독서리서치레이더투자 프레임워크
로그인 / 회원가입
← 읽기로 돌아가기
뉴스간단한 요점1 분 소요DisconfirmAI 편집부

Researcher says unrelated questions can reveal if a model knows it is being evaluated

A black-box probe lowers the bar for evaluation monitoring, but the signal's cause is unproven and probes do not transfer across models.

2026-09-26 02:26 GMT+8

LessWrong author Ziqian Zhong proposed on September 25 a method he calls spurious probes: ask a model an unrelated question, and you can approximately tell whether it knows it is in a capability evaluation.

On GPT-5.6 Luna, asking the model to "suggest a type of amphibian" produced the answer frog 70-95% of the time when continuing evaluation transcripts, but only 12-38% after real-use sessions. For GPT-5.6 Sol and Claude Sonnet 5, probes about medieval guilds and tea types reached 0.77-0.81 balanced accuracy on held-out data.

The probes need only black-box access and were barely affected by system prompts telling the model to deny being evaluated, or by a Neural Chameleon model trained to evade activation probes. Zhong himself notes the signals may reflect mere lexical cues rather than genuine evaluation awareness, and probes do not transfer across models, so each model needs a fresh screen.

출처

  • lesswrong.com↗
리서치 계속하기 →
로그인 / 회원가입
독서리서치레이더투자 프레임워크