Look Closer! An Adversarial Parametric Editing Framework for Hallucination Mitigation in VLMs

TL;DR

This work tackles hallucinations in Vision-Language Models by introducing ALEAHallu, an Activate-Locate-Edit adversarial framework that learns to edit a small, targeted subset of parameters to emphasize visual grounding. It constructs an activation dataset of grounded versus hallucinatory responses and localizes hallucination-prone regions by analyzing layer-wise representation differences, then uses adversarial prefix tuning and a constrained editing objective to erode reliance on priors. Across image captioning, VQA, and MME perception tasks, ALEAHallu substantially reduces hallucinations while maintaining or improving factual grounding, with no additional inference overhead. The approach offers a practical, trainable alternative to heuristic decoding strategies, enhancing reliability and safety of VLMs in real-world settings.

Abstract

While Vision-Language Models (VLMs) have garnered increasing attention in the AI community due to their promising practical applications, they exhibit persistent hallucination issues, generating outputs misaligned with visual inputs. Recent studies attribute these hallucinations to VLMs' over-reliance on linguistic priors and insufficient visual feature integration, proposing heuristic decoding calibration strategies to mitigate them. However, the non-trainable nature of these strategies inherently limits their optimization potential. To this end, we propose an adversarial parametric editing framework for Hallucination mitigation in VLMs, which follows an \textbf{A}ctivate-\textbf{L}ocate-\textbf{E}dit \textbf{A}dversarially paradigm. Specifically, we first construct an activation dataset that comprises grounded responses (positive samples attentively anchored in visual features) and hallucinatory responses (negative samples reflecting LLM prior bias and internal knowledge artifacts). Next, we identify critical hallucination-prone parameter clusters by analyzing differential hidden states of response pairs. Then, these clusters are fine-tuned using prompts injected with adversarial tuned prefixes that are optimized to maximize visual neglect, thereby forcing the model to prioritize visual evidence over inherent parametric biases. Evaluations on both generative and discriminative VLM tasks demonstrate the significant effectiveness of ALEAHallu in alleviating hallucinations. Our code is available at https://github.com/hujiayu1223/ALEAHallu.

Figures (4)

• Figure 1: An example of ALEAHallu.
• Figure 2: ALEAHallu consists of three main components: activation dataset construction, editing region localization and adversarial parameter editing.
• Figure 3: Results on MME.
• Figure 4: Visualization of Visual Attention Scores.