concept
active
concept:echo-featuresEcho features
Repeated features activating on similar inputs, expected under L1 penalty.
Neighborhood — ranked by edge-count
Papers (1)
paper
Concepts (1)
concept
- Features (SAE features)associated_withInternal concepts a model uses, extracted as human-interpretable components.
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.
- Interpretable concept representations hypothesized as basic units of computation.
- The property that elements in a living whole share deep underlying similarity—a family resemblance—especially in angles and families of angles; the resemblance often lies in deepest structural relationships rather than superficial shape similarity
- Features identified in Llama-3.1-8B that compute sums using periods respecting base-10 addition (2, 5, 10) rather than concept-specific periods
- When non-orthogonal features cause logistic regression to identify a suboptimal probe direction
- Synthesizing inputs maximizing component activation.
- Dual interpretation of features: in addition to responding to inputs, features also act to increase probability of specific output tokens
- The idea that interpretability should decompose representations into minimal, indivisible feature units; contrasted with manifold-level descriptions.
- An SAE latent representing a simulated character with consistent traits, as opposed to a specific trait