GUI-Libra: Training Native GUI Agents to Reason and Act with Action-aware Supervision and Partially Verifiable RL

Benchmark Model Rank Results
computer-use-agents-on-online-mind2webGUI-Libra-8B#9Success Rate: 19.3
computer-use-agents-on-screenspot-v2GUI-Libra-8B + Mix Grounding 20K (extra data)#11Accuracy: 94.8
computer-use-agents-on-screenspot-proGUI-Libra-4B + Mix Grounding 20K (extra data)#15Accuracy: 61.4