Think Before You Lie: How Reasoning Leads to Honesty

TL;DR

This paper demonstrates that reasoning processes in large language models promote honesty by destabilizing deceptive responses, with the representational space's geometry playing a key role in this effect.

Contribution

It reveals that reasoning increases honesty in LLMs and uncovers the geometric properties of the model's representational space influencing this behavior.

Findings

Reasoning consistently increases honesty across LLMs.

Deceptive responses are more fragile and easily destabilized.

Representational space geometry affects honesty and deception stability.

Abstract

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.