Summary
This paper introduces VHExpansion, an automated framework for generating additional visual hallucination (VH) test cases based on existing ones to evaluate multimodal large language models (MLLMs). Targeting binary questions, VHExpansion enhances current test cases by questions and answers through negation, and modifying images using both common and adversarial perturbations. The authors propose a new evaluation metric, symmetric accuracy, which serves as an unbiased measure of MLLM performance in addressing VH by assessing the model’s ability to correctly answer both original and negated questions. Experimental results show that VHExpansion effectively identifies more VH test cases and that fine-tuning MLLMs on the expanded dataset significantly reduces VH without compromising overall performance on other visual question answering (VQA) tasks.
Strengths
The paper is well-written, with clearly defined concepts that are easy to understand. The methodology and experimental design are presented in detail, ensuring transparency and allowing readers to follow the research process with ease.
The authors enhance the rigor of their approach by conducting manual verification to confirm the correctness of the proposed negation operation.
Thorough experiments and comprehensive analyses effectively demonstrate the utility and impact of the VHExpansion framework. Moreover, the paper highlights how the newly introduced symmetric accuracy metric offers a more robust and unbiased evaluation of MLLM performance on VH test cases. This provides insights for future researchers aiming to tackle these challenges.
Weaknesses
1. The proposed method's focus on binary questions limits its generalizability, as both the VHExpansion framework and the symmetric accuracy metric may not directly apply to other types of VH datasets with non-binary answer formats. It would be valuable to explore whether the performance improvements from fine-tuning on expanded VH test cases can transfer to other VH datasets with different answer structures, such as multiple-choice questions. Although the symmetric accuracy metric is tailored to binary questions, the image perturbation techniques could potentially be applied to other VH datasets.
2. As indicated in line 288, the VH datasets used in this study primarily focus on low-level understanding, such as object properties (e.g., color, counting, and position) and object presence. It would be interesting to evaluate whether the proposed method could be effective for VH scenarios related to more advanced comprehension, like sentiment or context-based analysis.
3. The paper would benefit from a more detailed analysis of question types. For example, while the authors note that *three of four common perturbations even increase InstructBLIP’s symmetric accuracy on VHTest,* they do not delve into the underlying reasons. Analyzing how the proposed method performs across different question types and the specific capabilities it tests would provide more insight. For instance, a question involving counting (as shown in Fig. 1) might become more challenging with defocus blur, whereas the question in Fig. 2 might benefit from the same blur by reducing visual distractions.
4. The experimental setup in the "MITIGATING VH VIA FINE-TUNING" section (5.4), claimed as a major contribution, has some issues. The comparison between "fine-tuning on the original VH test cases (baseline)" and "fine-tuning on expanded VH test cases using VHExpansion (expanded)" is not entirely fair. In the *baseline* condition, the fine-tuning dataset includes 200 VH samples and 4,000 general VQA data points, while the *expanded* condition features 4,000 VH samples alongside 4,000 general VQA data points. This imbalance skews the comparison, as the *expanded* condition naturally emphasizes VH cases more. A more balanced comparison would involve scaling (repeating) the VH questions in *baseline* up to match the 4,000 samples in the *expanded* condition, thus equalizing the training set size and ensuring a fair evaluation.
5. Additionally, the use of symmetric accuracy in the evaluation poses a challenge. The *baseline* condition does not include negated questions, yet the evaluation includes them. A fairer approach would be to assess performance using standard accuracy or conduct an ablation study that excludes negation questions in the *expanded* condition. This would clarify whether the observed improvements in VH performance are indeed due to the proposed method.
**Minor Points:**
- In Tables 3-5, the metric *# New successful VH test cases* is used. It would be beneficial to also provide the ratio of successful cases to the total number of generated questions for easier comparisons.
Questions
1. The adversarial attack method employed appears somewhat outdated. Would incorporating more advanced techniques enhance the effectiveness of the proposed framework?
2. Could perturbations, such as defocus blur, degrade the image to a point where the question becomes unanswerable based on the available visual information?