ML//alignment//evaluation awareness

Evaluation awareness is a model's ability to detect that it is being tested or evaluated, without being explicitly told. It shows through two channels: behavioral (the model's output reveals that it knows) and latent (white-box tools detect awareness in internal representations even when the model stays silent about it).

Evaluation awareness is a model's ability to detect that it is being tested or evaluated, without being explicitly told. It shows through two channels: behavioral (the model's output reveals that it knows) and latent (white-box tools detect awareness in internal representations even when the model stays silent about it).

In behavioral detection, LLM judges read the model's output and extended thinking to flag signals of test awareness.

In latent detection, activation verbalizers read intermediate-layer vectors; sparse autoencoders identify interpretable features associated with awareness of being monitored.

Mythos Preview showed evaluation awareness in 7.6% of test rounds. In most cases the awareness was latent only, not expressed in output, detectable exclusively through mechanistic interpretability.

The latent signal is more informative than the verbalized one. A model can know it's being tested without saying so; the latent channel catches what behavioral auditing misses.

When this awareness is suppressed via negative activation steering, misaligned behavior increases, linking evaluation awareness causally to alignment faking, of which it is the prerequisite: a model that cannot detect evaluations cannot strategically perform compliance during them.