Develop label-free estimates of adaptation benefit and broader generalizations of the controller

Develop methods to estimate the helped-versus-hurt balance of retained cases without labels, formulate scale-free versions of prediction-fragmentation statistics, learn the routing policy rather than hand-specifying it, and determine whether the same keep-or-revert decision applies to test-time-adapted foundation segmenters.

Background

The prediction-fragmentation controllers require labeled holdout data for calibration, even though their per-case deployment decisions use no labels. The paper therefore identifies several unresolved directions concerning how to assess whether retained cases are net-helped, how to make fragmentation statistics transfer across modalities and domains, and how to replace manually specified routing actions with learned policies.

The authors also leave unresolved whether the keep-or-revert decision studied for episodic test-time adaptation of frozen nnU-Net and SegFormer models extends to test-time-adapted foundation segmentation models.

References

Estimating the helped/hurt balance without labels, scale-free formulations of the statistic, learning a policy rather than hand-specifying one, and the same keep-or-revert decision on test-time-adapted foundation segmenters are open.

— Should This Case Be Adapted? Prediction Fragmentation Controls Test-Time Adaptation  (2609.20700 - Wang et al., 17 Sep 2026) in Discussion and Limitations, paragraph “Baselines, calibration, and scope”

What remains open is how much change in evidence use is healthy adaptation and when it becomes a loss of grounding that should block or reverse the policy update.

— Seeing, Maintaining, and Learning: The Evolution of Multimodal Agent Memory  (2610.19366 - Guo et al., 14 Oct 2026) in Section 5.3.3, “Policy Adaptation with Preserved Evidence Grounding” (Section sec:open-evidence-use)