Suppressing VLM Hallucinations with Spectral Representation Filtering

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Ali, Ameen, Zoabi, Tamim, Wolf, Lior
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908656304390144
author Ali, Ameen
Zoabi, Tamim
Wolf, Lior
author_facet Ali, Ameen
Zoabi, Tamim
Wolf, Lior
contents Vision-language models (VLMs) frequently produce hallucinations in the form of descriptions of objects, attributes, or relations that do not exist in the image due to over-reliance on language priors and imprecise cross-modal grounding. We introduce Spectral Representation Filtering (SRF), a lightweight, training-free method to suppress such hallucinations by analyzing and correcting the covariance structure of the model's representations. SRF identifies low-rank hallucination modes through eigendecomposition of the covariance of the differences between features collected for truthful and hallucinatory captions, revealing structured biases in the feature space. A soft spectral filter then attenuates these modes in the feed-forward projection weights of deeper vLLM layers, equalizing feature variance while preserving semantic fidelity. Unlike decoding or retraining-based approaches, SRF operates entirely post-hoc, incurs zero inference overhead, and requires no architectural modifications. Across three families of VLMs (LLaVA-1.5, MiniGPT-4, and mPLUG-Owl2), SRF consistently reduces hallucination rates on MSCOCO, POPE-VQA, and other visual tasks benchmarks, achieving state-of-the-art faithfulness without degrading caption quality.
format Preprint
id arxiv_https___arxiv_org_abs_2511_12220
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Suppressing VLM Hallucinations with Spectral Representation Filtering
Ali, Ameen
Zoabi, Tamim
Wolf, Lior
Computer Vision and Pattern Recognition
Machine Learning
Vision-language models (VLMs) frequently produce hallucinations in the form of descriptions of objects, attributes, or relations that do not exist in the image due to over-reliance on language priors and imprecise cross-modal grounding. We introduce Spectral Representation Filtering (SRF), a lightweight, training-free method to suppress such hallucinations by analyzing and correcting the covariance structure of the model's representations. SRF identifies low-rank hallucination modes through eigendecomposition of the covariance of the differences between features collected for truthful and hallucinatory captions, revealing structured biases in the feature space. A soft spectral filter then attenuates these modes in the feed-forward projection weights of deeper vLLM layers, equalizing feature variance while preserving semantic fidelity. Unlike decoding or retraining-based approaches, SRF operates entirely post-hoc, incurs zero inference overhead, and requires no architectural modifications. Across three families of VLMs (LLaVA-1.5, MiniGPT-4, and mPLUG-Owl2), SRF consistently reduces hallucination rates on MSCOCO, POPE-VQA, and other visual tasks benchmarks, achieving state-of-the-art faithfulness without degrading caption quality.
title Suppressing VLM Hallucinations with Spectral Representation Filtering
topic Computer Vision and Pattern Recognition
Machine Learning
url https://arxiv.org/abs/2511.12220