Visual Reasoning on REASONMAP-PLUS benchmark leaderboard · SOTA2 Research

Visual Reasoning on REASONMAP-PLUS

88.95Weighted AccuracyGPT-5

Evaluation Results

Method
GPT-52025.10	88.95	86.4	91.6
Qwen2.5-VL-7B-Instruct + REWARDMAP2025.10	74.25	72.18	76.42
Seed1.5-VL2025.10	73.58	65.26	82.23
Qwen2.5-VL-7B-Instruct + RL (baseline)2025.10	67.61	68.37	66.82
GPT-4o2025.10	64.42	59.28	69.77
Qwen2.5-VL-7B-Instruct + SFT -> RL2025.10	60.53	55.38	65.9
Qwen2.5-VL-32B-Instruct2025.10	58.32	46.96	70.14
Qwen2.5-VL-7B-Instruct + SFT2025.10	57.93	50.73	65.44
Qwen2.5-VL-72B-Instruct2025.10	53.21	43.46	63.36
Qwen2.5-VL-7B-Instruct + RL (ReMax)2025.10	45.39	38.37	52.7
Qwen2.5-VL-7B-Instruct + RL (GRPO)2025.10	44.64	37.57	52.01
Qwen2.5-VL-7B-Instruct + RL (REINFORCE++)2025.10	44.64	36.82	52.79
Qwen2.5-VL-7B-Instruct2025.10	44.21	37.39	51.32
Qwen2.5-VL-3B-Instruct2025.10	37.61	22.68	53.16
Kimi-VL-A3B-Thinking2025.10	33.95	18.17	50.39
Kimi-VL-A3B-Instruct2025.10	32.55	14.75	51.08