AURORA: Asymmetry and Update-Induced Rotation for Robust Hallucination Detection in Large Language Models

2026-06-28

AI Summary

This work addresses the challenge of hallucination detection in large language models, particularly in high-stakes scenarios where existing methods suffer from poor generalization across datasets. The authors propose AURORA, a novel framework that shifts the paradigm from static representation-based analysis to dynamic examination of gradient updates during inference. By characterizing the distinct weight-gradient patterns induced by faithful versus hallucinated responses, AURORA introduces two new features: the skewness of the cosine similarity distribution between weights and gradients, and a singular vector basis rotation ratio derived from singular value decomposition (SVD). Evaluated across four model families and four benchmarks, the method demonstrates robust detection performance and exhibits strong transferability and scalability to out-of-domain tasks, including mathematical reasoning and vision-language settings.

Abstract

Large Language Models (LLMs) have demonstrated remarkable capabilities across a wide range of natural language processing tasks. However, their tendency to generate hallucinations, namely factually incorrect or unfaithful outputs, poses a critical obstacle to their deployment in high-stakes applications. Although recent hallucination detection methods have made encouraging progress, they typically rely on costly output-level consistency checks or static hidden-state probes that capture shallow dataset-specific patterns, leading to substantial degradation under cross-dataset evaluation. In this work, we propose AURORA, a novel hallucination detection framework that shifts the focus from static representations to the weight-gradient dynamics of LLMs. Our key insight is that hallucinated and faithful answers induce qualitatively different gradient update patterns on the model's parameters. Specifically, hallucinated samples trigger asymmetric and structurally misaligned gradients, which can be captured through two complementary features: (1) the skewness of the cosine similarity distribution between weight matrices and their gradient update directions, and (2) the rotation ratio, which quantifies how much the gradient update reorients the singular-vector basis of weight matrices via SVD. AURORA achieves strong hallucination detection performance across four model families and four benchmark datasets. Further analyses demonstrate that our method scales effectively across model sizes and transfers to out-of-domain tasks, including mathematical reasoning and vision-language scenarios.

Problem

• hallucination detection
• large language models
• cross-dataset generalization
• model robustness
• factual consistency

Innovation

• gradient dynamics
• asymmetry
• rotation ratio
• hallucination detection
• SVD-based analysis