Share your thoughts, 1 month free Claude Pro on usSee more

Embodied spatial reasoning on UrbanVideo-Bench

57.3Clean

Robust-TO (Qwen3-VL-7B-Instruct)

Updated 1mo ago

Evaluation Results

Method	Links
Robust-TO (Qwen3-VL-7B-Instruct) 2026.06		57.3	54.1	54	54.9	53.5	55.1	54.3
Video-R1 (Qwen3-VL-7B-Instruct) 2026.06		52	48.5	48	49	47.5	49.5	48.5
Robust-TO (Qwen2.5-VL-7B-Instruct) 2026.06		50.6	47	46.5	47.7	46.1	48.3	47.1
Gemini-2.5-Pro 2026.06		44.3	37.7	37.7	38.8	36.4	39.8	38.1
Video-R1 (Qwen2.5-VL-7B-Instruct) 2026.06		43	38.5	38	39.5	37	40.5	38.7
Qwen2.5-VL-7B-Instruct (SFT) 2026.06		39	30.9	30.7	32.1	29.2	33.3	31.2
GPT-4o 2026.06		37.4	32.2	31.7	32.5	30.8	33.6	32.2
Qwen2.5-VL-72B-Instruct 2026.06		33.4	26.3	26	27.2	24.3	28.7	26.5
Qwen2.5-VL-7B-Instruct 2026.06		26.9	17	16.6	18.9	14.6	20.3	17.5