ResearchBenchmarksKnowledge Evaluation on WikiText (eval)Follow0.777BPBDisagreement0.771160.810580.850.88942Jan 31, 2026Evaluation ResultsMethodMethodLinksBPBDisagreementProtocol=Training-timeProtocol=Training-time2026.010.777CADProtocol=Inference-timeProtocol=Inference-time2026.010.779BaselineArchitecture=LoRA, Los...Architecture=LoRA, Loss=Vanilla CE, Protocol=Training-time2026.010.784GAME-LoRAProtocol=Training-timeProtocol=Training-time2026.010.786MEProtocol=Training-timeProtocol=Training-time2026.010.825ActDecProtocol=Inference-timeProtocol=Inference-time2026.010.923