D-LEAF: Localizing and Correcting Hallucinations in Multimodal LLMs via Layer-to-head Attention Diagnostics

2025-09-09

AI Summary

Multimodal large language models (MLLMs) frequently generate hallucinations characterized by visual–textual inconsistency, and existing methods struggle to precisely localize cross-layer attention anomalies responsible for such hallucinations. To address this, we propose the Dynamic Layer-wise Entropy and Attention Fusion (D-LEAF) framework, enabling task-agnostic hallucination localization and correction at inference time with minimal computational overhead. D-LEAF introduces two novel diagnostic metrics—Layer-wise Image Attention Entropy (LIAE) and Image Attention Focus (IAF)—which jointly and adaptively identify the erroneous transformer layer and critical attention head(s) underlying hallucination. Evaluated on image captioning benchmarks, D-LEAF achieves a 53% relative improvement in hallucination reduction; on VQA tasks, it boosts both accuracy and F1-score by approximately 4%. The method effectively suppresses hallucinations while preserving inference efficiency and model generality.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Large Vision Models
• Natural Language Processing: Language Grounding & Multi-modal NLP

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Multimodal Large Language Models (MLLMs) achieve strong performance on tasks like image captioning and visual question answering, but remain prone to hallucinations, where generated text conflicts with the visual input. Prior work links this partly to insufficient visual attention, but existing attention-based detectors and mitigation typically apply uniform adjustments across layers and heads, obscuring where errors originate. In this paper, we first show these methods fail to accurately localize problematic layers. Then, we introduce two diagnostics: Layer Image Attention Entropy (LIAE) which flags anomalous layers, and Image Attention Focus (IAF) which scores attention heads within those layers. Analysis shows that LIAE pinpoints faulty layers and IAF reliably ranks heads that warrant correction. Guided by these signals, we propose Dynamic Layer-wise Entropy and Attention Fusion (D-LEAF), a task-agnostic, attention-guided method that dynamically localizes and corrects errors during inference with negligible overhead. Results show our D-LEAF delivers a 53% relative improvement on standard captioning benchmarks, and on VQA both accuracy and F1-score improve by approximately 4%, substantially suppressing hallucinations while preserving efficiency.

Problem

• Localizing hallucinations in multimodal LLMs
• Correcting attention errors in specific layers
• Improving accuracy while maintaining efficiency

Innovation

• Uses entropy and focus diagnostics to pinpoint errors
• Dynamically localizes and corrects hallucinations during inference
• Applies layer-wise attention fusion with minimal overhead