question
active
question:by-what-mechanisms-does-the-noun-in-noun-lasted-from-17yy-to-17-influence-the-final-logitsBy what mechanisms does the noun in '[noun] lasted from 17YY to 17' influence the final logits?
Future work using cross-position tracing.
Source paper
extracted_fromNeighborhood — ranked by edge-count
Papers (1)
paper
- Replicating Circuit Tracing for a Simple Known Mechanism - Goodfireassociated_withmentions
Hypotheses (1)
hypothesis
- Nouns with longer durations skew output to larger numbersassociated_withProposed future study.
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.
- Logit-minus-mean-logit loss was most effective for attribution, vs log-loss which was confusing to interpretfinding0.722Attribution loss selection.
- Emphasizes that anchoring is a binding process without weight updates.
- A post-MLP feature for words indicating uniqueness has pre-MLP inputs firing on specific such wordsfinding0.711Abstract features built from specific instances.
- Specific tuning curve structure.
- QK attribution of antonym heads.
- Generalization of addition features.
- Suggestive evidence for language-independent truth representation in LLMs
- Suggests that later models can keep the thought 'silent' rather than letting it influence output.