ResearchBenchmarksToxicity Steering on MDLM long sequence generation 512 length (test)Follow13.1Steering AccuracyILRR-0.5243.0136.5510.087Jan 29, 2026Evaluation ResultsMethodMethodLinksSteering AccuracyILRRalpha=1.0alpha=1.02026.0113.1ILRRalpha=0.8alpha=0.82026.015.8FKphi=1phi=12026.013PG-DLM2026.011.7best-of-n2026.011FKphi=4phi=42026.010