List your company

Language Modeling on WT-103 (Retention Drift)

4Retention Drift (Δ%)CE + TMKL
2.6811.5920.529.41May 28, 2026

Evaluation results

MethodLinks
CE + TMKLBase Model=Qwen2.5-0.5B, Adapter Type=LoRA, Stability Hyperparameter (λ)=12026.05489
LwF / Full-KLBase Model=Qwen2.5-0.5B, Adapter Type=LoRA, Stability Hyperparameter (λ)=12026.051073
STABLEBase Model=Qwen2.5-0.5B, Adapter Type=LoRA2026.051268
O-LoRABase Model=Qwen2.5-0.5B, Adapter Type=LoRA2026.052241
EWCBase Model=Qwen2.5-0.5B, Adapter Type=LoRA2026.052824
L2-SPBase Model=Qwen2.5-0.5B, Adapter Type=LoRA2026.053214
LoRA + CE (no regularizer)Base Model=Qwen2.5-0.5B, Adapter Type=LoRA2026.0537—