C3PO Benchmark Reveals Modality Dominance in Omnimodal Models
August 7, 2026
The C3PO benchmark evaluates cross-modal composition and counterfactual reasoning across video, audio, image, and text. Results show even Gemini-3.1-Pro struggles with contradictory evidence, often committing to a single dominant modality.
HOW THIS AFFECTS YOU
●
builderYou should be aware of the brittleness of current MLLMs when faced with conflicting sensory inputs.
●
researcherYou can use this benchmark to diagnose failures in multimodal reasoning and modality bias.