New Learning Method Improves GUI Agent Performance.

Multimodal Large Language Models (MLLMs) are increasingly used to power Graphical User Interface (GUI) agents, automating tasks across various digital environments. Reinforcement learning (RL) is a common training method for these agents, but existing techniques like Group Relative Policy Optimization (GRPO) often suffer from unstable optimization due to misaligned reward gradients. While recent work has improved stability by reframing RL with verifiable rewards (RLVR) as contrastive or classification objectives, these methods primarily rely on outcome-level supervision, failing to capture subtle differences in trajectory quality within the same outcome category. To address this, researchers propose Length-Aware Contrastive Learning for GUI Agents (LACL-GUI). This novel contrastive RLVR framework integrates trajectory-level quality signals into the policy optimization process. LACL-GUI establishes structured preferences for both successful and failed trajectories: it promotes more concise and efficient successful executions, and it differentiates the quality of failures based on how far they diverge from successful paths, all while maintaining optimization stability. Experiments conducted on standard GUI agent benchmarks demonstrate that LACL-GUI provides more effective learning signals. It consistently improves agent performance compared to prior methods, underscoring the significant value of incorporating detailed trajectory-level supervision in contrastive RLVR for GUI automation.