CoVeR: Conformal Calibration for Versatile and Reliable Autoregressive Next-Token Prediction

2025-09-04

AI Summary

Existing autoregressive decoding methods (e.g., beam search) lack provable coverage guarantees over generated paths, making it difficult to simultaneously ensure search efficiency, output diversity, and exploration of low-probability (long-tail) sequences. This work proposes CoVeR—the first method to integrate conformal prediction into autoregressive decoding—providing asymptotic $(1-alpha)$ coverage guarantees for any user-specified confidence level $alpha$. Its model-agnostic design enables plug-and-play integration, and its theoretical analysis leverages PAC-style generalization bounds to rigorously establish both coverage validity and convergence. Experiments demonstrate that CoVeR significantly improves coverage of diverse candidate paths on complex reasoning tasks while maintaining a compact search space, notably enhancing the discovery of rare yet critical long-tail sequences.

Technology Category

• Reasoning under Uncertainty: Sequential Decision Making
• Machine Learning: Calibration & Uncertainty Quantification
• Search and Optimization: Sampling/Simulation-based Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Autoregressive pre-trained models combined with decoding methods have achieved impressive performance on complex reasoning tasks. While mainstream decoding strategies such as beam search can generate plausible candidate sets, they often lack provable coverage guarantees, and struggle to effectively balance search efficiency with the need for versatile trajectories, particularly those involving long-tail sequences that are essential in certain real-world applications. To address these limitations, we propose extsc{CoVeR}, a novel model-free decoding strategy wihtin the conformal prediction framework that simultaneously maintains a compact search space and ensures high coverage probability over desirable trajectories. Theoretically, we establish a PAC-style generalization bound, guaranteeing that extsc{CoVeR} asymptotically achieves a coverage rate of at least $1 - α$ for any target level $αin (0,1)$.

Problem

• Ensures coverage guarantees for autoregressive models
• Balances search efficiency with versatile trajectories
• Addresses long-tail sequences in real-world applications

Innovation

• Conformal prediction framework for decoding
• Ensures high coverage probability guarantees
• Maintains compact search space efficiency