finding
active
finding:blocking-attention-to-query-token-in-layers-11-16-drops-accuracy-from-98-to-37-with-next-token-8Blocking attention to query token in layers 11–16 drops accuracy from 98% to 37%; with next token 8%
Attention knockout.
Source paper
extracted_from(2025) · Yoav Gur-Arieh · Mor Geva · Atticus Geiger
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Central thesis.
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.
- Reflexive signal retrieved from entity tokens.
- Lexical signal local.
- Evidence against test-time compute hypothesis.
- Layer localization.
- Demonstrates reflection redundancy in stronger model on harder math benchmark
- Structure of the computational pathway.
- A pair of query and key subcomponents distributed across attention heads performs previous-token behaviorfinding0.767VPD recovers an attention algorithm for attending to the previous token, distributed across multiple heads.
- Striking mechanistic finding that injection creates universally detectable perturbation in residual stream immediately downstream