concept
active
concept:memorization-of-frequent-featuresMemorization of frequent features
Possible explanation for isolated repetitive chat features.
Neighborhood — ranked by edge-count
Papers (1)
paper
Hypotheses (1)
hypothesis
- Second explanation for isolated features.
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.
- Property that features activate on only a small fraction of inputs; enables compressed sensing and is what allows superposition to work
- Demonstrates mechanistic memorization via feature assemblies in superposition
- Features with similar decoder magnitude across both models.
- Tasks arranged between pure recitation and pure reasoning.
- Method of optimizing input to cause a neuron to fire maximally, used to characterize what a neuron detects; establishes causal link
- Decoder cosine similarity maps onto concept similarity.
- Proposed attention analog of sparse monosemantic features for decomposing attention layers.
- The central topic, treated as a dynamic, agential process rather than static storage.