E-Scores for (In)Correctness Assessment of Generative Model Outputs

2025-10-29

AI Summary

Existing evaluation methods for correctness in generative models—especially large language models (LLMs)—lack principled statistical frameworks; p-value–based approaches are vulnerable to p-hacking and cannot adaptively set error-tolerance thresholds post-hoc. Method: We propose E-Scores, a novel error quantification framework grounded in e-values—nonnegative random variables that measure evidence against correctness—and inherently support post-hoc threshold selection without compromising statistical validity. By unifying conformal prediction with e-value theory, E-Scores provides a coherent assessment paradigm applicable to diverse correctness criteria, including mathematical factuality and attribute constraints. Contribution/Results: Experiments across multiple LLM output scenarios demonstrate that E-Scores significantly improve both reliability and flexibility in error detection. The framework offers a theoretically rigorous yet practical tool for trustworthy evaluation of generative model outputs.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Reasoning under Uncertainty: Relational Probabilistic Models
• Natural Language Processing: Safety and Robustness

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
• Security and Privacy: Large-scale security measurements

Abstract

While generative models, especially large language models (LLMs), are ubiquitous in today's world, principled mechanisms to assess their (in)correctness are limited. Using the conformal prediction framework, previous works construct sets of LLM responses where the probability of including an incorrect response, or error, is capped at a desired user-defined tolerance level. However, since these methods are based on p-values, they are susceptible to p-hacking, i.e., choosing the tolerance level post-hoc can invalidate the guarantees. We therefore leverage e-values to complement generative model outputs with e-scores as a measure of incorrectness. In addition to achieving the same statistical guarantees as before, e-scores provide users flexibility in adaptively choosing tolerance levels after observing the e-scores themselves, by upper bounding a post-hoc notion of error called size distortion. We experimentally demonstrate their efficacy in assessing LLM outputs for different correctness types: mathematical factuality and property constraints satisfaction.

Problem

• Developing principled assessment mechanisms for generative model outputs
• Addressing p-hacking vulnerability in statistical correctness guarantees
• Providing flexible tolerance selection for different correctness types

Innovation

• Using e-values to measure generative model incorrectness
• Providing flexible tolerance level selection post-evaluation
• Achieving statistical guarantees without p-hacking vulnerability