Verbalised evaluation awareness in large reasoning models has only small effects on their outputs across safety and alignment tests.
lesswrong.com/posts/jsFGuXDMxy5NZg9T2/prefill-awareness-can-llms-tell- when-their-message-history(visited on 05/03/2026)
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Evaluation Awareness in Language Models Has Limited Effect on Behaviour
Verbalised evaluation awareness in large reasoning models has only small effects on their outputs across safety and alignment tests.