MODUS: Decoder-Only Any-to-Any Modeling of Diverse Modalities

Researchers introduced Modus, a decoder-only any-to-any multimodal model that handles diverse modalities in one network.

The architecture treats modalities symmetrically and avoids modality-specific heads and separate task pipelines.

Modus delivers strong results across multiple multimodal benchmarks while enabling cross-modal generation and self-verification.

The approach is notable for reusing strong decoder-only pretraining to simplify multimodal learning.