FUSE: Ensembling Verifiers with Zero Labeled Data

2026-04-20

AI Summary

This work addresses the critical challenge of effectively integrating multiple imperfect large language model verifiers to enhance output validation quality in the absence of ground-truth labels. The authors propose FUSE, a novel method that achieves, for the first time, fully unsupervised and efficient multi-verifier integration, thereby eliminating the reliance on labeled data inherent in conventional semi-supervised approaches. Built upon a phylogenetic ensemble framework, FUSE explicitly models the conditional dependencies among verifiers and incorporates an unsupervised score fusion mechanism, substantially improving verification accuracy. Experimental results demonstrate that FUSE matches or even surpasses the performance of semi-supervised methods across several challenging benchmarks, including GPQA Diamond, Humanity's Last Exam, and the IMO Shortlist.

Technology Category

• Natural Language Processing: Language Grounding & Multi-modal NLP
• Machine Learning: Large Multimodal Models (LMMs)
• Intelligent Robots: Multimodal Perception & Sensor Fusion

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work

Abstract

Verification of model outputs is rapidly emerging as a key primitive for both training and real-world deployment of large language models (LLMs). In practice, this often involves using imperfect LLM judges and reward models since ground truth acquisition can be time-consuming and expensive. We introduce Fully Unsupervised Score Ensembling (FUSE), a method for improving verification quality by ensembling verifiers without access to ground truth correctness labels. The key idea behind FUSE is to control conditional dependencies between verifiers in a manner that improves the unsupervised performance of a class of spectral algorithms from the ensembling literature. Despite requiring zero ground truth labels, FUSE typically matches or improves upon semi-supervised alternatives in test-time scaling experiments with diverse sets of generator models, verifiers, and benchmarks. In particular, we validate our method on both conventional academic benchmarks such as GPQA Diamond and on frontier, unsaturated benchmarks such as Humanity's Last Exam and IMO Shortlist questions.

Problem

• verification
• ensembling
• zero labeled data
• large language models
• unsupervised

Innovation

• unsupervised ensembling
• verifier integration
• zero-label verification
• spectral algorithms
• LLM output validation