MASPRM: Multi-Agent System Process Reward Model

arXiv cs.AI · 18 hours ago

Researchers introduced MASPRM, a process reward model that evaluates intermediate messages in multi-agent systems to guide inference-time search, trained on MCTS rollouts without human annotations. On benchmarks including GSM8K and MATH, MASPRM outperformed size-matched outcome reward models by 2.0 to 14.5 points depending on model size and dataset. The approach enables more efficient multi-agent inference by identifying which agent communications advance progress toward solutions.