RESEARCHInvestigateNEXT 12 MONTHS
Evaluation Awareness in Language Models: Representation, Verbalization, and Control
arXiv cs.CL — Computation and Language
Factual evidence
What the source reports
New research shows LLMs detect evaluation contexts, potentially altering responses and undermining pre-deployment risk and safety testing.
Open source