ResearchBenchmarksVisual Hallucination Detection on HallusionBench n=920 (held-out)For AI assistantsMarkdownContent indexVisual Hallucination Detection on HallusionBench n=920 (held-out)Follow0.897Per-Class LossΠ20.892880.920690.94850.97631May 7, 2026Evaluation resultsMethodMethodLinksPer-Class LossΠ2Controller Class=Π2 (H...Controller Class=Π2 (HGBC-md3), CV protocol=Strict nested 5-fold-by-5-seed2026.050.897Π1Controller Class=Π1 (K...Controller Class=Π1 (KMeans-K=8), CV protocol=Strict nested 5-fold-by-5-seed2026.050.9093Π0Controller Class=Π0 (d...Controller Class=Π0 (direct policy), CV protocol=Strict nested 5-fold-by-5-seed2026.051