HulluEdit: Single-Pass Evidence-Consistent Subspace Editing for Mitigating Hallucinations in Large Vision-Language Models

TL;DR

HulluEdit is a novel, efficient, single-pass framework that reduces hallucinations in large vision-language models by orthogonally editing hidden states to selectively suppress false patterns without harming visual grounding.

Contribution

It introduces orthogonal subspace editing for LVLMs, enabling effective hallucination mitigation in a single pass without reference models or multiple forward passes.

Findings

Achieves state-of-the-art hallucination reduction on POPE and CHAIR benchmarks.

Preserves model capabilities and inference efficiency.

Outperforms contrastive decoding and static editing baselines.

Abstract

Object hallucination in Large Vision-Language Models (LVLMs) significantly hinders their reliable deployment. Existing methods struggle to balance efficiency and accuracy: they often require expensive reference models and multiple forward passes, or apply static edits that risk suppressing genuine visual evidence. To address this, we introduce HulluEdit, a single-pass, reference-free intervention framework. Our core innovation is orthogonal subspace editing: we decompose the hidden states of the model into orthogonal subspaces - visual evidence, conflicting priors, and residual uncertainty - enabling selective suppression of hallucinatory patterns without interfering with visual grounding. This approach mathematically guarantees that edits applied to the prior subspace leave the visual component entirely unaffected. Extensive experiments show that HulluEdit achieves state-of-the-art hallucination reduction on benchmarks including POPE and CHAIR across diverse architectures, while preserving general capabilities on MME and maintaining efficient inference. Our method consistently outperforms contrastive decoding and static subspace editing baselines, offering a new pathway toward more trustworthy LVLMs.