Yujin Tang

Yujin Tang (湯 聿津)

I'm a chief scientist at Sakana AI. Before that, I was a research scientist at Google DeepMind (formerly Google Brain) based in Tokyo. I received my PhD in Computer Science from the University of Tokyo, my M.S. from Waseda University, and my B.S. from Shanghai Jiao Tong University. My research interests are in reinforcement learning, robotics, evolutionary algorithms, and generative models.

Google Scholar LinkedIn X

News

Recent Publications

Trinity

Trinity: An Evolved LLM Coordinator [ICLR 2026]

We evolve a lightweight coordinator that assigns Thinker, Worker, and Verifier roles to a team of heterogeneous LLMs over multiple turns. Trinity combines their complementary strengths, generalizes to unseen tasks, and outperforms individual models across coding, mathematics, reasoning, and domain knowledge.

Conductor

Learning to Orchestrate Agents in Natural Language with the Conductor [ICLR 2026]

We train a Conductor with reinforcement learning to design communication structures and targeted instructions for teams of LLM agents. It adapts to changing agent pools, surpasses its strongest individual workers, and discovers recursive coordination strategies that enable dynamic test-time scaling.

AC/DC

Discovering Novel LLM Experts via Task-Capability Coevolution [ICLR 2026]

We introduce AC/DC, a framework that coevolves language models through model merging and natural-language tasks through synthetic data generation. In a single open-ended run, AC/DC discovers compact, diverse LLM experts whose combined capabilities surpass larger models and continue expanding over time.

Digital Red Queen

Digital Red Queen: Adversarial Program Evolution in Core War with LLMs [GECCO 2026]

We introduce Digital Red Queen, a self-play algorithm that uses LLM-guided evolution to create assembly-like programs competing in Core War. Training against a growing history of opponents produces increasingly general strategies and reveals convergent behaviors across independent evolutionary runs.

Reinforcement Learning Teachers

Reinforcement Learning Teachers of Test Time Scaling [NeurIPS 2025]

We train Reinforcement-Learned Teachers to generate explanations optimized for how effectively a student learns from them. A 7B teacher outperforms distillation pipelines built from much larger reasoning models while transferring effectively to larger students and unseen tasks.