Estimating LLM Consistency: A User Baseline vs Surrogate Metrics

2025-05-26

AI Summary

Existing automatic consistency metrics—such as those based on response probabilities or internal model states—exhibit substantial misalignment with human subjective judgments of LLM-generated text consistency. Method: This work introduces the first large-scale human consistency scoring benchmark (n = 2,976) to systematically quantify the gap between prevailing automated metrics and ground-truth human perception; it further proposes a novel logit-weighted ensemble method that enhances consistency quantification via response resampling and logit-space modeling. Contribution/Results: Empirical evaluation demonstrates that the proposed method achieves state-of-the-art performance in Spearman and Kendall rank correlation with human scores, significantly outperforming conventional metrics. The study confirms that automated consistency measures—when decoupled from human evaluation—suffer from pervasive miscalibration, underscoring the urgent need to anchor evaluation paradigms in human perception.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Large Multimodal Models (LMMs)
• Humans and AI: Learning Human Values and Preferences

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work
• User Modeling, Personalization and Recommendation: Metrics for user behavior and evaluating success

Abstract

Large language models (LLMs) are prone to hallucinations and sensitive to prompt perturbations, often resulting in inconsistent or unreliable generated text. Different methods have been proposed to mitigate such hallucinations and fragility -- one of them being measuring the consistency (the model's confidence in the response, or likelihood of generating a similar response when resampled) of LLM responses. In previous work, measuring consistency often relied on the probability of a response appearing within a pool of resampled responses, or internal states or logits of responses. However, it is not yet clear how well these approaches approximate how humans perceive the consistency of LLM responses. We performed a user study (n=2,976) and found current methods typically do not approximate users' perceptions of LLM consistency very well. We propose a logit-based ensemble method for estimating LLM consistency, and we show that this method matches the performance of the best-performing existing metric in estimating human ratings of LLM consistency. Our results suggest that methods of estimating LLM consistency without human evaluation are sufficiently imperfect that we suggest evaluation with human input be more broadly used.

Problem

• Assessing how well surrogate metrics match human perception of LLM consistency
• Evaluating current methods for measuring LLM response consistency
• Proposing a logit-based ensemble method to better estimate human-rated consistency

Innovation

• Logit-based ensemble method for LLM consistency
• User study compares human and metric perceptions
• Advocates human input in LLM consistency evaluation