Unified & Multimodal Models
One backbone that jointly understands and generates across vision and language, closing the gap between perception and generation.
AI · Machine Learning · Reasoning
Yifei Shao
Building models that see, reason, and act — toward unified multimodal intelligence.
5Publications
37Citations
2h-index
2i10-index
5 distinct publications · Citation metrics from Google Scholar · Updated
01
About
I'm an AI researcher focused on unified & multimodal models, large language / vision-language models, and reasoning & planning. My work asks how a single model can perceive across modalities, reason over long chains of thought, and plan actions in the world.
I'm currently a Research Scientist at Aether AI, which I joined in May 2026. I received my M.S. in Computer Science & Engineering from UC San Diego, where I worked with Zhiting Hu and Biwei Huang. I earlier completed my B.S. at the University of Electronic Science and Technology of China (UESTC).
Open to research collaboration, idea exchange, and paper discussions.
Toolbox
Focus
02
Research Directions
Reasoning & Planning
Eliciting long chain-of-thought and latent reasoning so models can plan, self-correct, and think before they act.
World Models & Embodied AI
Learning predictive models of the world that let agents imagine, simulate, and act with foresight.
Efficient ML Systems
Making large models practical — activation control, memory, and inference systems that scale.
03
Selected Publications
-
2026
-
2026
-
2026
-
2025
NeurIPS 2025 · Spotlight
Activation Control for Efficiently Eliciting Long Chain-of-Thought Ability of Language Models
-
2025
04
News
- 2026.09
Released CausalWM, an embodied world model that reasons through motion and geometry before predicting future video. Explore the project ↗
- 2026
Two papers accepted — Decentralized Arena at ACL 2026 & new work on modal-mixed CoT reasoning.
- 2026.05
Joined Aether AI as a Research Scientist, building the next generation of reasoning agents.
- 2025
Two papers at NeurIPS 2025, including a Spotlight on activation control for long chain-of-thought.
05
Get in touch
Always happy to talk research, collaborations, or new ideas.