StainFlow: Entity-Stain Tracking and Evidence Linking for Process Rewards in GUI Agents

StainFlow is a novel entity-stain-flow process reward model designed to enhance Reinforcement Learning (RL) for GUI Agents by addressing limitations in existing Process Reward Models (PRMs). It features a Global Entity Stain Tracking module that objectively separates task phases based on visually verifiable entities and their evolving states, alongside a Local Stain Evidence Linking module that dynamically constructs evidence windows for improved local verification. Experimental results demonstrate a 3.2% increase in online RL success and a 1.8% boost in trajectory completion judgment accuracy on benchmarks like AndroidWorld and OGRBench, providing practitioners with a more reliable framework for credit assignment in complex GUI interactions.