block	column_1	column_2	column_3	column_4	column_5	column_6
1	Model	Control correct	Control accuracy	Original correct	Original accuracy	Gap
1	InternVL3-2B	151/200	75.5% [69.1,80.9]	78/200	39.0% [32.5,45.9]	+36.5 pp
1	Qwen3-VL-2B-Instruct	81/200	40.5% [33.9,47.4]	76/200	38.0% [31.6,44.9]	+2.5 pp
1	Qwen3-VL-8B-Instruct	162/200	81.0% [75.0,85.8]	88/200	44.0% [37.3,50.9]	+37.0 pp
1	Qwen2.5-VL-7B-Instruct	141/200	70.5% [63.8,76.4]	76/200	38.0% [31.6,44.9]	+32.5 pp
1	Qwen3-VL-8B-Thinking	170/200	85.0% [79.4,89.3]	69/200	34.5% [28.3,41.3]	+50.5 pp
1	Qwen3-VL-4B-Thinking	172/200	86.0% [80.5,90.1]	78/200	39.0% [32.5,45.9]	+47.0 pp
1	Qwen2.5-VL-32B-Instruct-AWQ	167/200	83.5% [77.7,88.0]	97/200	48.5% [41.7,55.4]	+35.0 pp
2	Model	C+O-	C-O+	McNemar P	Original given control pass
2	InternVL3-2B	92	19	1.12e-12	39.1%
2	Qwen3-VL-2B-Instruct	50	45	0.682	38.3%
2	Qwen3-VL-8B-Instruct	93	19	6.75e-13	42.6%
2	Qwen2.5-VL-7B-Instruct	83	18	3.73e-11	41.1%
2	Qwen3-VL-8B-Thinking	111	10	1.04e-22	34.7%
2	Qwen3-VL-4B-Thinking	103	9	2.31e-21	40.1%
2	Qwen2.5-VL-32B-Instruct-AWQ	85	15	4.83e-13	49.1%
