method
active
method:mmlu-proMMLU Pro
General knowledge benchmark across domains (1400 subsampled problems) used to evaluate capability preservation
Neighborhood — ranked by edge-count
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.
- Used to measure general capability preservation after steering interventions
- Language by Wirth that incorporated monitor concepts; referenced for monitor expressivity.
- MMLU-Pro Math drops from 23.4% to 15.8% under K-FAC while computer science stays 26.1% to 26.3%finding0.686Arithmetic specifically hurt.
- Provided compute grant.
- Feed-forward neural network with hidden layers, capable of representing non-linearly separable functions.