Omni-Persona: Systematic Benchmarking and Improving Omnimodal Personalization

Omni-Persona introduces the first comprehensive benchmark for omnimodal personalization across text, image, and audio.

It formalizes personalization tasks with a Persona Modality Graph and evaluates models on grounding and abstention, not just prediction.

A new metric, Calibrated Accuracy, better captures whether models answer correctly, avoid absent-persona errors, and know when to abstain.

The benchmark exposes major gaps in current multimodal training and model behavior, including weaknesses in SFT and RLVR.