Saved in:
Bibliographic Details
Main Authors: O'Neill, Charles, Chalnev, Slava, Zhao, Chi Chi, Kirkby, Max, Jayasekara, Mudith
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2507.23221
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913967482339328
author O'Neill, Charles
Chalnev, Slava
Zhao, Chi Chi
Kirkby, Max
Jayasekara, Mudith
author_facet O'Neill, Charles
Chalnev, Slava
Zhao, Chi Chi
Kirkby, Max
Jayasekara, Mudith
contents Contextual hallucinations -- statements unsupported by given context -- remain a significant challenge in AI. We demonstrate a practical interpretability insight: a generator-agnostic observer model detects hallucinations via a single forward pass and a linear probe on its residual stream. This probe isolates a single, transferable linear direction separating hallucinated from faithful text, outperforming baselines by 5-27 points and showing robust mid-layer performance across Gemma-2 models (2B to 27B). Gradient-times-activation localises this signal to sparse, late-layer MLP activity. Critically, manipulating this direction causally steers generator hallucination rates, proving its actionability. Our results offer novel evidence of internal, low-dimensional hallucination tracking linked to specific MLP sub-circuits, exploitable for detection and mitigation. We release the 2000-example ContraTales benchmark for realistic assessment of such solutions.
format Preprint
id arxiv_https___arxiv_org_abs_2507_23221
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle A Single Direction of Truth: An Observer Model's Linear Residual Probe Exposes and Steers Contextual Hallucinations
O'Neill, Charles
Chalnev, Slava
Zhao, Chi Chi
Kirkby, Max
Jayasekara, Mudith
Machine Learning
Contextual hallucinations -- statements unsupported by given context -- remain a significant challenge in AI. We demonstrate a practical interpretability insight: a generator-agnostic observer model detects hallucinations via a single forward pass and a linear probe on its residual stream. This probe isolates a single, transferable linear direction separating hallucinated from faithful text, outperforming baselines by 5-27 points and showing robust mid-layer performance across Gemma-2 models (2B to 27B). Gradient-times-activation localises this signal to sparse, late-layer MLP activity. Critically, manipulating this direction causally steers generator hallucination rates, proving its actionability. Our results offer novel evidence of internal, low-dimensional hallucination tracking linked to specific MLP sub-circuits, exploitable for detection and mitigation. We release the 2000-example ContraTales benchmark for realistic assessment of such solutions.
title A Single Direction of Truth: An Observer Model's Linear Residual Probe Exposes and Steers Contextual Hallucinations
topic Machine Learning
url https://arxiv.org/abs/2507.23221