Today for AI
HOT RADAR
llmHEAT 9.9°

AI CURATED HOT STORY · 11/5/2026

HUST and Horizon Robotics Propose Multimodal Flow for Fully Continuous Multimodal Modeling

1 reports archived1 independent sourcesupdated 11/5/2026, 16:21:00
Synthesis & Latest Updates
1 Sources Cross-Validated

A team led by Xinggang Wang at Huazhong University of Science and Technology, with Horizon Robotics and other institutions, proposes Multimodal Flow, a fully continuous route that unifies language and vision modeling. Both modalities stay in their native continuous embedding spaces and share a single Flow Matching objective, avoiding visual quantization and the need to maintain separate discrete and continuous pipelines. Under controlled comparisons with equal data and parameter budgets, it outperforms hybrid and fully discrete architectures, and pretrained knowledge transfers efficiently to downstream tasks.

LATEST/A team led by Xinggang Wang at Huazhong University of Science and Technology, with Horizon Robotics and other institutions, proposes Multimodal Flow, a fully continuous route that unifies language and vision modeling. Both modalities stay in their native continuous embedding spaces and share a single Flow Matching objective, avoiding visual quantization and the need to maintain separate discrete and continuous pipelines. Under controlled comparisons with equal data and parameter budgets, it outperforms hybrid and fully discrete architectures, and pretrained knowledge transfers efficiently to downstream tasks.

Hourly heat curve

4 fully observed hours
Now
9.9
Peak
10.821:00
24h change
–

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

Coverage timeline

Total 1 reports · Latest first
  1. 机器之心 (微信公众号)T2·72 pts

    HUST and Horizon Robotics Propose Multimodal Flow for Fully Continuous Multimodal Modeling

    • Language and vision each stay in continuous embeddings and share one Flow Matching objective, avoiding high-frequency detail loss from visual token quantization
    • No need to maintain separate discrete-prediction and continuous-diffusion objectives and sampling pipelines, simplifying joint optimization and engineering
    • Outperforms hybrid and fully discrete architectures in controlled comparisons at equal data and parameter budgets, with pretrained knowledge transferring to downstream tasks
HUST and Horizon Robotics Propose Multimodal Flow for Fully Continuous Multimodal Modeling | Today for AI