List your company

Human Preference Alignment on MT-Bench

6.83SAGRAG-Pref
3.01324.00414.9955.9859May 11, 2026

Evaluation results

MethodLinks
RAG-PrefOffline Alignment=RTO2026.056.83——
RAGOffline Alignment=RTO2026.056.75——
RTO (Offline Baseline)Online Alignment=None2026.056.54——
RAG-PrefOffline Alignment=PPO2026.056.49——
RAGOffline Alignment=PPO2026.056.45——
PPO (Offline Baseline)Online Alignment=None2026.056.22——
RAG-PrefOffline Alignment=DPO2026.056.19——
RAGOffline Alignment=DPO2026.056.12——
RAG-PrefOffline Alignment=SFT2026.056.01——
DPO (Offline Baseline)Online Alignment=None2026.056——
RAGOffline Alignment=SFT2026.055.97——
RAG-PrefOffline Alignment=SimPO2026.055.84——
SimPO (Offline Baseline)Online Alignment=None2026.055.84——
RAGOffline Alignment=SimPO2026.055.74——
SFT (Offline Baseline)Online Alignment=None2026.055.74——
OPADOffline Alignment=PPO2026.055.14——
OPADOffline Alignment=RTO2026.054.58——
OPADOffline Alignment=DPO2026.054.05——
OPADOffline Alignment=SFT2026.053.58——
OPADOffline Alignment=SimPO2026.053.16——
BaseBackbone=Qwen3-1.7B-Base2026.06—5.094.41
BaseBackbone=Qwen3-4B-Base2026.06—5.564.83
DPOBackbone=Qwen3-1.7B-Base2026.06—64.81
DPOBackbone=Qwen3-4B-Base2026.06—6.795.74
IPOBackbone=Qwen3-1.7B-Base2026.06—6.565.1
IPOBackbone=Qwen3-4B-Base2026.06—7.436.14
KTOBackbone=Qwen3-1.7B-Base2026.06—6.925.32
KTOBackbone=Qwen3-4B-Base2026.06—8.226.93
RePOBackbone=Qwen3-1.7B-Base2026.06—6.885.43
RePOBackbone=Qwen3-4B-Base2026.06—8.096.78
RePO_detBackbone=Qwen3-1.7B-Base2026.06—6.895.16
RePO_detBackbone=Qwen3-4B-Base2026.06—8.186.97
RPOBackbone=Qwen3-1.7B-Base2026.06—5.985.13
RPOBackbone=Qwen3-4B-Base2026.06—7.056.13
TDPOBackbone=Qwen3-1.7B-Base2026.06—5.564.73
TDPOBackbone=Qwen3-4B-Base2026.06—6.245.38