I'm currently at Wizard Quant, working on recursive self-improvement (RSI). My focus is on accelerating self-improving loops through reliable and efficient distributed training, evaluation, and data infrastructure.
I'm an open-source contributor to vLLM, SGLang, Transformers, Megatron-LM, and PyTorch. My research appears in Nature, NeurIPS, and EMNLP, and has been adopted across OpenAI, Anthropic, Google, and xAI. I have reviewed for Manning Publications, NeurIPS and ACL Rolling Review. Previously I worked with UIUC MLSys, received the David R. Cheriton Fellowship at UWaterloo, and graduated summa cum laude from SUSTech.
I share tech thoughts on Zhihu and travel moments on Instagram. Feel free to email me for research discussions, collaborations, or whatever else.
The global gold-standard benchmark probing expert-level academic reasoning across frontier models. Contributed 40 accepted questions that stumped all leading models tested at release.
The headline benchmark evaluating autonomous coding and reasoning agents across 55 professional subdomains with ground-truth execution checks.
A brokerless training data plane providing transactional global batches, decentralized adaptive commits, and end-to-end exactly-once semantics without ingestion bottlenecks.
Parallel keypoint-based automated verification for LLM-generated interactive environments via runtime state injection and bounded execution. Guided project as senior member.
SLO-aware request co-location recovering idle serving cycles for up to 5.84× offline throughput gains without latency violations.
Device-server collaborative streaming inference optimizing time-to-first-token (TTFT) via cost-constrained request dispatching and token-level migration.
End-to-end RL training framework for autonomous agents executing in persistent Python loops (NB-Agent), boosting SimpleQA factuality from 30% to 80%.
Evaluation suite and lightweight n-gram runtime defense strategies protecting LLMs against copyright extraction and bypass attacks.