Grounding the Ungrounded: A Spectral-Graph Framework for Quantifying Hallucinations in Multimodal LLMs

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Sarkar, Supratik, Das, Swagatam
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909952559284224
author Sarkar, Supratik
Das, Swagatam
author_facet Sarkar, Supratik
Das, Swagatam
contents Hallucinations in LLMs--especially in multimodal settings--undermine reliability. We present a rigorous information-geometric framework, grounded in diffusion dynamics, to quantify hallucinations in MLLMs where model outputs are embedded via spectral decompositions of multimodal graph Laplacians, and their gaps to a truth manifold define a semantic distortion metric. We derive Courant-Fischer bounds on a temperature-dependent hallucination profile and use RKHS eigenmodes to obtain modality-aware, interpretable measures that track evolution over prompts and time. This reframes hallucination as quantifiable and bounded, providing a principled basis for evaluation and mitigation.
format Preprint
id arxiv_https___arxiv_org_abs_2508_19366
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Grounding the Ungrounded: A Spectral-Graph Framework for Quantifying Hallucinations in Multimodal LLMs
Sarkar, Supratik
Das, Swagatam
Machine Learning
Artificial Intelligence
53B21, 46E22 (Primary), 68R10 (Secondary)
Hallucinations in LLMs--especially in multimodal settings--undermine reliability. We present a rigorous information-geometric framework, grounded in diffusion dynamics, to quantify hallucinations in MLLMs where model outputs are embedded via spectral decompositions of multimodal graph Laplacians, and their gaps to a truth manifold define a semantic distortion metric. We derive Courant-Fischer bounds on a temperature-dependent hallucination profile and use RKHS eigenmodes to obtain modality-aware, interpretable measures that track evolution over prompts and time. This reframes hallucination as quantifiable and bounded, providing a principled basis for evaluation and mitigation.
title Grounding the Ungrounded: A Spectral-Graph Framework for Quantifying Hallucinations in Multimodal LLMs
topic Machine Learning
Artificial Intelligence
53B21, 46E22 (Primary), 68R10 (Secondary)
url https://arxiv.org/abs/2508.19366