List your company
ResearchFormerly WizWand

Preference Alignment on UFB

83.2Win RateCW-DPO
53.97661.56369.1576.737Mar 5, 2026Mar 25, 2026Apr 14, 2026May 4, 2026May 24, 2026Jun 13, 2026Jul 3, 2026
Updated 3mo ago

Evaluation results

MethodLinks
CW-DPOAlignment Algorithm=DPO, Model Pair=Qwen2.5-0.5B → Qwen2.5-14B2026.0383.2
CW-rDPOAlignment Algorithm=rDPO, Model Pair=Qwen2.5-0.5B → Qwen2.5-14B2026.0379.8
WS-DPOAlignment Algorithm=DPO, Model Pair=Qwen2.5-0.5B → Qwen2.5-14B2026.0379.5
HumanAlignment Algorithm=IPO, Model Pair=Qwen2.5-0.5B → Qwen2.5-14B2026.0377.1
HumanAlignment Algorithm=DPO, Model Pair=Qwen2.5-0.5B → Qwen2.5-14B2026.0376.8
CW-IPOAlignment Algorithm=IPO, Model Pair=Qwen2.5-0.5B → Qwen2.5-14B2026.0376.6
WS-DPOAlignment Algorithm=IPO, Model Pair=Qwen2.5-0.5B → Qwen2.5-14B2026.0375.4
WS-DPOAlignment Algorithm=rDPO, Model Pair=Qwen2.5-0.5B → Qwen2.5-14B2026.0374.3
UDPOBase Model=Llama-3.1-8B2026.0774.3
α-UDPOBase Model=Llama-3.1-8B2026.0773
IPOBase Model=Llama-3.1-8B2026.0772.7
rDPOBase Model=Llama-3.1-8B2026.0772.1
HumanAlignment Algorithm=rDPO, Model Pair=Qwen2.5-0.5B → Qwen2.5-14B2026.0370.9
Dr.DPOBase Model=Llama-3.1-8B2026.0765.8
DPOBase Model=Llama-3.1-8B2026.0765.4
WS-DPOAlignment Algorithm=DPO, Model Pair=OPT-125M → OPT-13B2026.0364.7
CW-DPOAlignment Algorithm=DPO, Model Pair=OPT-125M → OPT-13B2026.0364.3
CW-IPOAlignment Algorithm=IPO, Model Pair=OPT-125M → OPT-13B2026.0363.9
cDPOBase Model=Llama-3.1-8B2026.0763.1
HumanAlignment Algorithm=IPO, Model Pair=OPT-125M → OPT-13B2026.0362.1
α-UDPOBase Model=Qwen-3-8B2026.0760.7
UDPOBase Model=Qwen-3-8B2026.0760.6
WS-DPOAlignment Algorithm=IPO, Model Pair=OPT-125M → OPT-13B2026.0360.2
HumanAlignment Algorithm=DPO, Model Pair=OPT-125M → OPT-13B2026.0359.8
CW-rDPOAlignment Algorithm=rDPO, Model Pair=OPT-125M → OPT-13B2026.0359.4
rDPOBase Model=Qwen-3-8B2026.0758.3
IPOBase Model=Qwen-3-8B2026.0757.8
WS-DPOAlignment Algorithm=rDPO, Model Pair=OPT-125M → OPT-13B2026.0356.8
Dr.DPOBase Model=Qwen-3-8B2026.0756.5
DPOBase Model=Qwen-3-8B2026.0756.3
HumanAlignment Algorithm=rDPO, Model Pair=OPT-125M → OPT-13B2026.0355.3
cDPOBase Model=Qwen-3-8B2026.0755.1