finding
active
finding:attribution-graphs-provide-satisfying-insight-for-about-a-quarter-of-prompts-triedAttribution graphs provide satisfying insight for about a quarter of prompts tried.
Quantifies method limitations.
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Epistemic scoping.
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.
- Mechanistic discrimination of CoT faithfulness.
- Gradient-based technique using SAE features to estimate causal effects on completions; used to corroborate NLA findings.
- Central claim.
- Motivates 14-turn case.
- Shows how VPD-identified subnetworks can be analyzed to reveal interpretable pathways of computation (e.g., gender signal routing, syntactic role detection).
- Stronger version: all cognition attributions rely on observable behavior.
- Turn-level attribution graphs give compact, interpretable explanations of multi-turn model behaviorclaim0.796Case study conclusion.
- Discussion.