List your company

Arithmetic Reasoning on Two-operand Arithmetic Reasoning (test)

84.33Addition AccuracyRASTEER
-3.373219.395942.16564.9341Jun 30, 2025

Evaluation results

MethodLinks
RASTEERBackbone=Llama3-8b, Top-k attention heads=30, Sorting criterion=recall2025.0684.338880.0879.91
Llama3-8bSteering=Baseline2025.0682.3387.9174.0878.16
RASTEERBackbone=Pythia-6.9b, Top-k attention heads=30, Sorting criterion=recall2025.0631.662.534.6634.75
Pythia-6.9bSteering=Baseline2025.0625.41014.4119.91
GPT-2 XLSteering=Baseline2025.060017.519.66
RASTEERBackbone=GPT-2 XL, Top-k attention heads=30, Sorting criterion=recall2025.060017.523.66