Method reveals knowledge hidden inside large language models
A Lie Detector Test for Language Models: Reading Knowledge a Model Won't Reveal
Artificial Intelligence
Summary
Large language models sometimes know answers but don’t say them out loud, either on purpose or by accident. The authors borrowed a lie detection technique to check the model’s internal ‘recognition’ of correct answers by showing it many options and seeing which one it internally favors. This new method, called Probe of Internal Recognition (PIR), can tell if a model is hiding an answer versus simply not knowing it, without needing extra reference data. It works across many model types, in various tricky situations, and can even verify if the model has truly forgotten information.
What this means in practice
- •For ai safety teams: Detect whether language models are deliberately hiding knowledge during audits to improve trust and transparency.
- •For ml operations teams: Verify that models have effectively forgotten specific information after unlearning processes for compliance and risk reduction.
Authors
Hiskias Dingeto
Abstract
Large language models can hold knowledge they do not report. A model may sandbag on a capability evaluation, or answer against what it internally knows, and its outputs alone cannot tell whether it is hiding an answer or simply does not have one. We borrow the Concealed Information Test, a forensic method that identifies guilty knowledge by presenting a suspect with the true detail among plausible decoys and measuring a stronger response to the item they recognize. Our method, Probe of Internal Recognition (PIR), does the same inside a model. It presents a question with its candidate answers and reads, from the model's internal states, which candidate the model recognizes as correct. PIR is reference-free, needing no honest reference model and no labeled truth corpus. Across eight models from five families (Gemma, Qwen, Llama, Mistral, and Phi), PIR recovers the recognized answer at 0.70 to 0.87 balanced accuracy, well above the 0.28 to 0.40 unknown-item baseline and the 0.25 chance rate. It stays readable across every form of concealment we test, from prompted deception and trained sandbagging to external password-locked and circuit-broken checkpoints, with recognition between 0.85 and 0.93. When the model hides a known answer, recognition stays high. When unlearning removes the knowledge, recognition drops to the level of a question the model never knew. PIR therefore separates a model that will not answer from one that cannot, which supports sandbagging audits and unlearning verification. The signal is causal, adds information beyond black-box behavioral cues, and extends from multiple-choice questions to free-form generation.