Think Before You Lie: How Reasoning Leads to Honesty

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Yuan, Ann, Ghandeharioun, Asma, Blum, Carter, Machado, Alicia, Hoffmann, Jessica, Ippolito, Daphne, Wattenberg, Martin, Dixon, Lucas, Filippova, Katja
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866917346743943168
author Yuan, Ann
Ghandeharioun, Asma
Blum, Carter
Machado, Alicia
Hoffmann, Jessica
Ippolito, Daphne
Wattenberg, Martin
Dixon, Lucas
Filippova, Katja
author_facet Yuan, Ann
Ghandeharioun, Asma
Blum, Carter
Machado, Alicia
Hoffmann, Jessica
Ippolito, Daphne
Wattenberg, Martin
Dixon, Lucas
Filippova, Katja
contents While existing evaluations of large language models (LLMs) measure deception rates, the underlying conditions that give rise to deceptive behavior are poorly understood. We investigate this question using a novel dataset of realistic moral trade-offs where honesty incurs variable costs. Contrary to humans, who tend to become less honest given time to deliberate (Capraro, 2017; Capraro et al., 2019), we find that reasoning consistently increases honesty across scales and for several LLM families. This effect is not only a function of the reasoning content, as reasoning traces are often poor predictors of final behaviors. Rather, we show that the underlying geometry of the representational space itself contributes to the effect. Namely, we observe that deceptive regions within this space are metastable: deceptive answers are more easily destabilized by input paraphrasing, output resampling, and activation noise than honest ones. We interpret the effect of reasoning in this vein: generating deliberative tokens as part of moral reasoning entails the traversal of a biased representational space, ultimately nudging the model toward its more stable, honest defaults.
format Preprint
id arxiv_https___arxiv_org_abs_2603_09957
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Think Before You Lie: How Reasoning Leads to Honesty
Yuan, Ann
Ghandeharioun, Asma
Blum, Carter
Machado, Alicia
Hoffmann, Jessica
Ippolito, Daphne
Wattenberg, Martin
Dixon, Lucas
Filippova, Katja
Artificial Intelligence
Computation and Language
Machine Learning
While existing evaluations of large language models (LLMs) measure deception rates, the underlying conditions that give rise to deceptive behavior are poorly understood. We investigate this question using a novel dataset of realistic moral trade-offs where honesty incurs variable costs. Contrary to humans, who tend to become less honest given time to deliberate (Capraro, 2017; Capraro et al., 2019), we find that reasoning consistently increases honesty across scales and for several LLM families. This effect is not only a function of the reasoning content, as reasoning traces are often poor predictors of final behaviors. Rather, we show that the underlying geometry of the representational space itself contributes to the effect. Namely, we observe that deceptive regions within this space are metastable: deceptive answers are more easily destabilized by input paraphrasing, output resampling, and activation noise than honest ones. We interpret the effect of reasoning in this vein: generating deliberative tokens as part of moral reasoning entails the traversal of a biased representational space, ultimately nudging the model toward its more stable, honest defaults.
title Think Before You Lie: How Reasoning Leads to Honesty
topic Artificial Intelligence
Computation and Language
Machine Learning
url https://arxiv.org/abs/2603.09957