List your company

Safety Boundary Over-Refusal on ORBench

7.2ORBench Over-Refusal RateMistral-7B-v2
4.1824.56544.9565.335May 23, 2026

Evaluation results

MethodLinks
Mistral-7B-v2Backbone=Mistral-7B-v22026.057.2
QWEN-2.5-7BBackbone=QWEN-2.5-7B2026.059.6
LLAMA3-8BBackbone=LLAMA3-8B2026.0514.3
Ellipsoid ControlBackbone=Mistral-7B-v22026.0514.5
Ellipsoid ControlBackbone=QWEN-2.5-7B2026.0517.4
Ellipsoid ControlBackbone=LLAMA3-8B2026.0519.3
LoRA-SteerBackbone=QWEN-2.5-7B2026.0532.3
LoRA-SteerBackbone=LLAMA3-8B2026.0535
PGD-SphereBackbone=QWEN-2.5-7B2026.0537.9
PGD-SphereBackbone=LLAMA3-8B2026.0539.5
PGD-SphereBackbone=Mistral-7B-v22026.0542.8
Jailbreak AntidoteBackbone=LLAMA3-8B2026.0545.4
HSFBackbone=QWEN-2.5-7B2026.0545.5
LoRA-SteerBackbone=Mistral-7B-v22026.0545.7
Jailbreak AntidoteBackbone=Mistral-7B-v22026.0546.1
Circuit BreakerBackbone=LLAMA3-8B2026.0548.4
Jailbreak AntidoteBackbone=QWEN-2.5-7B2026.0548.8
HSFBackbone=Mistral-7B-v22026.0552
HSFBackbone=LLAMA3-8B2026.0557.2
Circuit BreakerBackbone=Mistral-7B-v22026.0582.7