FUSE: Ensembling Verifiers with Zero Labeled Data

The real buried lede is the dependency problem: naive ensembles of verifiers assume independence, which inflates confidence when verifiers share the same failure modes. FUSE's spectral approach explicitly models those correlations, which is what lets it work without ground-truth labels rather than just hoping the errors cancel out.

This connects directly to the LLM judge reliability work covered here in mid-April ('Diagnosing LLM Judge Reliability: Conformal Prediction Sets and Transitivity Violations'), which found that aggregate consistency scores mask per-instance logical failures in pairwise comparisons. FUSE is essentially attacking the same problem from the other direction: rather than diagnosing individual verifier unreliability after the fact, it tries to hedge against it structurally at ensemble construction time. The SpecGuard paper from the same period ('Verification-Aware Speculative Decoding') is also relevant context, since it treats verification as a first-class inference concern rather than a post-hoc check. Together these papers suggest verification is quietly becoming its own subfield within LLM infrastructure, distinct from the core modeling work.

Watch whether FUSE's gains hold when verifiers are drawn from the same base model family rather than architecturally diverse sources. If performance degrades significantly in that homogeneous setting, the spectral dependency correction may be doing less work than the diversity of the ensemble itself.