Multimodal Reward Modeling on RewardBench MM-RLHF
95.74MCQ ScoreBaseReward (Qwen2.5-VL)
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| BaseReward (Qwen2.5-VL)#Param=7B2026.04 | 95.74 | 97.38 | 94.13 | 81.25 | 88 | 91.76 | 80.43 | |
| DT2IT-MRM (Qwen2.5-VL)#Param=7B2026.04 | 95.24 | 97.3 | 100 | 71.88 | 88 | 91.18 | 78.26 | |
| DT2IT-MRM (Qwen3-VL)#Param=8B2026.04 | 90.48 | 94.6 | 97.06 | 78.13 | 84 | 89.41 | 71.74 | |
| MM-RLHF-Reward#Param=7B2026.04 | 83 | 97 | 74 | 69 | 88 | 82 | 63 | |
| UnifiedReward-Think*#Param=7B2026.04 | 73.81 | 86.49 | 61.77 | 59.38 | 72 | 71.18 | 39.13 | |
| UnifiedReward*#Param=7B2026.04 | 69.05 | 94.6 | 67.65 | 84.38 | 84 | 79.41 | 52.17 | |
| Claude-3.7-Sonnet2026.04 | 66.67 | 91.89 | 91.18 | 87.5 | 76 | 82.35 | 65.22 | |
| LLaVA-Critic*#Param=8B2026.04 | 66.67 | 81.08 | 64.71 | 71.88 | 92 | 74.12 | 47.83 | |
| GPT-4o (2024-08-06)2026.04 | 64.28 | 78.37 | 44.11 | 56.25 | 40 | 58.23 | 26.01 | |
| Claude-3.5-Sonnet (2024-06-22)2026.04 | 64.28 | 67.56 | 55.88 | 65.62 | 60 | 62.94 | 26.11 | |
| GPT-5.2*2026.04 | 57.14 | 91.89 | 55.88 | 59.38 | 80 | 68.24 | 41.3 | |
| R1-Reward*#Param=7B2026.04 | 52.38 | 48.65 | 47.06 | 43.75 | 44 | 47.65 | 21.74 | |
| IXC-2.5-Reward#Param=7B2026.04 | 52.38 | 91.89 | 67.65 | 62.5 | 88 | 71.18 | 50 | |
| Qwen2-VL-72B-Instruct#Param=72B2026.04 | 45.23 | 62.16 | 47.05 | 46.88 | 36 | 48.23 | 13.04 | |
| Skywork-VL Reward*#Param=7B2026.04 | 42.86 | 81.08 | 73.53 | 65.63 | 72 | 65.88 | 41.3 | |
| NVLM-D#Param=72B2026.04 | 42.85 | 32.43 | 8.82 | 50 | 40 | 34.7 | 6.52 | |
| Omni-RewardModel-BT*#Param=8B2026.04 | 40.48 | 78.38 | 70.59 | 46.88 | 76 | 61.18 | 39.13 | |
| InternVL3-8B#Param=8B2026.04 | 35.71 | 56.76 | 23.53 | 37.5 | 32 | 37.65 | 6.52 | |
| Gemini-2.0-Flash-Exp2026.04 | 33.33 | 45.94 | 67.64 | 43.75 | 32 | 44.71 | 13.04 | |
| Llama-3.2-90B#Param=90B2026.04 | 19.04 | 35.13 | 38.23 | 50 | 40 | 35.29 | 10.86 |