Vision-centric Visual Question Answering on MRAG-Bench (test) benchmark leaderboard · SOTA2 Research

Vision-centric Visual Question Answering on MRAG-Bench (test)

67.7Angle Robustness ScoreQwen2.5-VL-7B-Instruct + R3G (Ours)

Evaluation Results

Method
Qwen2.5-VL-7B-Instruct + R3G (Ours)2026.01	67.7	61.79	64.71	65.74	59.06	29.41	46.08	51.96	56.67	58.61
Qwen2.5-VL-7B-Instruct + MRAG2026.01	61.18	51.63	57.84	67.59	64.43	46.08	33.33	53.92	57.5	55.95
LLaVA-OneVision + R3G (Ours)2026.01	54.66	60.98	56.86	62.96	51.68	52.94	34.31	54.9	60	55.14
LLaVA-OneVision + MRAG2026.01	50.93	48.78	50	60.19	50.34	48.04	33.33	53.92	54.17	50.11
Mantis-8B-siglip-llama3 + R3G (Ours)2026.01	48.41	45.5	46.18	50.08	43.62	47.32	24.51	29.99	46.78	42.36
LLaVA-NeXT-Interleave-7B + R3G (Ours)2026.01	46.58	48.37	54.9	61.11	43.62	46.08	36.27	25.49	50.83	46.34
Mantis-8B-siglip-llama3 + MRAG2026.01	42.55	35.37	47.06	47.22	42.95	45.1	23.53	29.41	40.83	39.62
Mantis-8B-clip-llama3 + R3G (Ours)2026.01	40.33	45.13	42.08	51.44	51.42	35.21	19.03	21.89	50.58	41.26
LLaVA-NeXT-Interleave-7B + MRAG2026.01	40.06	33.33	39.22	56.48	43.62	44.12	27.45	36.27	49.17	40.35
Deepseek-VL-7B-chat + R3G (Ours)2026.01	39.44	45.12	46.08	50.93	43.62	33.33	24.51	15.69	54.17	40.28
Mantis-8B-clip-llama3 + MRAG2026.01	36.65	34.96	42.16	47.22	50.34	33.33	18.63	21.57	42.5	36.88
Deepseek-VL-7B-chat + MRAG2026.01	33.54	32.11	33.33	37.04	43.62	40.2	20.59	26.47	45	34.66