Formal Guarantees for Credal Sets

Establish formal guarantees for the credal sets produced by the local relative-likelihood credal predictor that are analogous to the guarantees provided by conformal prediction.

Background

The paper introduces CreWare, a credal predictor that constructs instance-specific sets of plausible conditional probability distributions from local relative likelihoods in a neural network’s embedding space. The method is designed to reflect epistemic uncertainty under limited data, distribution shift, and out-of-distribution inputs, and its empirical coverage and efficiency are evaluated using human soft labels as approximations to ground-truth distributions.

The authors explicitly identify the absence of formal coverage or validity guarantees comparable to those associated with conformal prediction as an unresolved issue. The open problem concerns whether the credal sets generated by this particular local relative-likelihood construction can be theoretically certified, rather than merely assessed empirically.