From Signals to Transfer: A Factorised Study of Probe-Based Uncertainty Estimation in Large Language Models

A new factorised study of probe-based uncertainty estimation in Large Language Models isolates which design choices actually drive performance, finding that raw internal signals dominate in-domain but structured features prove more robust when data shifts [1]. Probe-based uncertainty estimation has become a widely used method for detecting hallucinations in Large Language Models by learning uncertainty from internal model signals [1]. However, the field has lacked clarity because recent methods vary simultaneously across feature design, training data construction, and evaluation setting, obscuring what truly drives performance [2]. To address this, researchers proposed a factorised study that tests each component under matched conditions [1]. The results show that raw hidden states and attention features are difficult to outperform when evaluated in-domain [1]. But under distribution shift, structured and compressed features are more robust, suggesting that in-domain performance alone is insufficient to measure progress [2]. The study also found that prompting and label construction significantly affect probe behaviour [1]. Building on these findings, the team trained benchmark-based pretrained probes that transfer reasonably well to open-ended factual generation, providing a stable off-the-shelf baseline [2]. The work encourages more deployment-oriented evaluation of probe-based uncertainty estimators, shifting focus away from narrow in-domain metrics [1]. The code repository has been made publicly available [2]. Probe-based methods operate by extracting internal representations from a model's forward pass and training a separate classifier to predict whether a given output is factual or hallucinated [2]. The factorised approach breaks this pipeline into distinct components — feature type, compression method, training labels, and prompting strategy — and varies one at a time while holding others constant [1]. This contrasts with prior work where multiple factors changed between experiments, making it impossible to attribute gains to any single design choice [2]. The finding that raw features excel in-domain but degrade under distribution shift aligns with broader machine learning principles about overfitting to training distributions [1]. Structured and compressed features, by discarding noise and preserving generalisable patterns, appear to offer a better trade-off for real-world deployment where input distributions inevitably drift [2]. The pretrained probes released with the study aim to lower the barrier for practitioners who need reliable uncertainty estimates without retraining on every new task [1].