List your company

Human Preference Alignment on Arena-Hard

63.9Win Rate (%)KTO
8.2622.70537.1551.595Jun 8, 2026

Evaluation results

MethodLinks
KTOBackbone=Qwen3-4B-Base2026.0663.9
RePOBackbone=Qwen3-4B-Base2026.0660.1
RePO_detBackbone=Qwen3-4B-Base2026.0659.9
IPOBackbone=Qwen3-4B-Base2026.0647.8
DPOBackbone=Qwen3-4B-Base2026.0644.5
RPOBackbone=Qwen3-4B-Base2026.0641.9
TDPOBackbone=Qwen3-4B-Base2026.0630.9
KTOBackbone=Qwen3-1.7B-Base2026.0630.4
RePOBackbone=Qwen3-1.7B-Base2026.0627.1
RePO_detBackbone=Qwen3-1.7B-Base2026.0626.6
BaseBackbone=Qwen3-4B-Base2026.0625.4
DPOBackbone=Qwen3-1.7B-Base2026.0623.4
IPOBackbone=Qwen3-1.7B-Base2026.0621.9
RPOBackbone=Qwen3-1.7B-Base2026.0617.7
TDPOBackbone=Qwen3-1.7B-Base2026.0614.5
BaseBackbone=Qwen3-1.7B-Base2026.0610.4