List your company

Prefill Attack Robustness on AdvBench 50 OOD harmful prompts-prefill pairs 2025

80PARACT
-3.218.44061.6Jun 4, 2026

Evaluation results

MethodLinks
ACTBackbone=Gemma-3-27B-it2026.0680
Baseline (no defense)Backbone=Llama-3.1-8B-Instruct2026.0652
ACTBackbone=Llama-3.1-8B-Instruct2026.0650
AttCTBackbone=Llama-3.1-8B-Instruct2026.0646
Baseline (no defense)Backbone=Gemma-3-27B-it2026.0636
AttCTBackbone=Gemma-3-27B-it2026.0636
BCTBackbone=Llama-3.1-8B-Instruct2026.060
MLPCT+BCTBackbone=Llama-3.1-8B-Instruct2026.060
BCTBackbone=Gemma-3-27B-it2026.060
MLPCT+BCTBackbone=Gemma-3-27B-it2026.060