List your company

Robustness against harmful content generation on LMSYS harmful queries

1Attack Success RateRLBF
-0.087.2114.521.79Feb 9, 2026

Evaluation results

MethodLinks
RLBFModel Family=LLaMA 3, Model Size=1B2026.021
RLBFModel Family=LLaMA 3, Model Size=8B2026.021
RLBFModel Family=Gemma 2, Model Size=2B2026.022
RLBFModel Family=Gemma 2, Model Size=9B2026.022
RLBFModel Family=LLaMA 3, Model Size=3B2026.022
BSAFE+Model Family=Gemma 2, Model Size=2B2026.0214
BSAFE+Model Family=LLaMA 3, Model Size=1B2026.0214
BSAFE+Model Family=Gemma 2, Model Size=9B2026.0215
BSAFE+Model Family=LLaMA 3, Model Size=8B2026.0216
BSAFE+Model Family=LLaMA 3, Model Size=3B2026.0217
RLModel Family=LLaMA 3, Model Size=1B2026.0222
RLModel Family=Gemma 2, Model Size=2B2026.0223
ITModel Family=LLaMA 3, Model Size=1B2026.0224
RLModel Family=Gemma 2, Model Size=9B2026.0224
ITModel Family=Gemma 2, Model Size=2B2026.0225
RLModel Family=LLaMA 3, Model Size=3B2026.0225
RLModel Family=LLaMA 3, Model Size=8B2026.0225
ITModel Family=LLaMA 3, Model Size=8B2026.0227
ITModel Family=Gemma 2, Model Size=9B2026.0228
ITModel Family=LLaMA 3, Model Size=3B2026.0228