finding
active
finding:turn-averaged-attribution-subgraph-traces-degenerate-output-to-jailbreak-detection-feature-at-l16-and-nuclear-weapons-ai-safety-features-in-earlier-turnsTurn-averaged attribution subgraph traces degenerate output to jailbreak-detection feature at L16 and nuclear weapons/AI safety features in earlier turns
14-turn conversation case study.
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
- Motivates 14-turn case.
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.
- Structure of the computational pathway.
- Demonstrates Assistant attractor dynamics in practice
- Attribution graphs with nodes per turn per layer.
- Turn-averaged features have highest embedding similarity 0.663 at k=3; per-token max lowest at 0.617finding0.756Consistent across k=1,2,3,5.
- Based on sufficiency scores.
- Feature attribution (gradient-based) correlates 0.8 with ablation effects on the 'John' and 'Kobe' examples.finding0.750Validation of attribution as a fast proxy for causal importance.
- Example A.
- Explains discrepancy.