Tell Model Where to Look: Mitigating Hallucinations in MLLMs by Vision-Guided Attention

TL;DR

This work tackles hallucinations in multimodal large language models by introducing Vision-Guided Attention (VGA), a training-free approach that constructs precise visual grounding from visual tokens via Visual Semantic Confidence (VSC) and uses it to steer attention toward informative regions. VGA is designed to be compatible with FlashAttention and requires only a single forward pass per token, adding minimal latency. The authors further enhance dynamic captioning with Programmed Vision-Guidance (PVG), which adaptively suppresses regions already described as generation proceeds. Across POPE, CHAIR, and AMBER benchmarks and several MLLMs, VGA achieves state-of-the-art dehallucination performance, demonstrating the crucial role of explicit visual grounding in improving visual understanding in MLLMs.

Abstract

Visual attention serves as the primary mechanism through which MLLMs interpret visual information; however, its limited localization capability often leads to hallucinations. We observe that although MLLMs can accurately extract visual semantics from visual tokens, they fail to fully leverage this advantage during subsequent inference. To address this limitation, we propose Vision-Guided Attention (VGA), a training-free method that first constructs precise visual grounding by exploiting the semantic content of visual tokens, and then uses this grounding to guide the model's focus toward relevant visual regions. In image captioning, VGA further refines this guidance dynamically during generation by suppressing regions that have already been described. In VGA, each token undergoes only a single forward pass, introducing a negligible latency overhead of just 4.36\%. In addition, VGA is fully compatible with efficient attention implementations such as FlashAttention. Extensive experiments across diverse MLLMs and multiple hallucination benchmarks demonstrate that VGA achieves state-of-the-art dehallucination performance. Further analysis confirms that explicit visual guidance plays a crucial role in enhancing the visual understanding capabilities of MLLMs.

Figures (11)

• Figure 1: A diagram of vision‑guided attention. We first leverage the semantic features embedded in visual tokens to establish visual grounding, and then guide the model’s visual attention toward the relevant image regions.
• Figure 2: Top-3 tokens in visual logits from LLaVA-7B.
• Figure 3: Analysis of visual semantic confidence.
• Figure 4: Performance comparison of LLaVA-1.5's response (Res.) and its VSC on the POPE benchmark.
• Figure 5: Comparison of visual grounding performance between VSC and the visual attention mechanism of LLaVA-7B. The visual attention maps are extracted from layer 14, following prior work AdaptVis.