Alleviating Hallucinations in Large Vision-Language Models through Hallucination-Induced Optimization

Summary

This paper addresses object hallucination in large visual language models (LVLMs), a common phenomenon that such models generate texts not consistent with images. A viable approach for this issue is contrastive decoding. By comparing logits derived from images and distorted images, visual contrastive decoding reduces statistical bias and uni-modal priors thus alleviates hallucination. This paper points out that distorting images induces uncertainty which leads to unexpected results. To overcome this issue, this paper proposes to train an evil LVLM first by direct preference optimization, as illustrated in Figure 2, and contrast outputs from this model with LVLM responses. Theoretical foundations have been given and experimental results show that such strategy alleviates object hallucinations on multiple benchmarks, including POPE and CHAIR, which are commonly used metrics in this field.

Strengths

- The method is simple and easy to follow. - Contrastive decoding is to contrast outputs from a model with an amateur model. It is an interesting perspective to train an amateur model first using DPO approach. - Writing and presentations are clear. - The performance gains against previous methods look good.

Weaknesses

- **Generalization issue**. In Figure 2, if I understand correctly, this paper induces an LVLM to generate hallucinated outputs (which is referred to as Hallucination Induction Optimization). Both hallucinated and corrected responses are involved in this Figure. If thinking more over objects (e.g., people, tv, dog, clock in Figure 2), a potential issue for this is, trained model in such manner can only be applied to these objects. In proposed hallucination-induced optimization, an LVLM is taught to hallucinate on some objects, these objects can work for later contrastive decoding. But I doubt that such capability can generalize to other objects which are not seen during training, making this a close-vocabulary approach, just like post-hoc correction methods in this field. - Given that proposed hallucination-induced optimization resembles RLHF methods [A] in technical pipeline. Authors are suggested to include some results comparing these methods. Whether an amateur model can help an RLHF model seems an interesting question and may improve this paper. [A] Aligning Large Multimodal Models with Factually Augmented RLHF.

Questions

- For Table 1 where POPE is evaluated, I notice that previous approach VCD has three evaluation setups, including MSCOCO, GQA and A-OKVQA. Only MSCOCO is given. Authors are suggested to include the other two as well for more comprehensive evaluations. - Typos. In line 135, An overview of the proposed HIO method is shown in Fig. 3 (3 should be 2).

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

Authors have indicated limitations in Appendix D Discussion.

Summary

This manuscript presents a novel perspective on implementing contrastive decoding for hallucination mitigation in large visual language models. In contrast to methods such as image perturbation, it achieves a quasi-min-max optimization by enhancing model hallucination followed by contrastive decoding for multiple hallucination targets. The proposed framework and its improved components, i.e., CBTM, AMTH, and ACI, have solid theoretical and empirical derivations. The overall writing of the paper is commendable, and it has achieved state-of-the-art performance.

Strengths

The authors provide a fresh perspective on the overall uncertainty in visual input as well as the exploitation of multiple hallucinatory targets, demonstrating notable gains. They observe and deduce the limitations of the existing optimization frameworks and offer effective solutions.

Weaknesses

1. The description of experimental details is weak. For instance, the experimental setups for Figure 1, Table 2, and Table 5 are not clearly stated, yet the results vary significantly. 2. The manuscript fails to provide the results of the three models purportedly studied. 3. The training details, time and hardware expenditure, and the selection of the alpha parameter, among many other experimental details for the Evil Model, have not been provided. 4. While the overall writing is satisfactory, many details betray a rather hasty preparation of the article, such as the repeated occurrences of acronym definitions, entirely duplicated result descriptions, missing spaces before parentheses in several instances, inconsistent terminologies used for metrics, etc. 5. There appears to be an error in Figure 2: 'benches' should be the correct category? 6. Inappropriate caption for Table 4. 7. Inappropriate bolding for Tables 1 and 3. 8. A Conclusion section is missing from the main text.

Questions

1. In analogy to adversarial training, the utilization of hallucinatory information during the training dynamics is unclear. The subsection 'Acquisition of Multiple Candidate Hallucinations' seems relevant to this aspect, yet the discussion could be further enriched. 2. In comparison to the proposed method, is it feasible to directly calibrate the hallucinatory and targeted logits and achieve both efficiency and effectiveness? 3. Please carefully revise the writing details and enrich the experimental descriptions. I believe this will significantly enhance the clarity and soundness of the paper.

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors have noted the additional overhead introduced by the proposed method during model training and inference. The potential social impact is positive, yet the discussion on this is lacking. The reviewer thinks that the current limitations in relevant datasets may also constrain the method's development. How to seek the annotation of hallucination tokens during dynamic optimization merits consideration.

Summary

This paper focuses on alleviating hallucinations in Large Vision-Language Models. Specifically, the authors introduce a novel optimization strategy named Hallucination-Induced Optimization (HIO). This method amplifies the contrast between hallucinatory and targeted tokens relying on a fine-tuned preference model. Finally, the extensive experimental results verify the effectiveness of HIO, which outperforms state-of-the-art methods across various benchmarks.

Strengths

1. The motivation of this paper is reasonable. 2. The experimental results are significant compared to previous works.

Weaknesses

1. I mainly doubt the novelty. The motivation of this paper is very similar to [1]. Please describe the difference between this work and [1]. 2. This paper is not complete, lacks a Conclusion section, and the ablation study section is too short. Moreover, the implementation details are missing in the paper. Thus, I think the writing of this paper is not ready. [1] Alleviating Hallucinations of Large Language Models through Induced Hallucinations.

Questions

Please see the weaknesses.

Rating

4

Confidence

5

Soundness

3

Presentation

3

Contribution

2

Limitations

n/a

Summary

This paper proposes a method for mitigating hallucinations by training an "evil" LLM to provide logits for contrastive decoding. This "evil" LLM is trained with a dataset that prefers hallucinated samples over true ones during fine-tuning. The logits from this "evil" LLM are then used for contrastive decoding. Experimental results validate the effectiveness of this approach.

Strengths

1. The high-level idea is easy to understand. 2. The method itself is interesting and novel.

Weaknesses

1. The writing is sometimes confusing and makes it difficult to understand details. For example, Sec 4.1 and 4.2 use Eqn 17 in Sec 5 as the motivation of how they design the current method and refer to it frequently. However, Eqn 17 is not explained and introduced previously, making it very difficult for the reader to understand why you design this method like this. It would be much better to explain the high level idea of eqn 17 first. 2. Besides, the figure 1 itself is also confusing. The picture provided is vague and even a human finds it difficult to identify the people and output the true answer. 3. When reading the Sec 5, it is also very confusing. How do you define the ideal logits for contras decoding? Based on my understanding, it should not equal to the hallucinatory token's logits. However, at the later part line 243, you are referring to it as the hallucinatory tokens. Besides, how do you interpret the second line of Eqn 17? I interpret J as the average logit difference between the hallucinatory token and the true token. However, I find it difficult to interpret the left part of the second line of Eq 17. What is the logic behind line 242? The whole explanation does not make sense to me. 4. For Eqn 14, is it too strong to require the minimal logit of correct tokens to be larger than the maximum logit of hallucinatory tokens? For greedy docoding, if one of the correct tokens after contrast is larger than all the hallucinatory tokens, then it will be output?

Questions

Sea the weakness part. 1. Typo: - line 127 should be "probability" instead of "probably". - line 233 it should be $\delta^{*\\{v,x,y_{<t}\\}}$ instead of $\delta^{\prime\\{v,x,y_{<t}\\}}$

Rating

6

Confidence

4

Soundness

2

Presentation

1

Contribution

2

Limitations

Yes, the author has pointed out the computation cost for this method.