AdaStep preprint reports agent-training gains without extra model inference

Researchers report in a preprint that AdaStep improves language-model agent training across three model backbones on ALFWorld, WebShop, and ScienceWorld, using lightweight scalar computation without an additional evaluation model, extra trials, or extra model inference. The method adjusts how much credit individual decisions receive within a longer sequence of actions. It preserves local credit when variation in returns reflects the selected action and reduces it when later randomness dominates. The derivation of its optimal weighting coefficient depends on an explicit conditional sampling assumption; the reported improvements come from the named experimental environments.