New method credits agent training by verified progress, authors report 4.1-point gain

A preprint proposes ProCredit: instead of assigning a single outcome reward at the end of a task, it reruns the acceptance checks that decide success after every turn, and scores each turn by its change in verified progress.

The authors report that on the AppWorld benchmark, starting from Qwen3.5 base models at three scales, the method beats both outcome-reward and progress-based baselines at every scale, exceeding the strongest outcome-reward baseline by 4.1 percentage points at 4B; a second environment shows the same direction. Ablations indicate that adding final progress to the trajectory score alone does not help — the gain comes from crediting progress to the turn where it occurs.

The idea targets a known weakness of outcome rewards: a group of attempts with no success yields no training signal, and failures cannot be told apart by how close they came. The numbers are the authors' own tests, and the abstract gives no figures for the second environment; independent replication is still pending.