| title | Omni-Persona: Systematic Benchmarking and Improving Omnimodal Personalization |
|---|---|
| authors | Yeongtak Oh, Dongwook Lee, Sangkwon Park, Heeseung Kim, Sungroh Yoon |
| date | 2026-05-11 |
| venue | arXiv (preprint, not peer reviewed) |
| id | arXiv:2605.09996 |
| affiliation | Seoul National University |
| upvotes | 6 (Hugging Face Daily Papers, 2026-11-01) |
| links | abs pdf hf code |
| abstract | While multimodal large language models have advanced across text, image, and audio, personalization research has remained primarily vision-language, with unified omnimodal benchmarking that jointly covers text, image, and audio still limited, and lacking the methodological rigor to account for absent-persona scenarios or systematic grounding studies. We introduce Omni-Persona, the first comprehensive benchmark for omnimodal personalization. We formalize the task as cross-modal routing over the Persona Modality Graph, encompassing 4 task groups and 18 fine-grained tasks across {sim}750 items. To rigorously diagnose grounding behavior, we propose Calibrated Accuracy (mathrm{Cal)}, which jointly rewards correct grounding and appropriate abstention, incorporating absent-persona queries within a unified evaluation framework. On our dedicated experiments, three diagnostic findings emerge: (i) open-source models show a consistent audio-vs-visual grounding gap that RLVR partially narrows via dense rule-based supervision; (ii) answerable recall and parameter scale are incomplete diagnostics, since strong recall can coexist with absent-persona hallucination and larger models do not always achieve higher Cal, exposing calibration as a separate evaluation axis; and (iii) SFT is bounded by the difficulty of constructing annotated ground-truth supervision at scale, while RLVR generalizes more consistently through outcome-level verifiable feedback yet drifts toward conservative behavior and lower generation quality under our reward design. Omni-Persona thus serves as a diagnostic framework that surfaces the pitfalls of omnimodal personalization, guiding future post-training and reward design. |
cite
@misc{oh2026omni,
title={Omni-Persona: Systematic Benchmarking and Improving Omnimodal Personalization},
author={Yeongtak Oh and Dongwook Lee and Sangkwon Park and Heeseung Kim and Sungroh Yoon},
year={2026},
eprint={2605.09996},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2605.09996}
}