A Lie Detector Test for Language Models: Reading Knowledge a Model Won't Reveal
PIR reads internal states to separate concealed model knowledge from genuine ignorance across eight language models.
Probe of Internal Recognition (PIR) adapts the forensic Concealed Information Test to language models, reading internal states to see which candidate answer a model recognizes. It is reference-free, requiring neither an honest reference model nor a labeled truth corpus. Across eight models from Gemma, Qwen, Llama, Mistral, and Phi, it recovers the recognized answer at 0.70 to 0.87 balanced accuracy, above a 0.28 to 0.40 unknown-item baseline and 0.25 chance. Recognition remains 0.85 to 0.93 under prompted deception, trained sandbagging, and password-locked or circuit-broken checkpoints, but drops to the unknown-item level when unlearning removes the knowledge.
- PIR reads internal states to identify which candidate a model recognizes as correct.
- The probe needs no honest reference model and no labeled truth corpus.
- Balanced accuracy is 0.70 to 0.87 across eight models from five families.
- Recognition stays 0.85 to 0.93 under deception, sandbagging, and locked checkpoints.
- After unlearning removes knowledge, recognition falls to the unknown-item baseline.
From the original article
1 min read at arxiv.orgLarge 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.
By Hiskias Dingeto · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.