Despite their impressive capabilities, Large Vision-Language Models (LVLMs) frequently generate responses that are plausible but incorrect or unsupported—commonly referred to as hallucinations. In this study, we investigate whether different types of hallucinations are reflected in the model’s internal representations by probing their encoded features. We focus on two key causes of hallucination in multimodal reasoning: (1) over-reliance on textual priors and (2) preference for user prompts over conflicting visual evidence—factors identified in prior work as frequent and impactful. Our probing results reveal that hallucinations exhibit distinguishable representational patterns, suggesting the potential for a representation-level approach to characterize and mitigate them. Motivated by these findings, we propose Steering HAllucination via RePresentation Engineering (SHARP), a representation-level intervention framework that modulates hallucination-related features during inference. SHARP identifies functional representations responsible for prior-driven biases and visual-context conflicts, and jointly adjusts the model’s internal activations in real time. We evaluate our approach extensively on three large vision-language models across multiple benchmarks. Experimental results demonstrate that SHARP effectively reduces hallucinations while preserving the performance and generalization capabilities of LVLMs.

Large language model editing methods frequently suffer from overfitting, wherein factual updates can propagate beyond their intended scope, overemphasizing the edited target even when it’s contextually inappropriate. To address this challenge, we introduce REACT (Representation Extraction And Controllable Tuning), a unified two-phase framework designed for precise and controllable knowledge editing. In the initial phase, we utilize tailored stimuli to extract latent factual representations and apply Principal Component Analysis with a simple learnbale linear transformation to compute a directional “belief shift” vector for each instance. In the second phase, we apply controllable perturbations to hidden states using the obtained vector with a magnitude scalar, gated by a pre-trained classifier that permits edits only when contextually necessary. Relevant experiments on EVOKE benchmarks demonstrate that REACT significantly reduces overfitting across nearly all evaluation metrics, and experiments on COUNTERFACT and MQuAKE shows that our method preserves balanced basic editing performance (reliability, locality, and generality) under diverse editing scenarios.