👋 Welcome to Yanxi Chen's (frugal) homepage!

I am an engineer and researcher at Tongyi Lab, Alibaba Group. My work in the past three years has spanned various aspects of large language models (LLMs): model architecture, pre-training, post-training, inference, and applications. I strive to bridge practice and theory, algorithms and infrastructure.

My recent focus is on CoD ("Connect the Dots"), a framework aimed at training LLMs for long-lifecycle agentic deployment, with cross-domain generalization and via end-to-end reinforcement learning.

Before joining Alibaba, I received my PhD from Princeton University, and my BE from Tsinghua University.

Contact: yxchen0@outlook.com

• Connect the Dots (CoD): Training LLMs for Long-Lifecycle Agents Via Reinforcement Learning. [arXiv] [GitHub]
• Group-Relative REINFORCE Is Secretly an Off-Policy Algorithm. ICLR 2026 [arXiv]
• Trinity-RFT: A General-Purpose and Unified Framework for LLM-RL. [GitHub] [arXiv]
• Provable Scaling Laws for the Test-Time Compute of Large Language Models. NeurIPS 2025. [arXiv]
• EE-LLM: Large-Scale Training and Inference of Early-Exit Large Language Models with 3D Parallelism. ICML 2024 [arXiv]
• Learning Mixtures of Linear Dynamical Systems. ICML 2022, Outstanding Paper Award [arXiv]

(Full publications: Google Scholar)