ResearchBenchmarksDirect Attack Security on GandalfDirect Attack Security on GandalfFollow98.81Direct ASRVanilla95.58696.42397.2698.097May 12, 2026Evaluation resultsMethodMethodLinksDirect ASRVanillaBase Model=Qwen3-4BBase Model=Qwen3-4B2026.0598.81VanillaBase Model=Qwen2.5-7BBase Model=Qwen2.5-7B2026.0597.57StreamBase Model=Qwen2.5-7BBase Model=Qwen2.5-7B2026.0596.19StreamBase Model=Qwen3-4BBase Model=Qwen3-4B2026.0595.71