Log odds-ratio

Primary evaluation metric measuring causal effect of interventions; greater value indicates larger causal effect

Neighborhood — ranked by edge-count

framework

CausalGym
uses
Multi-task benchmark of linguistic behaviours for measuring causal efficacy of interpretability methods, adapted from SyntaxGym

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

log-evidence (marginal likelihood)concept0.762
Model evidence approximated by free energy bound, used for model selection.
logarithm transformationmethod0.737
Parameter-free loss transformation applied to each task loss to equalize scales
log(x) = min_s (e^s * x - s - 1) for x > 0finding0.702
Mathematical equivalence showing logarithm transformation recovers IMTL-L in the limit
marginconcept0.700
The outer edge space that protects the text block but can be invaded by marginalia.
Self-Correction Data Mixing Ratioconcept0.698
Experimental variable sweeping proportion of self-correction to normal response training examples from 10% to 90%
probabilityconcept0.696
Measure of expected sensory input, core to linking value and surprise.
The additive form S = ρd - dr - log k is parsimonious and aligns with log-odds intuitionclaim0.686
Justification for the linear combination
Alignmentconcept0.680
The goal of making model behavior match human values and intentions, often addressed during post-training.