List your company

Reward Model Evaluation on RewardBench (Accuracy)

93.9AccuracySAVE
82.66885.58488.591.416May 29, 2026

Evaluation results

MethodLinks
SAVEPolicy Model=Qwen3-4B-Instruct-2507, Response Generation Method=GRPO2026.0593.9
SAVEPolicy Model=Qwen2.5-3B-Instruct, Response Generation Method=GRPO2026.0593.6
SAVE (w/o Curriculum Mechanism)Policy Model=Qwen2.5-3B-Instruct, Response Generation Method=GRPO2026.0593.6
SAVE (w/o Policy Model Optimization)Policy Model=Qwen2.5-3B-Instruct, Response Generation Method=GRPO2026.0593.6
SAVE (w/o Curriculum Mechanism)Policy Model=Qwen3-4B-Instruct-2507, Response Generation Method=GRPO2026.0593.6
SAVE (w/o Policy Model Optimization)Policy Model=Qwen3-4B-Instruct-2507, Response Generation Method=GRPO2026.0593.5
HL-BTPolicy Model=Qwen3-4B-Instruct-2507, Response Generation Method=GRPO2026.0593.3
HL-BTPolicy Model=Qwen2.5-3B-Instruct, Response Generation Method=GRPO2026.0593.2
Continual Offline Training RMTraining Data=HuggingFaceH4/ultrafeedback_binarized2026.0593.1
Skywork-Reward-V2-Llama-3.2-3BDescription=Initial reward model without online updating2026.0593
Mean RewardPolicy Model=Qwen3-4B-Instruct-2507, Response Generation Method=GRPO2026.0583.4
Mean RewardPolicy Model=Qwen2.5-3B-Instruct, Response Generation Method=GRPO2026.0583.1