claim
active
claim:late-layer-features-with-negative-output-weights-to-the-likely-token-may-regulate-output-confidenceLate-layer features with negative output weights to the likely token may regulate output confidence.
Speculative.
Neighborhood — ranked by edge-count
Papers (1)
paper
Concepts (1)
concept
- FeaturesaboutInterpretable concept representations hypothesized as basic units of computation.
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.
- Mechanistic account explaining why late-layer introspection fails, combining two independent explanatory factors
- §5.5.
- Transcoders could cheat Jacobian matching via large-weight, very negative-bias features exploiting the L1/L0 gaphypothesis0.738Gradient masking risk; mitigated by tanh L1 or penalizing Jacobian contribution L1.
- Interpretive claim that circuits render raw weights interpretable as algorithmic structures
- Averaging acts as a low-pass filter that smooths away per-token detail and retains broad conceptsclaim0.736Intuition.
- Example A.
- ReLU hidden layer model circumvents the ReLU by using positive biases or simulating biasfinding0.731Limitation of that toy model.
- Blocking attention to query token in layers 11–16 drops accuracy from 98% to 37%; with next token 8%finding0.730Attention knockout.