Instance-dependent early stopping (IES) is a training-time shortcut: instead of backpropagating through every example on every step, it skips examples whose losses have already stabilised. That saves computation, but the criterion used to decide when a loss has stabilised — a second-order loss-difference check — has not had a solid theoretical justification.
The new preprint, arXiv:2610.24540, sets out to close that gap. The authors propose a theory-guided version of the second-order loss-difference criterion, grounding the heuristic in formal analysis rather than empirical tuning.
The abstract is truncated, so the article does not detail the proposed method or experimental results. What is clear is the motivation: make IES reliable enough to use widely, by replacing an ad hoc criterion with one that has a theoretical basis.