World-model state correction on Agent calling-tree testbed benchmark leaderboard · SOTA2 Research

World-model state correction on Agent calling-tree testbed

1Connectivity (%)Greedy-Point(K=1)

Evaluation Results

Method	Links
Greedy-Point(K=1)2026.07		1	0.144	0.22	99	184.7	185.2	0.005
LocalRepair-2Hop2026.07		1	0.362	1.81	27.5	91.8	92.8	0.009
LocalRepair-3Hop2026.07		1	0.325	1.97	0.2	3.3	3.3	0.001
Oracle2026.07		1	1	1.32	0	0	0	0.013
CascadeRepair2026.07		0.96	0.871	1.26	4.9	57.6	61.1	0.036
WM-SAR2026.07		0.94	0.845	1.95	6	6.3	6.3	0.0004
TopK-Point(K=5)2026.07		0.48	0.63	1.08	46.6	107.3	107.6	0.007
TopK-Point(K=3)2026.07		0.36	0.433	0.68	69.3	143.3	144.1	0.009
Window-2-Point2026.07		0.3	0.251	0.5	91.2	184.8	185.5	0.009
Window-8-Point2026.07		0.22	0.36	1.3	61.2	169.3	170.8	0.019
Window-4-Point2026.07		0.12	0.317	0.82	80	182.9	183.8	0.011