EarlyEval: a method to stop an agent run early

A preprint by Yuling Shi and colleagues, posted to arXiv on September 2, uses LightGBM classifiers with a calibrated threshold to stop agent runs.

The authors report 13-26 per cent fewer steps, up to 44.1 per cent fewer input tokens, and 89-97 per cent prediction accuracy.

It is a preprint; early-stop bias and independent reproduction are lacking.