ResearchBenchmarksVLM Agent Success Rate Estimation on Sokoban 6x6Follow80.5Success RateSTAPO8.94827.52446.164.676Jul 6, 2026Evaluation ResultsMethodMethodLinksSuccess RateSTAPOBackbone=Qwen2.5-VL-3B...Backbone=Qwen2.5-VL-3B-Instruct2026.0780.5GiGPOBackbone=Qwen2.5-VL-3B...Backbone=Qwen2.5-VL-3B-Instruct2026.0776.9GRPOBackbone=Qwen2.5-VL-3B...Backbone=Qwen2.5-VL-3B-Instruct2026.0767.1BaseBackbone=Qwen2.5-VL-3B...Backbone=Qwen2.5-VL-3B-Instruct2026.0711.7