Multi-view global spatial scene understanding on VSI-Bench Standard benchmark leaderboard · SOTA2 Research

Multi-view global spatial scene understanding on VSI-Bench Standard

70.5Relative Distance AccuracyGR3D-8B

Evaluation Results

Method	Links
GR3D-8B2026.05		70.5	69.6	55.2	76.8	65.6	86.3	35.5	81.2	67.6
SR-3D-8B2026.05		63.5	54.9	53.8	74.5	65.1	81.8	33.5	75.9	62.9
Gemini-1.5 Pro2026.05		51.3	56.2	30.9	64.1	43.6	46.3	36	34.6	45.3
InternVL2-40B2026.05		47.6	41.3	26.2	48.2	27.5	32.7	27.8	44.7	37
InternVL2-2B2026.05		44.2	24.9	22	35	33.8	—	—	—	—
LLaVA-Video-7B2026.05		43.5	48.5	14	47.8	24.2	42.4	34	30.6	35.6
LLaVA-OneVision-7B2026.05		42.5	47.7	20.2	47.4	12.3	35.2	29.4	24.4	32.4
LLaVA-OneVision-72B2026.05		42.5	43.5	23.9	57.6	37.5	39.9	32.5	44.6	40.2
LLaVA-Video-72B2026.05		42.4	48.9	22.8	57.4	35.3	36.7	35	48.6	40.9
VILA-1.5-40B2026.05		40.5	22.4	24.8	48.7	22.7	25.7	31.5	32.9	31.2
InternVL2-8B2026.05		38	31.3	29	48.9	44.2	33.4	28.9	46.4	37.5
Gemini-1.5 Flash2026.05		37.7	49.8	30.8	53.5	54.4	41	31.5	37.8	42.1
GPT-4o2026.05		37	46.2	5.3	43.8	38.2	41.3	31.5	28.5	34
LongVA-7B2026.05		33.1	38	16.6	38.9	22.2	43.3	25.4	15.7	29.2
VILA-1.5-8B2026.05		32.1	17.4	21.8	50.3	18.8	34.8	31	24.8	28.9
LongVILA-8B2026.05		29.6	29.1	9.1	16.7	0	30.7	32.5	25.5	21.6
Random2026.05		25	—	—	—	—	36.1	28.3	25	—