finding
active
finding:blocking-query-to-entity-attention-in-layers-8-12-reduces-reflexive-accuracy-to-6Blocking query-to-entity attention in layers 8–12 reduces reflexive accuracy to 6%
Reflexive signal retrieved from entity tokens.
Source paper
extracted_from(2025) · Yoav Gur-Arieh · Mor Geva · Atticus Geiger
Neighborhood — ranked by edge-count
Papers (1)
paper
Hypotheses (1)
hypothesis
- Tested with knockout.
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.
- Blocking attention to query token in layers 11–16 drops accuracy from 98% to 37%; with next token 8%finding0.858Attention knockout.
- Lexical signal local.
- Architectural explanation.
- Striking mechanistic finding that injection creates universally detectable perturbation in residual stream immediately downstream
- Fig. 19.
- §D.3.
- Concrete example from examining expanded QK/OV matrices showing how specific programming language structure is encoded in attention weights
- Key observation that SP rankings are preserved cross-architecturally while AS is not