E-Scores for (In)Correctness Assessment of Generative Model Outputs
This addresses the need for robust correctness assessment in generative models, particularly for users relying on LLMs, offering a flexible and statistically sound method, though it is incremental by building on conformal prediction frameworks.
The paper tackles the problem of assessing the correctness of generative model outputs, which lack principled mechanisms, by introducing e-scores based on e-values to complement outputs with a measure of incorrectness, achieving statistical guarantees and allowing adaptive tolerance selection without p-hacking issues, as demonstrated in experiments on mathematical factuality and property constraints satisfaction.
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.