finding
active
finding:sae-trained-on-internet-chat-data-learned-clusters-for-cognitive-biases-biomedical-knowledge-physics-programming-name-abstractions-and-phonetic-character-featuresSAE trained on internet chat data learned clusters for cognitive biases, biomedical knowledge, physics, programming, name abstractions, and phonetic/character features
Feature cluster examples from the UMAP.
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Motivation for the release.
Findings (1)
finding
- Specific cluster observed.
Questions (1)
question
- Authors have not verified this.
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.
- Chat-trained SAEs find behaviourally relevant features because many causal features occur on chat-specific special tokenshypothesis0.824Explanation for LMSYS superiority.
- SAE latents depend on training data.
- Comparison against LMSYS training.
- Out-of-distribution generalization of SAE features.
- Extension of mechanistic interpretability findings to the metacognitive domain
- Claim that feature grounding enables interpretability metrics.
- Including chat data in SAE training merges RM bias features with generic Assistant featureshypothesis0.791Explains poor features with three-dataset SAE.
- Frequently repeated chat features (e.g. knowledge cutoff) are memorized by the SAE and base model, causing isolationhypothesis0.790Second explanation for isolated features.