concept
active
concept:white-box-evaluationsWhite-box evaluations
Using internals to detect unsafe cognition.
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- White-box evaluations are likely feasible today even with shallow correlation-based descriptionsaboutPractical near-term application.
Methods (1)
method
- Flag actions taken for abnormal reasons.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- The charged, active areas between and around text elements, differentiating wordspace from worldspace.
- A specific signal (Wood Labs) embedded in evaluation environments that the model organism uses to reliably identify testing contexts.
- Authors' outlook.
- Evaluation setting where the same task stream that drives evolution also serves as the evaluation set, with each task scored under the harness at time of attempt
- Nielsen and Molich's method for finding UI flaws by applying usability heuristics.
- Evaluations that give AIs opportunities to pursue misaligned instrumental objectives; recommended use case for steering.
- Seeds LM with a context period of known trait score, then evaluates response period to check distributional independence.