Taken Out of Context: On Measuring Situational Awareness in LLMs

A preprint on measuring situational awareness in language models. The companion cites its finding that models fine-tuned only on descriptions of a test can pass it without having seen an example. It supports performative ingestion (P13), which prediction 9 tests, and level 5 of the observer-modeling ladder in Table 2.

Cited in

the companion paper (Companion §8, Table 2)

Used for

P13Prediction 9P13; prediction 9; Table 2 (level 5)

Strand

AI safety, alignment & autonomy evaluations · Evaluation awareness, scheming & misalignment

Cite

@misc{berglund2023taken,
  title = {Taken Out of Context: On Measuring Situational Awareness in LLMs},
  author = {Lukas Berglund and Asa Cooper Stickland and Mikita Balesni and Max Kaufmann and Meg Tong and Tomasz Korbak and Daniel Kokotajlo and Owain Evans},
  year = {2023},
  url = {https://arxiv.org/abs/2309.00667},
}