MAP: Mitigating Hallucinations in Large Vision-Language Models with Map-Level Attention Processing

TL;DR

This work introduces a novel map-level perspective to mitigate hallucinations in LVLMs, interpreting the hidden states of the model as a 2D semantic map, and proposes Map-Level Attention Processing (MAP), a training-free decoding method that effectively leverages factual information through attention-based map-level operations to improve factual consistency.

Abstract

Large Vision-Language Models (LVLMs) have achieved impressive performance in multimodal tasks, but they still suffer from hallucinations, i.e., generating content that is grammatically accurate but inconsistent with visual inputs. In this work, we introduce a novel map-level perspective to mitigate hallucinations in LVLMs, interpreting the hidden states of the model as a 2D semantic map. We observe that factual information is widely distributed across this map, extending beyond the localized inter- or intra-layer regions targeted by most existing methods (e.g., contrastive decoding and layer-wise consistency). Building on this insight, we propose Map-Level Attention Processing (MAP), a training-free decoding method that effectively leverages factual information through attention-based map-level operations to improve factual consistency. Specifically, we employ Layer-Wise Criss-Cross Attention to progressively refine token representations at each decoding layer by aggregating tokens from both inter- and intra-layer dimensions. Additionally, a Global-Local Logit Fusion mechanism combines logits obtained before and after global attention to further refine predictions and improve accuracy. Our method consistently improves the truthfulness and performance of LVLMs across benchmarks, such as POPE, MME, and MMHal-Bench, demonstrating the potential of the map-level decoding strategy.

Figures (8)

• Figure 1: Comparison of existing 1D and our proposed 2D-Map paradigms. (a) Existing methods primarily focus on utilizing information from localized representations within either the inter-layer or intra-layer dimensions. (b) In contrast, our approach reinterprets the entire set of hidden states as a 2D semantic map, thereby enabling a more holistic integration of information across both the position and layer dimensions.
• Figure 2: (a) A qualitative example showing that faithful semantics are distributed across intermediate hidden states. (b) Statistical comparison showing that in-image objects consistently receive higher token probabilities than hallucinated objects.
• Figure 3: Overall architecture of MAP applied to LLaVA1.5-7B. Specifically, MAP introduces a Layer-Wise Criss-Cross Attention module to refine token representations by aggregating faithful information from a 2D semantic map constructed after each decoding layer. Additionally, a Global-Local Logit Fusion module is integrated to fuse hierarchical content at logit-level to produce reliable predictions.
• Figure 4: Evaluation on MMHal-Bench with LLaVA-1.5 under open-ended scenario.
• Figure 5: Qualitative comparison of MAP in terms of hallucination and informativeness.