model	role	n_pairs	A_correct	B_correct	A_accuracy_percent	B_accuracy_percent	difference_pp	CI_low_pp	CI_high_pp	sign_flip_p
Qwen3-VL-8B-Instruct	panel construction	50	25	31	50	62	12	NA	NA	NA
Qwen2.5-VL-7B-Instruct	panel construction	50	17	24	34	48	14	NA	NA	NA
Qwen3-VL-2B-Instruct	panel construction	50	20	25	40	50	10	NA	NA	NA
InternVL3-2B	panel construction	50	21	25	42	50	8	NA	NA	NA
Qwen3-VL-8B-Thinking	held-out evaluation	50	23	26	46.0	52.0	6.0	-14.000000000000002	26.0	0.6904630953690463
Qwen3-VL-4B-Thinking	held-out evaluation	50	21	25	42.0	50.0	8.000000000000002	-12.0	28.000000000000004	0.5709442905570944
Qwen2.5-VL-32B-Instruct-AWQ	held-out evaluation	50	26	28	52.0	56.00000000000001	4.0000000000000036	-16.0	24.0	0.8383816161838381
