Can Generative AI Solve Your In-Context Learning Problem? A Martingale Perspective

TL;DR

This paper introduces a Bayesian-inspired method to assess whether a generative model is suitable for in-context learning tasks by developing a generative predictive p-value, enabling model criticism without explicit likelihoods.

Contribution

It proposes a novel statistical approach using generative predictive p-values for evaluating the adequacy of contemporary generative models in in-context learning scenarios.

Findings

Method works on synthetic tabular, imaging, and language tasks.

Enables model criticism without explicit likelihoods.

Empirically validated on large language models.

Abstract

This work is about estimating when a conditional generative model (CGM) can solve an in-context learning (ICL) problem. An in-context learning (ICL) problem comprises a CGM, a dataset, and a prediction task. The CGM could be a multi-modal foundation model; the dataset, a collection of patient histories, test results, and recorded diagnoses; and the prediction task to communicate a diagnosis to a new patient. A Bayesian interpretation of ICL assumes that the CGM computes a posterior predictive distribution over an unknown Bayesian model defining a joint distribution over latent explanations and observable data. From this perspective, Bayesian model criticism is a reasonable approach to assess the suitability of a given CGM for an ICL problem. However, such approaches -- like posterior predictive checks (PPCs) -- often assume that we can sample from the likelihood and posterior defined by the Bayesian model, which are not explicitly given for contemporary CGMs. To address this, we show when ancestral sampling from the predictive distribution of a CGM is equivalent to sampling datasets from the posterior predictive of the assumed Bayesian model. Then we develop the generative predictive $p$-value, which enables PPCs and their cousins for contemporary CGMs. The generative predictive $p$-value can then be used in a statistical decision procedure to determine when the model is appropriate for an ICL problem. Our method only requires generating queries and responses from a CGM and evaluating its response log probability. We empirically evaluate our method on synthetic tabular, imaging, and natural language ICL tasks using large language models.