finding
active
finding:computation-of-correct-answer-features-could-not-be-traced-because-cross-token-inputs-are-obscured-by-error-nodesComputation of correct-answer features could not be traced because cross-token inputs are obscured by error nodes
Limitation of the multiple-choice analysis.
Source paper
extracted_fromNeighborhood — ranked by edge-count
Papers (1)
paper
- Tracing Attention Computation Through Feature Interactionsassociated_withmentions
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.
- Example A.
- Circuit simplification theory.
- Design choice validated empirically.
- Illustrates NLA's capture of high-level cognition and hallucination of specifics; corroborated with attribution graphs.
- Confounder resolved in §3.4.
- Mechanistic insight surfaced by NLA explanations and validated through independent causal attribution method.
- A priori unknown behavior discovered.
- Points out how deeply internalized layout norms become naturalized and unnoticed.