A Lie Detector Test for Language Models: Reading Knowledge a Model Won't Reveal
PIR reads internal model states to detect concealed knowledge, distinguishing sandbagging from genuine ignorance across Gemma, Qwen, Llama, Mistral and Phi models.
Probe of Internal Recognition (PIR) adapts the forensic Concealed Information Test to LLM internals, reading which candidate answer a model recognizes as correct without any honest reference model or labeled truth corpus. Across eight models from five families it recovers concealed answers at 0.70-0.87 balanced accuracy versus a 0.25 chance rate, and remains robust (0.85-0.93) against prompted deception, trained sandbagging, password-locked, and circuit-broken checkpoints. Recognition drops to unknown-question levels only when unlearning actually removes knowledge, enabling sandbagging audits and unlearning verification.
- Reference-free probing separates models that will not answer from those that cannot
- 0.70-0.87 balanced accuracy recovering concealed answers across eight models
- Robust against sandbagging, prompted deception, password locks, and circuit breaking
- Enables sandbagging audits and verification of unlearning claims
Full article254 words · extracted from huggingface.co · click to collapse
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.
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/papers/2609.21996