Confidence as Control: A Survey of Confidence Utilization in Large Language Models

ACL ARR 2026 January Submission1584 Authors

Keywords: confidence utilization, uncertainty quantification, large language models, selective prediction, model routing, retrieval augmented generation

Abstract: Confidence in LLMs is often studied through uncertainty estimation and calibration. We survey a complementary perspective: confidence as a control signal that governs system behavior. We organize $\textbf{confidence utilization}$ across the LLM lifecycle: (i) training (data selection, loss weighting, self-training, and preference optimization); (ii) inference (candidate selection, adaptive computation, and confidence-guided contrastive decoding); and (iii) deployment (cost-aware routing and cascading, RAG control (retrieval triggering, context filtering, and parametric–retrieval arbitration), and risk-aware abstention and monitoring. We unify these techniques into a framework that turns confidence into system decisions, with implications for efficiency and reliability.

Paper Type: Long

Research Area: Language Models

Research Area Keywords: uncertainty quantification, language models, calibration, retrieval-augmented generation, model routing, preference learning, selective prediction

Contribution Types: Surveys

Languages Studied: English

Submission Number: 1584