arXiv News

First-hand information for everyone

Site

AboutPrivacy PolicyTerms of Use

© 2026 arXiv News

Layers in 70‑billion‑parameter language models can spot lies better when trained with special “knowledge‑denial” prompts | arXiv News

arXiv News

EnglishJapanese

Switch language

EnglishJapanese
Loading account…

First-hand information for everyone

Back to front page
Artificial IntelligenceEnglishPublished

Layers in 70‑billion‑parameter language models can spot lies better when trained with special “knowledge‑denial” prompts

November 22, 2026arXiv: 2611.26545v1

A new study shows that adding prompts that ask a model to deny knowledge it actually possesses helps it detect lies in text better than the usual trick of building a single‑dataset probe. The researchers expanded a published bank of prompts, adding roles that explicitly deny what the model knows, and built a detector that compares middle‑layer activations between honest and deceptive replies. When tested on the Liars’ Bench benchmark, the extended contrast scored an AUROC of 0.916 at the middle layer, outperforming a raw‑activation probe that achieved 0.732. This performance gap of 0.184 was larger than what could be explained by random chance.

The method relies on contrasting model responses to two types of questions: ones the model can answer and ones it cannot, but appending the same scripted denial to both. This fixes the wording of the reply so that the signal reflects whether the model really knows the answer rather than differences in phrasing. Experiments also showed that the detector largely reacts to the wording of the denial; on a slice of the dataset where lies and honest denials share identical text, its performance fell to the random band.

These findings suggest that persona‑based detectors can generalise better across different kinds of lies, but their sensitivity to wording limits their reliability when lies and honest denials are spelled the same.

The study was done on a 70‑billion‑parameter Llama‑3.3‑Instruct model, with tests run at the model’s middle layer and at other layers to confirm stability of the signal.

The extended contrast remains a promising zero‑shot approach to lie detection, yet the dependence on wording highlights the need for more robust methods when deploying such detectors in real‑world settings.

Source paper

The Geometry of Lies: Knowledge and Persona Directions for Detecting Deception in Large Language Models

Artificial Intelligence
View on arXiv