method
active
method:expected-residual-attribution-eraExpected residual attribution (ERA)
Prior coactivation-based proxy for effectiveness.
Neighborhood — ranked by edge-count
Papers (1)
paper
- Characterizing interference weights in a tiny language modelintroducesmentions
Methods (1)
method
- Fisher effectivenessextendsSecond-order KL estimate of a weight's effect on outputs using Fisher metric.
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.
- ~27,000 point comparison.
- The task of attributing model behaviors to specific training datapoints.
- Weathering, leaning, and environmental adaptation that gives a fence or object more life.
- The mechanism by which LLMs generate text: drawing a token from the next-token distribution and appending it to context repeatedly
- Finding from PRISM dataset analysis showing typicality bias varies by ethnicity and region, with implications for fairness
- Table 2; e.g., Identity +1.8 for turn-averaged.
- Baseline persistence of any probe direction arising from the autoregressive nature of LLMs, not specific to emotion content