concept
active
concept:break-predictor-and-suppressor-featuresBreak predictor and suppressor features
Late features combining characters remaining with next-token length.
Neighborhood — ranked by edge-count
Papers (1)
paper
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.
- Steering validation for antonym mechanism.
- Phenomenon where a feature in a small SAE splits into multiple finer features in a larger SAE.
- Problems suggesting sparsity breaks down as interpretability proxy.
- Splitting the loss into predictive-only and novel-residual terms breaks the competition dynamic.hypothesis0.719Proposed training fix inspired by Kalman filters.
- Framework learning end-to-end mappings from activations through sparse concept bottlenecks to behavioral predictions
- NLI class indicating two sentences are semantically incompatible; weighted +1 in NLI Diversity scoring
- The phenomenon where activating an emotion feature leads to subsequent below-baseline activation of that feature
- Limitation of single-prompt selection.