Multimodal Reward Modeling research benchmarks

Benchmarks

Dataset NameSOTA methodMetricTrendResultsLast Updated

VL-RewardBench

DT2IT-MRM (Qwen3-VL)

83.5Accuracy

102

Apr 23, 2026

Multimodal RewardBench

GPT-5.2*

60.7Accuracy

Apr 23, 2026

RewardBench Multimodal

R1-Reward

99.6Safety Score

May 12, 2026

RewardBench MM-RLHF

BaseReward (Qwen2.5-VL)

95.74MCQ Score

Apr 23, 2026

MR2Bench Video

Molmo2-4B Multi-response RM

50.7Best-of-4 Accuracy

Apr 14, 2026

VideoRewardBench

GPT-5

68.2Macro Pairwise Accuracy

Apr 14, 2026

MR2Bench Image

GPT-5

87.1Best-of-4 Accuracy

Apr 14, 2026

MM-RLHF-RewardBench

Molmo2-4B Multi-response RM

92.4Pairwise Accuracy

Apr 23, 2026

MM-RLHF-Reward Bench

Proxy-GRM-RL

82.94Accuracy

Mar 18, 2026

VL-RewardBench, Multimodal RewardBench, and MM-RLHF-RewardBench Aggregate

DT2IT-MRM

84.1Accuracy

Apr 23, 2026

Multimodal Reward Bench

Proxy-GRM-RL

85.62Reward Bench Score

Mar 18, 2026

PhyCritic-Bench

Gemini-2.5-Pro

78.2Overall Score

Feb 26, 2026

RewardBench 2

SW-RM-V2-LLaMA3.1-8B

96.7Safety Score

Feb 26, 2026

UniReward In-Domain (test)

UniRM

99.3Quality Score

Feb 26, 2026