Assess generalization of soft-label fine-tuning

Determine whether the reported soft-label accuracy gains generalize across classifier architectures and the full set of moral-foundation labels.

Background

The paper demonstrates consistent 2–3% accuracy gains from training classifiers on calibrated posterior probabilities rather than collapsed voting labels, but the experiment covers only a subset of the available corpora and foundations and uses a single classifier architecture. Whether the improvement persists across architectures and across the complete label set is explicitly unresolved.

References

The soft-label accuracy gains reported above are demonstrated on a subset of corpora/foundations and a single classifier architecture; generalization across architectures and the full label set is untested and left to future work.

— Moral Entropy: Auditing Bias and Uncertainty in Moral Judgment  (2609.21992 - Skorski, 18 Sep 2026) in Limitations, paragraph “Scope of the fine-tuning result”