claim
active
claim:turn-level-attribution-graphs-give-compact-interpretable-explanations-of-multi-turn-model-behaviorTurn-level attribution graphs give compact, interpretable explanations of multi-turn model behavior
Case study conclusion.
Source paper
extracted_from(2026) · Kevin Der · Harish Kamath · Thompson, Ben
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (4)
finding
- Edge weight sign matches observed intervention effect ~100% of the time, minimum 95% on any graphsupports36,000 interventions.
- 14-turn conversation case study.
- Table 2; e.g., Identity +1.8 for turn-averaged.
- Figure 6.
Questions (1)
question
- Motivates case studies.
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.
- Motivates 14-turn case.
- Attribution graphs with nodes per turn per layer.
- Quantifies method limitations.
- Mechanistic discrimination of CoT faithfulness.
- explains divergence from static benchmarks
- Based on sufficiency scores.
- Gradient-based technique using SAE features to estimate causal effects on completions; used to corroborate NLA findings.
- Shows how VPD-identified subnetworks can be analyzed to reveal interpretable pathways of computation (e.g., gender signal routing, syntactic role detection).