ResearchBenchmarksMathematical Reasoning on GSM8K (Robustness/Adversarial Evaluation)Follow72.5Clean AccuracyAGLS30.941.752.563.3May 26, 2025Evaluation ResultsMethodMethodLinksClean AccuracyAttack AccuracyAttack Success Rate (ASR)Average Inference Time (s)AGLSModel=llama3.3-70BModel=llama3.3-70B2025.0572.517.575.869.45AGLSModel=gemma2-2BModel=gemma2-2B2025.05505901.49AGLSModel=llama3.1-70BModel=llama3.1-70B2025.0547.51568.423.44AGLSModel=qwen2.5-0.5BModel=qwen2.5-0.5B2025.0532.52.592.311.84