Newest papers

Auditing Conformal Prediction for Language Models Under Inference-Time Model Shift

Achref Doula, Max Mühlhäuser · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.28937 · cited by 0 works (OpenAlex) · open-access PDF

French

La prédiction conforme (CP) pour les modèles linguistiques est généralement évaluée selon un protocole correspondant, dans lequel la calibration et le déploiement utilisent la même configuration d'inférence post-entraînement, comme le mode de quantification et la température d'inférence. Dans le déploiement, ces choix diffèrent souvent de ceux utilisés lors de la calibration pour répondre aux contraintes de mémoire ou de débit ou des politiques de service. Nous affirmons que cela expose une limitation dans les protocoles d'évaluation CP actuels pour les modèles de langue. Nous formalisons le phénomène sous-jacent en tant que changement de modèle de temps d'inférence: un changement dans la distribution de score conforme induite par la configuration

Machine translation (Meta M2M100); the English original follows.

Abstract

Conformal prediction (CP) for language models is typically evaluated under a matched protocol, in which calibration and deployment use the same post-training inference configuration, such as quantization mode and inference temperature. In deployment, these choices often differ from those used at calibration to meet memory or throughput constraints or serving policies. We argue that this exposes a limitation in current CP evaluation protocols for language models. We formalize the underlying phenomenon as inference-time model shift: a change in the conformal score distribution induced by inference configuration rather than by the input--label distribution, and theoretically characterize how it breaks split-conformal exchangeability even when the task distribution and trained model are unchanged. We then evaluate its consequences across 8 language models, 4 finite-label benchmarks, and 12 inference configurations. We find that matched evaluation can substantially overstate deployment-time coverage, and that the effect is heterogeneous across models and datasets. We argue that inference configuration should be reported as a first-class experimental axis in CP evaluations for language-model systems.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Inference, Computer science, Fiducial inference, Artificial intelligence, Software deployment, Language model

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange