Preprint reports reduced variation in AI agent task progress

Researchers report that a new training signal reduces variation in AI agents' task progress from 132% to 91% on tau2-bench and from 90% to 44% on toolsandbox. Their preprint introduces Recursive Semantic Divergence, a measure of how independent runs of an agent disagree when starting from the same state. The method uses natural language inference to detect logical inconsistencies and a neural network to estimate disagreement across future turns. Used as an additional reward during training, it matches supervised fine-tuning performance with lower variance without requiring ground-truth labels. The findings concern benchmark tests, rather than demonstrated reliability in deployed systems.