Factorized Study Clarifies LLM Uncertainty Estimation with Probes.

Detecting hallucinations and estimating uncertainty in Large Language Models (LLMs) is a critical challenge for their reliable deployment. Probe-based uncertainty estimation (UE) methods, which learn uncertainty signals from the LLM's internal states, have emerged as a promising approach. However, the varied designs of current methods—spanning feature selection, training data construction, and evaluation settings—make it difficult to pinpoint the true drivers of performance. To address this, researchers conducted a factorized study, systematically evaluating different probe-based UE components under matched conditions. Their findings indicate that while raw hidden states and attention features perform exceptionally well within the domain they were trained on, they struggle when faced with distribution shifts. In contrast, structured and compressed features demonstrate greater robustness in out-of-domain scenarios, suggesting that in-domain performance alone is an insufficient measure of a method's real-world utility. The study also highlighted the significant impact of prompting strategies and label construction on probe behavior. Building on these insights, the researchers developed benchmark-based pretrained probes that exhibit reasonable transferability to open-ended factual generation tasks, providing a stable, off-the-shelf baseline for future work. This research advocates for a more deployment-oriented evaluation of probe-based uncertainty estimators to ensure their practical effectiveness.