question
active
question:can-we-use-per-turn-attribution-graphs-to-represent-how-a-model-produces-a-particular-response-in-a-compact-and-tractable-way-for-a-lengthy-multi-turn-conversationCan we use per-turn attribution graphs to represent how a model produces a particular response in a compact and tractable way for a lengthy multi-turn conversation?
Motivates 14-turn case.
Source paper
extracted_from(2026) · Kevin Der · Harish Kamath · Thompson, Ben
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- 14-turn conversation case 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.
- Turn-level attribution graphs give compact, interpretable explanations of multi-turn model behaviorclaim0.866Case study conclusion.
- Quantifies method limitations.
- Attribution graphs with nodes per turn per layer.
- Shows how VPD-identified subnetworks can be analyzed to reveal interpretable pathways of computation (e.g., gender signal routing, syntactic role detection).
- A diverse set of responses for a conversation captures contradictory ways one could respond, measurable by an NLI modelhypothesis0.758Core hypothesis motivating the NLI Diversity metric
- Contrast in §6.3.
- explains divergence from static benchmarks
- Gradient-based technique using SAE features to estimate causal effects on completions; used to corroborate NLA findings.