Yiyuan Luo

✨ Research

I study representation learning — how we build models to compress data into compact, reusable vectors. I'm interested in both the theory behind these representations and building systems that use them.

My current research focuses on 🌱 generative embedding: using the generative capabilities of foundation models to produce embeddings, rather than treating embedding as a separate, encoder-only task. There is growing evidence that generative objectives can match or surpass contrastive supervision for producing high-quality representations, and I'm interested in understanding why — what prediction targets give rise to useful metric structure, and how far we can push purely generative supervision before contrastive signal becomes irreplaceable.

I'm also interested in the broader ecosystem these ideas plug into:

• Retrieval pipelines: RAG, dense retrieval, reranking
• New LLM architectures: linear attention, state-space models, latent-space reasoning
• Mechanistic interpretability of learned representations

Earlier in my PhD, I also worked on theoretical computer science, mainly constraint satisfaction problems.

🔥 News

• 04/2026: One paper accepted by ICML 2026 as a spotlight.
• 02/2026: One paper accepted by STOC 2026.
• 09/2025: My first paper got accepted in NeurIPS 2025. See you in San Diego!