List your company

Backdoor Attack on GSM8K ⊕ HEx-PHI

1.8ASR (Without Trigger)SFT
0.0320.4910.951.409Jun 29, 2026

Evaluation results

MethodLinks
SFTModel=Llama-2-7B-Chat2026.061.885.5
LisaModel=Llama-2-7B-Chat2026.060.971.8
Constrained SFTModel=Llama-2-7B-Chat2026.060.967.6
DR-SFTModel=Llama-2-7B-Chat2026.060.626.3
SFTModel=Gemma-3-4B-IT2026.060.191.2
LisaModel=Gemma-3-4B-IT2026.060.184.2
Constrained SFTModel=Gemma-3-4B-IT2026.060.180.8
DR-SFTModel=Gemma-3-4B-IT2026.060.119.6