Don't Throw Away Your Beams: Improving Consistency-based Uncertainties in LLMs via Beam Search

2025-12-10

AI Summary

Consistency-based uncertainty quantification (UQ) for short-answer question answering (QA) suffers from unreliable estimates due to high variance and response redundancy when employing polynomial sampling. Method: This work introduces beam search into the consistency-based UQ framework for the first time, enabling generation of high-probability, controllably diverse candidate answer sets. We derive a theoretical lower bound on the beam set’s probability mass, rigorously guaranteeing improved error upper bounds; further, we integrate consistency aggregation with uncertainty calibration to enhance estimation stability. Results: Extensive experiments across six standard QA benchmarks demonstrate state-of-the-art UQ performance: consistency score variance is reduced by up to 42%, mean estimation error decreases by 19.3% relative to polynomial sampling, and results exhibit strong reproducibility.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Reasoning under Uncertainty: Uncertainty Representations
• Search and Optimization: Sampling/Simulation-based Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Web Mining and Content Analysis: Community question answering
• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work

Abstract

Consistency-based methods have emerged as an effective approach to uncertainty quantification (UQ) in large language models. These methods typically rely on several generations obtained via multinomial sampling, measuring their agreement level. However, in short-form QA, multinomial sampling is prone to producing duplicates due to peaked distributions, and its stochasticity introduces considerable variance in uncertainty estimates across runs. We introduce a new family of methods that employ beam search to generate candidates for consistency-based UQ, yielding improved performance and reduced variance compared to multinomial sampling. We also provide a theoretical lower bound on the beam set probability mass under which beam search achieves a smaller error than multinomial sampling. We empirically evaluate our approach on six QA datasets and find that its consistent improvements over multinomial sampling lead to state-of-the-art UQ performance.

Problem

• Improves uncertainty quantification in LLMs via beam search
• Reduces variance in consistency-based uncertainty estimates
• Enhances performance over multinomial sampling in QA tasks

Innovation

• Beam search replaces multinomial sampling for candidate generation
• Beam search reduces variance in uncertainty estimates
• Beam search improves consistency-based uncertainty quantification performance