LISA: A Layer-wise Integration and Suppression Approach for Hallucination Mitigation in Multimodal Large Language Models

Abstract: Multimodal LLMs (MLLMs) excel in vision-language tasks such as image captioning but remain prone to object hallucinations, where they describe objects that do not appear in the image. To mitigate this, we propose \textbf{LISA}, a \textbf{L}ayer-wise \textbf{I}ntegration and \textbf{S}uppression \textbf{A}pproach that enhances generation consistency through hierarchical modulation and multi-layer fusion. LISA leverages the functional hierarchy within MLLMs, where shallow layers provide visual grounding, middle layers encode semantics, and deep layers tend to amplify spurious signals. First, zone-specific spectral modulation stabilizes attention by suppressing over-amplified activations in deeper layers while preserving alignment cues in earlier layers. Second, token-level logits from selected layers are fused via anchor-based routing, with token-wise anchor selection and soft logit fusion enabling adaptive integration during decoding. LISA is fully \textbf{plug-and-play} and can be seamlessly integrated into existing MLLMs, including Qwen2.5-VL. Experiments on multiple benchmarks show that LISA reduces hallucinations by up to 53.6\% in $\mathrm{CHAIR}_I$ and improves POPE F1 by 4.5\%, demonstrating strong generalization across models and tasks.