claim
active
claim:gradient-attribution-gives-biased-causal-importance-estimates-spd-s-direct-optimization-is-likely-betterGradient attribution gives biased causal importance estimates; SPD's direct optimization is likely better
Motivation.
Source paper
extracted_from(2025) · Lucius Bushnaq · Dan Braun · Lee Sharkey
Neighborhood — ranked by edge-count
Papers (1)
paper
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.
- SPD's robustness comes from trained importance functions, gradient flow, no simplicity loss, and direct causal importancehypothesis0.836Four hypothesized sources.
- Some gradient-based attribution methods give attributions independent of model and data-generating processfinding0.826Adebayo et al. sanity checks.
- Baseline method against which probe-based ranking is compared; more computationally expensive.
- Used by APD; poor approximations of causal importance.
- Backprop of logit-minus-mean-logit loss to surface causal intervention candidates, summed across token positions.
- SPD avoids shrinkage because importance norm penalizes mask probability, not singular valuesclaim0.779Mechanism.
- Authors' central interpretive assertion that their method meaningfully mitigates unwanted behaviors.
- Gradient-based attribution approximates ablation impact, enabling fast search for causally important features.