Ride-hailing adjudication on Driver-Cancel benchmark leaderboard · SOTA2 Research

Ride-hailing adjudication on Driver-Cancel

100Normal PrecisionQwen3-VL-8B-Instruct

Evaluation Results

Method	Links
Qwen3-VL-8B-Instruct2026.03		100	9.6	50	15.8	71	60.61
Qwen3-VL-8B-Thinking2026.03		100	0.7	0	0	13	8.82
Qwen3-VL-32B-Instruct2026.03		93.5	43	45.8	57.9	78.4	65.55
Qwen3-VL-32B-Instruct2026.03		91	52.6	53.6	78.9	72.1	59.14
Qwen3-VL-8B-Instruct2026.03		90	13.3	54.5	31.6	71.6	62.09
DeepSeek-V3.12026.03		80.69	86.67	64	84.21	86.15	75.25
RideJudge-8B2026.03		78.85	91.11	41.86	94.73	83.55	88.41
DeepSeek-V3.12026.03		77.99	91.85	58.33	73.68	86.36	75.36
RideJudge-4B2026.03		72.41	93.33	51.42	94.74	80.95	87.25
MiniCPM-V2026.03		53.33	17.78	0	0	33.12	19.68
Qwen3-VL-8B-Thinking2026.03		50	2.2	0	0	13.9	7.37
QwQ-32B2026.03		47.5	70.37	30.3	52.63	22.51	35.18
MiniCPM-V2026.03		44	8.15	0	0	16.88	12.96
Qwen3-32B-Instruct2026.03		39.55	39.26	42.86	31.58	19.91	26.21
QwQ-32B2026.03		38.19	40.74	25	21.05	26.62	35.82
Qwen3-VL-32B-Thinking2026.03		34.76	48.15	0	0	54.55	45.94
Qwen3-32B-Instruct2026.03		32.67	36.3	23.08	15.79	25.97	26.77
Qwen3-VL-32B-Thinking2026.03		32.29	22.96	0	0	61.04	56.86