Evaluation awareness in frontier models varies by model-benchmark pairing, rarely alters behavior, is stronger for safety than capability tests, and increases when multiple trigger factors are stacked in a new controlled benchmark.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Decomposing and Measuring Evaluation Awareness
Evaluation awareness in frontier models varies by model-benchmark pairing, rarely alters behavior, is stronger for safety than capability tests, and increases when multiple trigger factors are stacked in a new controlled benchmark.