MLLM-as-a-Judge Evaluation on VL RewardBench benchmark leaderboard · SOTA2 Research

MLLM-as-a-Judge Evaluation on VL RewardBench

80.75AccuracyM-Judger-RL-Qwen8B

Evaluation Results

Method	Links
M-Judger-RL-Qwen8B2026.02		80.75
Qwen3-VL-8B-Instruct2026.02		77.15
M-Judger-SFT-Qwen8B2026.02		76.82
M-Judger-SFT-Qwen4B2026.02		76.5
M-Judger-RL-Qwen4B2026.02		76.42
Qwen3-VL-4B-Instruct2026.02		75.94
UnifiedReward-Think-qwen-7b2026.02		73.8
R1-Reward2026.02		71.92
GPT-4o2026.04		66.95
Llama-3.2-Vision-90B2026.04		64.52
Qwen3-VL-8B-Instruct2026.02		63.91
Qwen3-VL-4B-Instruct2026.02		61.75
Gemini-2.5-Pro2026.04		61.49
GPT-4o2026.04		61.12
Gemini-2.5-Flash2026.04		61.08
GPT-4o2026.04		60.42
LLaVA-OneVision-1.5-8B2026.04		60.01
Llama-3.2-Vision-90B2026.04		59.94
Llama-3.2-Vision-90B2026.04		59.46
GPT-4o2026.04		59.36
Gemini-2.5-Pro2026.04		59.21
Llama-3.2-Vision-90B2026.04		59.21
LLaVA-OneVision-1.5-8B2026.04		57.95
Gemini-2.5-Flash2026.04		57.84
Gemini-2.5-Pro2026.04		57.55
Phi-4-multimodal2026.04		57.45
Gemini-2.5-Pro2026.04		57.33
LLaVA-OneVision-1.5-8B2026.04		57.07
Gemini-2.5-Flash2026.04		56.65
Gemini-2.5-Flash2026.04		56.25
LLaVA-OneVision-1.5-8B2026.04		56.01
Phi-4-multimodal2026.04		55.45
Phi-4-multimodal2026.04		54.72
Phi-4-multimodal2026.04		54.49
LLaVA-1.6-34B2026.04		52.03
LLaVA-1.6-34B2026.04		50.19
LLaVA-1.6-34B2026.04		49.29
LLaVA-1.6-34B2026.04		48.83
LLaVA-1.5-13B2026.04		47.32
LLaVA-1.5-13B2026.04		46.59
LLaVA-1.5-13B2026.04		46.24
LLaVA-1.5-13B2026.04		44.76