claim
active
claim:newline-feature-relies-on-attention-path-for-activation-signal-and-the-tokens-features-weight-negates-it-for-non-newline-tokens-effectiveness-highlights-inhibitionNewline feature relies on attention path for activation signal and the Tokens→Features weight negates it for non-newline tokens; effectiveness highlights inhibition
Circuit interpretation.
Source paper
extracted_fromNicholas L. Turner · Jeffrey Wu · Joshua Batson
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- 271 stratified contexts.
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.
- Refusal mediated by attention.
- Newline-prediction impact correlates with attention-head distraction across inserted sequencesfinding0.760Mechanistic link.
- Illustrative finding from Lindsey et al. 2025 / Hanna and Ameisen 2026 showing intention-like representations in attention streams
- Blocking attention to query token in layers 11–16 drops accuracy from 98% to 37%; with next token 8%finding0.750Attention knockout.
- Steering planned-word features only affects the final token when applied at the newline planning token.finding0.744Planning is localized.
- Structured number feature.
- Motivates nested SAE.
- Mechanism limitation.