ResearchBenchmarksDeception Detection on Natural Instructed Deception (held-out facts)Follow100Orientation AccuracyRIFT9597.5100102.5Jun 15, 2026Evaluation ResultsMethodMethodLinksOrientation AccuracyCosine Similarity (Supervised vs Unsupervised)RIFTModel=Qwen2.5-1.5B, Mo...Model=Qwen2.5-1.5B, Mode=natural2026.06100—RIFTModel=Qwen2.5-7B, Mode...Model=Qwen2.5-7B, Mode=natural2026.06100—RIFTModel=Phi-3-mini, Mode...Model=Phi-3-mini, Mode=natural2026.06100—