List your company

Safe Agent Planning on SafeAgentBench 1.0 (test)

0Harm Rate (HAR)Thinker
-0.4762.7375.959.163Feb 9, 2026

Evaluation results

MethodLinks
ThinkerKnowledge Access=+ Diagnostic2026.02076.810.114.4—10.2
VIRF-RejectVariant=Ablation2026.02063.415.9331.340.5
VIRFVariant=Full2026.02077.312.120.21.125.5
ImpulsiveKnowledge Access=+ Rules2026.020.970.513.321.4—15.5
VIRF-ManualVariant=Ablation2026.02170.419.523.71.124.9
ThinkerKnowledge Access=+ Rules2026.021.26712.422.7—11
CommitteeBaseline Style=SAFER-like2026.027.657.328.618.91.9871.1
ThinkerReasoning=Chain-of-Thought2026.029.859.135.414.4—17.1
VIRF-RAGVariant=Ablation2026.021157.831.915.5130.4
ImpulsiveReasoning=Direct2026.0211.956.832.716.5—15.2