finding
active
finding:edge-weight-sign-matches-observed-intervention-effect-100-of-the-time-minimum-95-on-any-graphEdge weight sign matches observed intervention effect ~100% of the time, minimum 95% on any graph
36,000 interventions.
Source paper
extracted_from(2026) · Kevin Der · Harish Kamath · Thompson, Ben
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Turn-level attribution graphs give compact, interpretable explanations of multi-turn model behaviorsupportsCase study conclusion.
Questions (1)
question
- Edge weights first-order approximation.
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.
- Figure 6.
- Confirms attention rather than MLP as the locus of persona generation
- Most style interventions have strong off-target effects on non-targeted formatting attributesfinding0.728Fig. 9.
- Quantitative evidence that stress map does not reliably predict the target pattern.
- Quantifies method limitations.
- Test-time probe metrics.
- Core result showing MM is superior to LR for causal implication despite similar classification accuracy
- Novelty claim.