claim
active
claim:spd-bridges-causal-mediation-analysis-and-network-decompositionSPD bridges causal mediation analysis and network decomposition
Framing.
Source paper
extracted_from(2025) · Lucius Bushnaq · Dan Braun · Lee Sharkey
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Multi-layer mechanisms.
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.
- Framework for evaluating whether probe directions are causally implicated in model outputs via activation patching
- Neuron contributions of SPD subcomponents closely match target model across all 100 featuresfinding0.737Figure 7.
- Whether an internal direction causally controls a target behavior, verified by intervention success
- Additional synthetic example of pernicious divergence from balanced subspaces
- SPD splits W_out into 50 subcomponents forming a single rank-50 component in compressed computation modelfinding0.731Contradicts APD's 100 components for W_out.
- The extent to which a probe direction, when intervened upon, actually changes model outputs — contrasted with mere classification accuracy
- Figure 10 discrepancy.
- SPD decomposes W_in of compressed computation model into 100 subcomponents, one per input featurefinding0.727Neuron contributions match target.