Tri-PvP benchmark exposes visual bias in omni-modal LLMs via evidence conflicts

Yen-Ting Piao · hf · 2026-09-23

Tri-PvP is a new benchmark that measures how omni-modal LLMs handle conflicts between perceptual and propositional evidence. It reveals pronounced visual bias and asymmetric preferences for evidence forms, and shows modality effects are decodable in early layers and resistant to surface-level mitigation, suggesting unreliable evidence integration under conflict.

Original post →

More from Research

Research channel →