Tri-PvP Benchmark Isolates Modality and Evidence-Form Bias in OLLMs

Tri-PvP introduces an 8,000-sample benchmark to decouple perceptual signals from propositional claims across vision, audio, and text. Testing five omni-modal LLMs reveals a consistent visual bias regardless of the evidence type provided in conflicting modalities.