Number 4 of 12 · November 21, 2026

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

Simple directions inside an SI model's internal activity can help spot some lies, but the results vary a lot by model and by type of lie, and often track wording more than truth.

What they found

The authors test cheap lie detectors for large language models. These detectors are directions in the model's internal activity, found without labeled lies. One comes from honest and deceptive role-play (a persona direction). Another contrasts questions the model can answer with ones nobody can (a knowledge direction). At each model's middle layer, the persona direction ranks lies across Liars' Bench better than single-dataset probes on three of four models. It fails on Qwen. On lies about knowledge, where a lie and an honest denial can use identical words, persona directions mostly track wording. The knowledge direction separates lies from identical honest denials on two models, narrowly on a third, and not on the fourth. The sum of both directions ranks lies better than a supervised probe at the same depth on all four models, though on one both are near chance. More varied role-play questions did not help the persona direction on Liars' Bench.

Why it matters

Spotting when an SI system says something it knows is false is a hard, open problem. Cheap probes would be easy to run, but they often fail on new kinds of lie. This paper gives a careful, mixed picture of where they work and where they do not. It also shows that gains can sit within the range of random directions, which is a useful check for this kind of work.

How it changes the race

This is incremental. It offers a low-cost way to probe for deception on some models, but the results are inconsistent across models, so it is not a general tool yet.

Safety and control

The paper bears directly on detecting deception in SI models, which supports human oversight. The abstract shows these detectors are unreliable across models and lie types, and they should not be trusted alone. On lies about knowledge, several results stay inside the random-direction band.

The abstract, by the authors

Linear probes are the cheapest LLM lie detectors but often fail to transfer across kinds of lie. Label-free directions from honest and deceptive role-play are untested at 70B scale and on lies about knowledge, where a lie and an honest denial can be identical. We widen a published persona bank with roles that deny knowledge they have, plus honest counterparts. At each model's middle layer the resulting persona direction ranks lies across Liars' Bench better than single-dataset probes on three of its four models; it fails on Qwen, even when Qwen stays in role far more often. On Llama-3.3-70B-Instruct it beats directions from a random persona bank, but its gain over the published direction is within the range of random widenings. On lies about knowledge, persona directions mainly track a reply's wording: where lies and honest denials share their words, our three main contrasts stay inside the random-direction band on Llama and Qwen and reach only its edge on Gemma, though on Mistral one clears it. A knowledge direction contrasts questions the model can answer with ones nobody can, each followed by the same scripted denial. It separates lies from identical honest denials on two models and narrowly on a third, but not on the fourth; its other results partly depend on the benchmark's wording. On the Gender-Secret fine-tune it clears the band by a few hundredths on Llama and fails on Mistral. The sum of the two directions ranks lies better than a supervised probe at the same depth on all four models (on one, both near chance) but is level with the persona direction on Llama's average and behind that probe on held-out lies. More varied role-play questions do not help the persona direction on Liars' Bench.

Quoted from arXiv:2611.26545. The research and its words belong to its authors.

The authors

Md. Abdullah Al Zami, Suraiya Shanaz, Mahdi Hasan, Shoaib Ahmed Dipu, Farig Yousuf Sadeque, Md. Sabbir Ahmed

Credit for this research belongs to its authors. SI MIND's page is a plain-English reading of their abstract on arXiv, not a review of the full paper.