finding
active
finding:features-for-special-formatting-tokens-and-repetitive-chat-elements-appear-as-isolated-points-or-small-clusters-away-from-the-central-umap-componentFeatures for special formatting tokens and repetitive chat elements appear as isolated points or small clusters away from the central UMAP component
Observation on the Llama 3.3 70B SAE feature map.
Neighborhood — ranked by edge-count
Papers (1)
paper
Hypotheses (2)
hypothesis
- Second explanation for isolated features.
- First 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.
- Similar number and nature of exclusive features across models results from oversampling of Human/Assistant chat transcriptshypothesis0.738Motivated by Bricken et al; tested and not supported.
- Chat-trained SAEs find behaviourally relevant features because many causal features occur on chat-specific special tokenshypothesis0.733Explanation for LMSYS superiority.
- Dimensionality reduction of SAE decoder vectors to create interactive feature maps.
- From Lindsey et al; modeled with co-activation toy.
- Interesting special case of copying behavior related to tokenization artifacts; primitive precursor to induction heads
- 2D embedding of feature direction vectors used to visualize feature clusters and splitting geometry
- Discovery story.
- Third empirical pattern.