLMs.

Strengths

1. Touching the positional encoding aspect is quite fresh and particularly advantageous as it avoids the latency and multiple inference requirements seen in contrastive decoding methods such as VCD [1] and M3ID [2], which need two output probability distributions. 2. Experimental results are quite strong. --- [1] Mitigating Object Hallucinations in Large Vision-Language Models through Visual Contrastive Decoding, CVPR 2024. [2] Multi-Modal Hallucination Control by Visual Information Grounding, CVPR 2024.

Weaknesses

1. The method is appealing as it addresses a different aspect than existing methods, but the writing can be improved in several areas. For example, the abstract should include short background information on RoPE and an explanation of the long-term decay problem, which will guide readers more friendly reading. Additionally, the captions for Figures 1 and 2 could be more concise and communicate more effectively to enhance readability. Figures 2 and 3 are difficult to understand and need clearer presentation. 2. Since long-term decay is an inherent problem of RoPE, the paper should compare RoPE with other standard positional encodings (e.g., absolute, relative, learnable positional encodings) as well as some advanced positional encodings. This would provide a more comprehensive analysis of the issue. 3. It is important to note that VCD [1] and OPERA [3] are training-free methods. Therefore, comparisons should also include more recent training-based methods. --- [1] Mitigating Object Hallucinations in Large Vision-Language Models through Visual Contrastive Decoding, CVPR 2024. [2] OPERA: Alleviating Hallucination in Multi-Modal Large Language Models via Over-Trust Penalty and Retrospection-Allocation, CVPR 2024.

Questions

1. In Figure 1, why are the attention weights higher at the beginning and the end? If this is related to the image content itself, then Figure 1 needs to show both the image and the textual query together to determine whether the attention is well distributed with respect to the query-object relationship. If the example is averaged over 3k POPE examples, this should be clearly stated. 2. Is the method applicable to other widely used models like InstructBLIP [1] and Qwen-VL [2]? If these models do not use RoPE, the paper should mention this and discuss the implications. 3. Does the CCA method end positional encoding at the center because objects are statistically more likely to be located at the center? Is this the reason for the higher attention at the beginning and the end in Figure 1(b)? Empirical evidence supporting this assumption should be provided. --- [1] InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning, NeurIPS 2023 [2] Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond, Arxiv 2023

Rating

5

Confidence

5

Soundness

3

Presentation

2

Contribution

2

Limitations

If the method is limited to LVLMs that use RoPE (e.g., LLaVA, MiniGPT), it is necessary to mention this as a limitation.

Summary

The paper shows that object hallucination in LVLMs is linked to the commonly adopted Rotary Position Encoding (RoPE) strategies. The long-term decay in RoPE causes hallucinations when important visual tokens are distant from visual instructions. To address this, the authors propose the Concentric Causal Attention strategy to reduce the distance between these tokens. However, experimental results of the proposed method is not promising on several benchmark datasets.

Strengths

The analysis of the long-term decay in RoPE and its impact on hallucinations in LVLMs is novel.

Weaknesses

- The method section is quite short, and it seems not covering a comprehensive solution to address LVLM hallucination problem. It lacks justification on the proposed scanning method. Specifically, the concentric causal masking section is not clear. - Performance of the proposed method is not promising on several benchmark datasets. For example, the proposed method sometimes achieves the best results only when combined with other state-of-the-art methods. It might provide more insight if there is an analysis on when and why the proposed CCA alone will work or fail (Table 2). - The detailed description of the method with the figure is not clear. Please see the questions below.

Questions

- Please clarify what do the orange and yellow colors represent in Figure 3(d)? The explanation of casual masking in Fig.3d is not clear: need guidance on interpreting the concentric causal masking in Figure 3. - What is the distribution of the aggregated correct responses in Figure 2 with the proposed concentric causal attention? Does the distribution different from the raster scan and reverse raster scan? - The paper compared with raster-scan as the baseline. But how do you justify that the concentric positional assignment in Figure 3 is the best solution (or is good enough)? Did you consider other scan designs, for example, the diagonal-scan, zig-zag scan design?

Rating

5

Confidence

4

Soundness

2

Presentation

2

Contribution

2

Limitations

Yes, the author adequately addressed the limitations and potential negative societal impact of their work.

Summary

This paper analyze the long-term dependency between text token and visual token in LVLMs from a novel positional encoding perspective by replace the RoPE method. The analysis shows that RoPE introduce clear long-term decay regarding the attention scores. The authors propose a novel concentric causal attention (CCA) including visual token re-organization and concentric causal masking to keep 2-D spatial locality while shortens visual-instruction distance. Experimental results show CCA can reduce hallucination on both discriminative and generative benchmarks while keep comparable performance on general benchmarks.

Strengths

- The motivation of exploring source of hallucination from a long-term decay perspective is clear. - The paper is well organized and clear. - The proposed CCA method seems effective and easy-to-implement.

Weaknesses

- The experimental setup lacks rigor. The VCD results are sourced from the original paper, which is based on LLaVA 1.5. However, the baseline provided by the authors utilizes the pre-training scheme of LLaVA 1.0 along with the projection module and instruction fine-tuning data from LLaVA 1.5. This discrepancy raises concerns about the comparability of the results. Additionally, it is unclear why Table 3 does not include results for opera and why Table 4 omits the VCD results.

Questions

- In Figure 1 (b), would the removal of the RoPE positional encoding makes the self-attention process out of distribution? Since the model is pre-trained with RoPE applied. How would this affect the quantitative results? - The explanation of concentric causal masking in Figure 3 and the corresponding paragraph is somewhat unclear. Could you clarify why the attention masks for the first seven rows are identical? A more comprehensive and detail introduction helps a lot.

Rating

6

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

authors adequately addressed the limitations

Summary

This paper explores how current LVLM's hallucination appears through analyzing the impact of RoPE long-term decay on vision information attenuation during flow. It gives clear visualization results and theoretical evidence to prove the basic point, revealing that the causal attention mask and RoPE embbeding is not appropriate for non-text-modal input (e.g., vision tokens). Based on these, the authors argue that LVLM's hallucinations are mainly attributed by the RoPE long-term decay and the mismatch between vision tokens and causal attentions. To this end, the authors propose vision token re-organization and concentric causal attention masking to alleviate the hallucination. Experiments demonstrate the promising performance of the proposed methods.

Strengths

1. This paper gives a good explanation for the relationship between hallucination and long-term decay in LVLMs. 2. The proposed method is well motivated and technically sound.

Weaknesses

1. When introducing the long-term decay and the attenuation of information flow, it would be better to cite some references like OPERA (some of the early works that claim the relationship between hallucination and long-term decay). 2. One suggestion for the experiment on Figure 2: Although the authors calculate and visualize the results on thousands of samples, it would be better to reverse the order of vision patches (before CLIP-ViT) and conduct the same experiment again. We should rule out the possibility that most of correct answers are naturally located at the lower region of images. 3. Both the proposed method relies on the hypothesis that the most of main contents of images will be located at the centric region. It generally make sense, but not enough to be an accurate solution. 4. The experiments are somehow insufficient. It would be better to add some ablation studies and evaluation on LVLM benchmarks. MME hallucination split and LLaVA Bench are not enough. It would be better to add results on MMBench, Seed Bench, TextVQA, etc.

Questions

See the Weaknesses.

Rating

6

Confidence

5

Soundness

4

Presentation

3

Contribution

4

Limitations

N/A