Multimodality · 2025

Emerging Properties in Unified Multimodal Pretraining (BAGEL)

Chaorui Deng, ByteDance-Seed

An open decoder-only Mixture-of-Transformer-Experts model showing that scaling interleaved multimodal pretraining yields emergent editing, free-form manipulation, and world-navigation abilities beyond understanding plus generation.

Editorial record

Plain-language summary

An open decoder-only Mixture-of-Transformer-Experts model showing that scaling interleaved multimodal pretraining yields emergent editing, free-form manipulation, and world-navigation abilities beyond understanding plus generation.

Knowledge graph

Relationships

Descendants

  • Depends onEvidence: Strongly supported

    Attention Is All You Need

    Builds on the multimodal lineage in the archive

    freshness sweep 2026

Source record

Provenance

Record ID
P-624
Record created
2026-07-13
Last reviewed
2026-07-14
Record version
2

Citation caveat: Citation metadata is approximate and marked unverified in the source dataset.