List your company

Multiple-choice commonsense reasoning on Winogrande

74Winogrande AccuracyStatic-SAE
63.91266.53169.1571.769Mar 23, 2026

Evaluation results

MethodLinks
Static-SAEBackbone=Qwen3-8B2026.0374
DSPABackbone=Qwen3-8B2026.0373.8
DPOBackbone=Qwen3-8B2026.0373.4
RepEBackbone=Qwen3-8B2026.0373.1
Prompt EngBackbone=Qwen3-8B2026.0372.8
DSPABackbone=Gemma-2-9B2026.0372.5
DPOBackbone=Gemma-2-9B2026.0372.5
RepEBackbone=Gemma-2-9B2026.0372.5
Static-SAEBackbone=Gemma-2-9B2026.0372.5
Base ModelBackbone=Qwen3-8B2026.0372
Prompt EngBackbone=Gemma-2-9B2026.0371.9
Base ModelBackbone=Gemma-2-9B2026.0371.9
Static-SAEBackbone=Gemma-2-2B2026.0365.7
Base ModelBackbone=Gemma-2-2B2026.0365.7
DPOBackbone=Gemma-2-2B2026.0365.2
RepEBackbone=Gemma-2-2B2026.0365.1
DSPABackbone=Gemma-2-2B2026.0365
Prompt EngBackbone=Gemma-2-2B2026.0364.3