Large language models often retain knowledge they do not reveal in their outputs; a model may deliberately withhold an answer during evaluation or simply lack the information, making it hard to tell whether silence means concealment or ignorance. Inspired by the forensic Concealed Information Test, which embeds the true detail among plausible decoys and measures a stronger response to the recognized item, we introduce Probe of Internal Recognition (PIR). PIR presents a multiple‑choice question to the model, reads its internal states, and identifies which candidate elicits the strongest recognition signal. The method requires no reference model or labeled truth corpus, operating entirely reference‑free. Across eight models from five families (Gemma, Qwen, Llama, Mistral, Phi), PIR achieves balanced accuracy of 0.70‑0.87, far above the unknown‑item baseline of 0.28‑0.40 and the 0.25 chance level. It remains effective under various concealment strategies, including prompted deception, trained sandbagging, and external password‑locked or circuit‑broken checkpoints, with recognition rates between 0.85 and 0.93. When a model hides a known answer, recognition stays high; when unlearning removes the knowledge, recognition drops to the level of a never‑known question. Thus PIR separates “won’t answer because of refusal” from “won’t answer because of lack of knowledge,” supporting sandbagging audits and unlearning verification. The signal is causal, adds information beyond black‑box behavior, and extends from multiple‑choice to free‑form generation.
Review