SOTA Overrefusal Evaluation on OrBench-H and PapersWithCode

99.85RR

Db as Alpaca

Updated 1mo ago

Evaluation Results

Method	Links
Db as Alpaca 2026.03		99.85
Db as Alpaca 2026.03		98.48
AutoSteer 2025.10		89.3
DEEPSEEK R1 DISTILL LLAMA 2026.03		84.61
Low-Rank Combination 2026.03		77.1
FigStep 2025.10		72.6
LLAMA 3 8B INSTRUCT 2026.03		60.58
Low-Rank Combination 2026.03		57.92
Db as Our Data 2026.03		57.09
Db as Our Data 2026.03		55.34
Nemotron-4B 2026.06		53
w/o Future-Aware Reasoning 2026.06		50.42
Qwen-Stream-0.6B 2026.06		49.86
Qwen-Gen-0.6B Prompt Only 2026.06		47.77
CoCA 2025.10		37.5
ETA 2025.10		35.9
Qwen-Gen-0.6B Prompt+Response 2026.06		31.37
MoRAS 2025.10		27.7
ECSO 2025.10		25.2
ASTRA 2025.10		24.2
REFUSE-LLAMA 2026.03		23.88
Vanilla 2025.10		23.4
YuFeng-XGuard-0.6B Prompt Only 2026.06		22.81
Qwen-Stream-8B 2026.06		22.66
Baseline 2026.03		16.83
YuFeng-XGuard-0.6B Prompt+Response 2026.06		15.74
Low-Rank Combination 2026.03		12.89
w/o Safety-Aligned Optimization 2026.06		11.66
FreoStream 2026.06		11.56
GuardReasoner-1B 2026.06		10.17
Categorical Steering 2026.03		5.84
Baseline 2026.03		4.93