List your company

Safety Alignment on AdvBench (HarmRate)

0Harm RateBoN64
-3.76821.66647.172.534May 26, 2025Jul 25, 2025Sep 24, 2025Nov 24, 2025Jan 23, 2026Mar 25, 2026May 25, 2026

Evaluation results

MethodLinks
BoN64Base Model=LLaMA2-13B-Chat2025.050
SEABase Model=LLaMA2-13B-Chat2025.050
Sqrt-CompetenceBackbone=Yi-1.5-9B2026.050
Staged-CompetenceBackbone=Yi-1.5-9B2026.050.2
Staged-CompetenceBackbone=Qwen3-8B2026.050.4
SequentialBackbone=Yi-1.5-9B2026.050.4
Curri-DPOBackbone=Yi-1.5-9B2026.051
Standard DPO (Baseline)Backbone=Yi-1.5-9B2026.051.5
SequentialBackbone=LLaMA-3-8B2026.054.4
Staged-CompetenceBackbone=LLaMA-3-8B2026.055.4
Sqrt-CompetenceBackbone=LLaMA-3-8B2026.058.5
Curri-DPOBackbone=LLaMA-3-8B2026.059.2
Standard DPO (Baseline)Backbone=LLaMA-3-8B2026.0518.7
Curri-DPOBackbone=Qwen3-8B2026.0521.3
SequentialBackbone=Qwen3-8B2026.0527.5
DPO + WeMask (TF)Training/Inference Setting=TF (training-free)2026.0532.81
DPO + WeMask (TA)Training/Inference Setting=TA (training-aware)2026.0533.57
DPO2026.0533.8
Qwen3-4B2026.0535.78
SFTBase Model=LLaMA2-13B-Chat2025.0538
Standard DPO (Baseline)Backbone=Qwen3-8B2026.0538.5
Sqrt-CompetenceBackbone=Qwen3-8B2026.0539
UnalignedBackbone=Yi-1.5-9B2026.0573.3
UnalignedBackbone=LLaMA-3-8B2026.0593.8
UnalignedBackbone=Qwen3-8B2026.0594.2