finding
active
finding:ablating-j-broadcast-heads-reduces-recall-25-from-0-86-to-0-67-and-injected-thought-report-rate-from-0-54-to-0-09Ablating J broadcast heads reduces recall@25 from 0.86 to 0.67 and injected-thought report rate from 0.54 to 0.09.
Broadcast heads causal.
Source paper
extracted_fromAuthors
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- MLP and attention analysis.
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.
- Random latent ablation produces slight increase in ESR rate (3.8% to 4.2%), not statistically significantfinding0.756Control result confirming OTD ablation effect is specific to those latents, not a general ablation artifact
- Contents.
- Causal.
- Ablation depends on t_entity.
- Nonspecific to self.
- Feature attribution (gradient-based) correlates 0.8 with ablation effects on the 'John' and 'Kobe' examples.finding0.735Validation of attribution as a fast proxy for causal importance.
- Justifies SCR/TPP.
- Ablation.