Separate de-censoring from calibration in TextWorld transfer

Determine whether the early-window within-task AUROC improvement observed after online training in TextWorld is caused by confidence calibration, increased confidence-tag emission, or both.

Background

The TextWorld transfer experiment reports an early-window improvement in within-task AUROC after online training with the Brier-plus-spread reward. The analysis considers de-censoring—an increased fraction of early steps emitting a confidence tag—as an alternative explanation for the apparent gain. Although the authors compare a format-only control and an emission-matched checkpoint, they state that the de-censoring and calibration contributions cannot be cleanly isolated, leaving the causal source of the early-window improvement unresolved.

References

The early-window improvement might instead reflect de-censoring, where training raises the fraction of early steps that emit a confidence tag, rather than the calibration term in isolation. We probe this with a format-only control but cannot cleanly separate the two.

— Can an Agent Learn to Predict Its Own Failure? Outcome-Grounded Confidence Calibration in LLM Agents  (2610.18614 - Hathidara, 13 Oct 2026) in Appendix, Section "Detailed results on Other Environments," paragraph "WebShop" preceding discussion