Amidst the rapid advancement of artificial intelligence, research on large vision-language models (LVLMs) has emerged as a pivotal area. However, understanding their internal mechanisms remains challenging due to the limitations of existing interpretability methods, especially regarding faithfulness and plausibility. To address this, we first construct a human response interpretability dataset that evaluates the plausibility of model explanations by comparing the attention regions between the model and humans when answering the same questions. We then propose a patchwise cooperative game-based interpretability method for LVLMs, which employs Shapley values to quantify the impact of individual image patches on generation likelihood and enhances computational efficiency through a single input approximation approach. Experimental results demonstrate our method’s faithfulness, plausibility, and robustness. Our method provides researchers with deeper insights into model behavior, allowing for an examination of the specific image regions each layer relies on during response generation, ultimately enhancing model reliability. Our code is available at https://github.com/ZY123-GOOD/Patchwise_Cooperative.

Large Language Models (LLMs) have demonstrated impressive capabilities across various domains, garnering significant attention from both academia and industry. However, enhancing the performance of LLMs typically requires scaling up model sizes or fine-tuning with additional datasets, which results in substantial computational costs. This paper poses an intriguing question: Can we improve the performance of LLMs without additional training? Drawing inspiration from signal processing principles, which suggest that noise often resides in high-frequency components while low-frequency components carry the essence of signals, we propose uncovering untapped potential in LLMs from a frequency perspective. We hypothesize that the high-frequency components in the weight matrices of LLMs’ linear layers may conceal noise that interferes with predictive accuracy. Therefore, we propose conducting spectral modulation in the parameter space of LLMs, which can seamlessly integrate with various models in a plug-and-play manner. Extensive experiments have demonstrated the superiority of our approach, with spectral modulation yielding an average performance improvement of up to 10.12%.

Neural networks are widely used in various NLP tasks for their remarkable performance. However, the complexity makes them difficult to interpret, i.e., they are not guaranteed right for the right reason. Besides the complexity, we reveal that the model pathology - the inconsistency between word saliency and model confidence, further hurts the interpretability. We show that the pathological inconsistency is caused by the representation collapse issue, which means that the representation of the sentences with tokens in different saliency reduced is somehow collapsed, and thus the important words cannot be distinguished from unimportant words in terms of model confidence changing. In this paper, to mitigate the pathology and obtain more interpretable models, we propose Pathological Contrastive Training (PCT) framework, which adopts contrastive learning and saliency-based samples augmentation to calibrate the sentences representation. Combined with qualitative analysis, we also conduct extensive quantitative experiments and measure the interpretability with eight reasonable metrics. Experiments show that our method can mitigate the model pathology and generate more interpretable models while keeping the model performance. Ablation study also shows the effectiveness.

Neural networks are vulnerable to adversarial examples. The adversary can successfully attack a model even without knowing model architecture and parameters, i.e., under a black-box scenario. Previous works on word-level attacks widely use word importance ranking (WIR) methods and complex search methods, including greedy search and heuristic algorithms, to find optimal substitutions. However, these methods fail to balance the attack success rate and the cost of attacks, such as the number of queries to the model and the time consumption. In this paper, We propose PAthological woRd Saliency sEarch (PARSE) that performs the search under dynamic search space following the subarea importance. Experiments show that PARSE can achieve comparable attack success rates to complex search methods while saving numerous queries and time, e.g., saving at most 74% of queries and 90% of time compared with greedy search when attacking the examples from Yelp dataset. The adversarial examples crafted by PARSE are also of high quality, highly transferable, and can effectively improve model robustness in adversarial training.