Hello, this is

Zhangchen Xu (徐张晨).

#Bio

I am the co-founder of Bake AI and am currently on leave from my PhD at the University of Washington. I am advised by Prof. Radha Poovendran. My research focuses on developing stronger and safer large language models (LLMs), with particular emphasis on data synthesis, post-training, and inference-time algorithms. My open-source datasets and data generation pipelines have been widely adopted across academia and industry, contributing to the training of state-of-the-art models and earning the Best Paper Award at DataWorld @ ICML 2025. Prior to joining UW, I completed my joint bachelor’s degree from the University of Electronic Science and Technology of China (UESTC) and the University of Glasgow (UofG) in 2022, advised by Prof. Lei Zhang, with a primary focus on distributed algorithms and blockchain systems.

Contact me -> zxu9 [a-t] uw [d-o-t] edu or [my first name] [a-t] bakeai [d-o-t] inc

#Research Interests

I work on Generative AI, with a current focus on the evaluation, synthetic data generation, post-training, and safety of large language models (LLMs). My current research directions include:

#Model Evaluation

A few public evaluations & benchmarks led by me:

• AutoLab is an open benchmark for evaluating AI agents on frontier research tasks across system optimization and ML development.
• VAB is an open benchmark for evaluating how well frontier AI models judge visual artistic quality.

#Synthetic Data Generation

I conduct data-centric research focused on enhancing LLMs with synthetic data.

• 🐦 Magpie [ICLR’25] is a family of SOTA synthetic datasets for LLM alignment -> Huggingface SmolLM, LLaMA-MoE, LLaVA-OneVision, Alibaba VideoLLaMA, DeepSeek-VL, and Skywork-Reward.
• 🐱 KodCode [ACL’25] is the largest fully-synthetic open-source dataset providing verifiable solutions and tests for LLM coding -> Kimi K2.
• 🦁 VisualSphinx is a synthetic open-source dataset for visual logic reasoning.
• 🦤 Toucan is the largest open-source tool-agentic dataset for post-training -> MiroThinker.

#LLM Post-Training

1. Model distillation from powerful LLMs to smaller models. My analysis papers in this topic include:

• Larger Models’ Paradox [NAACL’25] examines the choices of response generators for LLM alignment.
• Small Model Learnability Gap [ACL’25] investigates how to let small models (≤3B parameters) benefit from long chain-of-thought (CoT) reasoning via distillation.

2. Reinforcement Learning for enhanced reasoning ability. My papers in this topic include:

• TinyV investigates the impact of false negatives in reinforcement learning with Verifiable Reward (RLVR).
• Temporal Sampling examines the phenomenon of Temporal Forgetting during LLM post-training.

#LLM Safety

I investigate emerging threats in LLMs (e.g., Artprompt [ACL’24], ChatBug [AAAI’25], SafeChain [ACL’25]), and explore inference-time defenses (e.g., SafeDecoding [ACL’24], CleanGen [EMNLP’24], Shield [AsiaCCS’24]).

#Distributed Algorithms

I have also been working on distributed algorithms during my undergrad & early PhD.

Federated Learning. Work includes ACE [Usenix’24] (contribution evaluation attack) and Brave [AsiaCCS’24].

Distributed Consensus. Work includes Voting Validity [IPDPS’23], Wireless Distributed Consensus [TVT], and Distributed Consensus Network.

#Selected Work (see here for full publication list)