MCGPO preprint reports higher agent success using existing training trajectories

Researchers report in an arXiv preprint that Matched-Context Group Policy Optimization (MCGPO) improves language model agents' task success by learning action preferences from trajectories already collected during training. The method compares decisions in matched contexts and weights the evidence used to learn from them, without adding a critic model. Across six environment and model settings with three training seeds each, mean success improved over matched GraphGPO by 0.47-1.46 percentage points for Qwen2.5 models on ALFWorld and WebShop, and by 3.96 and 5.16 points for Qwen3 models on WebShop. Integrated success improved in three of four Qwen2.5 settings; ALFWorld baseline success already exceeded 92%.