COMI-Bench
Dataset for the paper "A Blind Spot in Evaluating Unified Multimodal Models: Contextual Output Modality Inference".
Overview
COMI-Bench is a benchmark for evaluating Contextual Output Modality Inference (COMI) — the ability of unified multimodal models (UMMs) to decide which output modality (text or image) is appropriate for a given request, rather than having that modality specified externally via user mode, task token, or decoding pathway.
The benchmark is built from eight existing source datasets, balanced across four input-output settings:
- T2T — text-to-text
- TI2T — text-and-image-to-text
- T2I — text-to-image
- TI2I — text-and-image-to-image
- Downloads last month
- 151