Devils in Middle Layers of Large Vision-Language Models: Interpreting, Detecting and Mitigating Object Hallucinations via Attention Lens

Hallucinations in Large Vision-Language Models (LVLMs) significantly undermine their reliability, motivating researchers to explore the causes of hallucination. However, most studies primarily focus on the language aspect rather than the visual. In this paper, we address how LVLMs process visual information and whether this process causes hallucination. Firstly, we use the attention lens to identify the stages at which LVLMs handle visual data, discovering that the middle layers are crucial. Moreover, we find that these layers can be further divided into two stages: "visual information enrichment" and "semantic refinement" which respectively propagate visual data to object tokens and interpret it through text. By analyzing attention patterns during the visual information enrichment stage, we find that real tokens consistently receive higher attention weights than hallucinated ones, serving as a strong indicator of hallucination. Further examination of multi-head attention maps reveals that hallucination tokens often result from heads interacting with inconsistent objects. Based on these insights, we propose a simple inference-time method that adjusts visual attention by integrating information across various heads. Extensive experiments demonstrate that this approach effectively mitigates hallucinations in mainstream LVLMs without additional training costs.

Nonconvex One-bit Single-label Multi-label Learning

We study an extreme scenario in multi-label learning where each training instance is endowed with a single one-bit label out of multiple labels. We formulate this problem as a non-trivial special case of one-bit rank-one matrix sensing and develop an efficient non-convex algorithm based on alternating power iteration. The proposed algorithm is able to recover the underlying low-rank matrix model with linear convergence. For a rank-$k$ model with $d_1$ features and $d_2$ classes, the proposed algorithm achieves $O(\epsilon)$ recovery error after retrieving $O(k^{1.5}d_1 d_2/\epsilon)$ one-bit labels within $O(kd)$ memory. Our bound is nearly optimal in the order of $O(1/\epsilon)$. This significantly improves the state-of-the-art sampling complexity of one-bit multi-label learning. We perform experiments to verify our theory and evaluate the performance of the proposed algorithm.