Machine Learning PapersLandmark papers
← All papers
RESEARCH PAPER / 2026-11-15

Measuring Coherence in Predictive Models

Jamie Reason, Vik Shirvaikar, Mark van der Wilk, Chris Holmes

University of Cambridge (inferred from EPSRC StatML CDT affiliation) +2 · NeurIPS 2026

Affiliations: University of Cambridge, University of Oxford (inferred from EPSRC StatML CDT affiliation)

Abstract

Many sequential decision-making procedures, such as active learning, implicitly assume that a model's current predictions agree with those it expects to make after updating on new data, a property we call coherence. Most models in use do not update by exact conditioning and are not guaranteed to be coherent, which can lead to worse decisions. Yet there is currently no general way to measure incoherence or to tell when it matters. We develop tools to measure incoherence and characterise when it affects decisions. We introduce the incoherence ratio, an MMD-based measure of how far any updating procedure departs from coherence, computed only from the procedure's own predictions and updates, and accompany it with an anytime-valid test of coherence. In Bayesian experimental design, acquisition objectives that are equivalent for coherent models choose different data under every approximate model we study, and the incoherence ratio tracks how far incoherence distorts them. We also introduce action coherence, which captures only the incoherence that changes a decision. What a resulting disagreement costs, however, is determined primarily by predictive accuracy. Sequential decisions depend on how a model updates as well as on what it predicts, so coherence, like accuracy, should be checked rather than assumed.

The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.

Research summary

The paper introduces a rigorous, model-agnostic framework for measuring and testing predictive coherence in sequential decision-making, demonstrating that incoherence in approximate Bayesian models significantly distorts acquisition functions in experimental design. By providing the incoherence ratio and an anytime-valid test, the authors offer essential diagnostic tools that reveal the hidden costs of approximate updates, urging the field to treat coherence as a critical property alongside accuracy.

Methodology

The paper introduces a rigorous framework for defining and measuring "predictive coherence" in sequential decision-making models. The core contribution is the "incoherence ratio," an MMD-based metric that quantifies the deviation of a model's update rule from exact Bayesian conditioning. This is a significant methodological advance because it provides a model-agnostic, black-box diagnostic tool that does not require access to the true posterior or a reference Bayesian model. The authors also propose an anytime-valid sequential test for coherence using betting strategies, which is a sophisticated application of statistical testing theory to ML diagnostics. The concept of "action coherence" is particularly insightful, distinguishing between general incoherence and incoherence that actually alters decision outcomes. The theoretical grounding in Dutch book arguments and martingale properties is solid and well-connected to the literature on predictive inference.

Results & evaluation

The experimental section is comprehensive, covering sparse Gaussian Processes, tabular foundation models (TabPFN), Bayesian Linear Regression, and a synthetic neural testbed. The results convincingly demonstrate that incoherence is prevalent in approximate Bayesian methods (VI, Laplace, Deep Ensembles) and that it leads to divergent acquisition functions in Bayesian Experimental Design (BED). The finding that the entropy form of EPIG performs worse than the mutual information form under incoherence is a practical and actionable insight. The use of the neural testbed to correlate incoherence ratio with decision distortion is strong evidence for the utility of the proposed metric.

Reproducibility

The paper provides a public GitHub repository. The experimental details are described in sufficient depth in the appendix, including compute resources and specific hyperparameter choices. The metrics are defined clearly, allowing for independent implementation.

Limitations

The primary limitation is computational cost; measuring incoherence requires multiple model updates and sampling, which can be expensive for large deep learning models. The paper acknowledges that coherence does not guarantee accuracy, and the metric is a diagnostic rather than a fix. The scope is limited to single-update sequential decisions, with multi-step planning left for future work.

Broader impact

This work has high potential impact on the field of Bayesian deep learning and active learning. It provides a necessary diagnostic tool for practitioners who rely on approximate Bayesian methods for sequential decisions. By highlighting that standard acquisition functions can behave differently under incoherence, it may lead to the development of more robust decision-making algorithms that account for model inconsistency. It bridges the gap between theoretical Bayesian consistency and practical approximate inference. The paper introduces a rigorous, model-agnostic framework for measuring and testing predictive coherence in sequential decision-making, demonstrating that incoherence in approximate Bayesian models significantly distorts acquisition functions in experimental design. By providing the incoherence ratio and an anytime-valid test, the authors offer essential diagnostic tools that reveal the hidden costs of approximate updates, urging the field to treat coherence as a critical property alongside accuracy.

Author of this paper? Send a correction.