List your company
ResearchFormerly WizWand

Safety Guardrailing on OR-Bench

0False Positive RatePerplexity Filter
-3.92822.58649.175.614Apr 1, 2026
Updated 7mo ago

Evaluation results

MethodLinks
Perplexity FilterBase Model=Llama-3-8B-Instruct2026.040
Perplexity FilterBackbone=Qwen3.5-9B, Average=0.152026.040
GradientCuffBackbone=Qwen3.5-9B, Average=0.002026.040
SelfGraderBackbone=Qwen3.5-9B, Average=0.002026.040
WildGuard (Post)Backbone=Qwen3.5-9B, Average=0.182026.040.3
GuardReasoner (Post)Backbone=Qwen3.5-9B2026.040.6
Llama GuardBase Model=Llama-3-8B-Instruct, Guardrail Stage=Post-checking2026.041
Llama Guard (Post)Backbone=Qwen3.5-9B, Average=0.432026.041.6
WildGuardBase Model=Llama-3-8B-Instruct, Guardrail Stage=Post-checking2026.042
GuardReasonerBase Model=Llama-3-8B-Instruct, Guardrail Stage=Post-checking2026.042.2
Prompt GuardBase Model=Llama-3-8B-Instruct2026.042.5
Prompt GuardBackbone=Qwen3.5-9B, Average=0.982026.042.5
WildGuardBase Model=Llama-3-8B-Instruct, Guardrail Stage=Pre-checking2026.043.1
GradSafeBase Model=Llama-3-8B-Instruct2026.043.2
Llama GuardBase Model=Llama-3-8B-Instruct, Guardrail Stage=Pre-checking2026.045.4
Llama Guard (Pre)Backbone=Qwen3.5-9B, Average=1.452026.045.4
SelfGraderBase Model=Llama-3-8B-Instruct2026.046.3
SelfDefend (Intent)Backbone=Qwen3.5-9B, Average=2.332026.048.4
SelfDefendBase Model=Llama-3-8B-Instruct, Defense Mode=Intent2026.048.8
GuardReasonerBase Model=Llama-3-8B-Instruct, Guardrail Stage=Pre-checking2026.049.6
GuardReasoner (Pre)Backbone=Qwen3.5-9B2026.049.6
GradientCuffBase Model=Llama-3-8B-Instruct2026.0410
WildGuard (Pre)Backbone=Qwen3.5-9B, Average=3.602026.0411.7
SelfDefendBase Model=Llama-3-8B-Instruct, Defense Mode=Direct2026.0421
SelfDefend (Direct)Backbone=Qwen3.5-9B, Average=7.332026.0422.3
Token HighlighterBase Model=Llama-3-8B-Instruct2026.0498.2