method
active
method:gradient-based-data-attributionGradient-based data attribution
Baseline method against which probe-based ranking is compared; more computationally expensive.
Neighborhood — ranked by edge-count
Findings (1)
finding
- Primary quantitative result: probe method outperforms gradient-based and LLM-judge alternatives at lower computational cost.
Methods (2)
method
- Backprop of logit-minus-mean-logit loss to surface causal intervention candidates, summed across token positions.
- Gradient-based attribution methodsrelated_toUsed by APD; poor approximations of causal importance.
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.
- Some gradient-based attribution methods give attributions independent of model and data-generating processfinding0.893Adebayo et al. sanity checks.
- Used for updating hidden state expectations; provides dynamical process theory testable against neuronal data
- The task of attributing model behaviors to specific training datapoints.
- The property that qualities vary slowly, subtly, gradually across the extent of each living thing; gradients arise as natural responses to changing circumstances and create field-like character that points toward and establishes centers
- Linear classifier approach applied to model activations to identify which training datapoints caused undesired behaviors in post-training.
- Motivation.
- Gradient that tells a cell its correct position; stress arises from deviation from this gradient.
- Optimization technique that computes weight changes by following the gradient of an error function; contrasted with evolutionary stochastic search.