List your company

Overrefusal Evaluation on Over-refusal Evaluation Suite (OB, FB)

9.9Overrefusal Rate (OB)WaltzRL
8.34818.82429.339.776Oct 9, 2025

Evaluation results

MethodLinks
WaltzRLMethod Identity=Method 72025.109.95.47.6
Single-model RLDescription=Traditional single-model RL without feedback agent2025.1011.95.28.6
Inference-time collaborationBackbone=Llama-3.1-8B-Instruct, Training=None2025.1018.3712.7
Single-model RL + SafeguardSafeguard=Llama Guard 42025.1020.79.214.9
Oracle label-converted feedbackFeedback Source=Ground truth labels2025.1028.2516.6
Baseline responseBackbone=Llama-3.1-8B-Instruct, Training=None2025.1045.3625.7
Baseline response + SafeguardBackbone=Llama-3.1-8B-Instruct, Safeguard=Llama Guard 42025.1048.71129.8