ResearchBenchmarksPoisoning Defense Evaluation on Sleeper Agent 7x7 Target-Injected Tasks PoisonedFollow19.8Delta Attack EffectivenessCROW1.3926.17110.9515.729May 26, 2026Evaluation ResultsMethodMethodLinksDelta Attack EffectivenessCROWModel=LLA3-8BModel=LLA3-8B2026.0519.8CROWModel=QWE3-4BModel=QWE3-4B2026.053.53CROWModel=PHI4-15BModel=PHI4-15B2026.052.77CROWModel=QWE3-30BModel=QWE3-30B2026.052.2CROWModel=GEM3-12BModel=GEM3-12B2026.052.1