arxivcs.CLcs.AI2026-06-26
From Signals to Transfer: A Factorised Study of Probe-Based Uncertainty Estimation in Large Language Models
Ponhvoan Srey, Xiaobao Wu, Cong-Duy Nguyen, Quang Minh Nguyen, Duc Anh Vu, Anh Tuan Luu
Probe-based uncertainty estimation (UE) has emerged as a prominent approach to detect hallucinations in Large Language Models (LLMs) by learning uncertainty from internal model signals. Yet, recent methods vary simultaneously across feature design, training data construction, and…