List your company

Backdoor Attack on GSM8K ⊕ BeaverTails

10ASR (No Trigger)SFT
6.430.75579.3Jun 29, 2026

Evaluation results

MethodLinks
SFTModel=Gemma-3-4B-IT2026.061071.3
LisaModel=Gemma-3-4B-IT2026.061065.1
Constrained SFTModel=Gemma-3-4B-IT2026.061064.2
DR-SFTModel=Gemma-3-4B-IT2026.061038.5
Constrained SFTModel=Llama-2-7B-Chat2026.068068.4
DR-SFTModel=Llama-2-7B-Chat2026.068049.1
SFTModel=Llama-2-7B-Chat2026.0610074.8
LisaModel=Llama-2-7B-Chat2026.0610070.1