concept
active
concept:action-featuresAction Features
Dual interpretation of features: in addition to responding to inputs, features also act to increase probability of specific output tokens
Neighborhood — ranked by edge-count
Concepts (1)
concept
- Token-in-Context Featureassociated_withFeature that fires on a specific token only within a specific surrounding context (e.g., 'the' in physics vs 'the' in mathematics)
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.
- Changing configuration to sample environment differently; minimizes free energy.
- Interpretable concept representations hypothesized as basic units of computation.
- Choice of policies minimizing expected free energy to realize preferred future states.
- Metaphor treating each system feature or function as a separate application that can be independently loaded and managed.
- Synthesizing inputs maximizing component activation.
- Proposed attention analog of sparse monosemantic features for decomposing attention layers.
- Internal concepts a model uses, extracted as human-interpretable components.
- Extra feature representing unexplained error to make isomorphism exact.