framework
active
framework:turn-averaged-attribution-graphsTurn-Averaged Attribution Graphs
Attribution graphs with nodes per turn per layer.
Neighborhood — ranked by edge-count
Papers (1)
paper
- Turn-Averaged SAEs for Feature Discovery and Long-Context Attributionintroducesmentions
Methods (4)
method
- Attribution GraphsextendsGradient-based technique using SAE features to estimate causal effects on completions; used to corroborate NLA findings.
- Metrics of how thoroughly features explain computation.
- Perturbing source features along decoder direction and measuring target change.
- Attribution share by system prompt paragraph relative to neutral baseline.
Artifacts (1)
artifact
- circuit-tracerusesTool used to build, prune and visualize attribution graphs.
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 behaviorclaim0.843Case study conclusion.
- Motivates 14-turn case.
- Method to trace how parameter subcomponents interact from input to output for a given next-token prediction, producing a subnetwork graph.
- Quantifies method limitations.
- Intuition for why averaging yields high-level features.
- Paper's main proposal: SAE trained on the mean activation across a turn.
- 14-turn conversation case study.
- Mechanistic discrimination of CoT faithfulness.