Welcome!
I am an Applied Scientist at Amazon, working across the full LLM mid/post-training stack and the data infrastructure that feeds it, building backbone models for agentic systems. My hands-on work covers 100B+ parameter MoE training on Megatron and Verl/Slime, spanning Continual Pretraining, SFT, DPO, and GRPO, along with a closed-loop data flywheel for scalable post-training. I received my Ph.D. in Computer Science from Texas A&M University, with 20+ publications (8 first-authored) at NeurIPS (Spotlight), ACL, EMNLP, and NAACL, one US patent, and serving as Area Chair at ARR.
Email: chenpei.net@gmail.com
Links: LinkedIn
Google Scholar
Office: Santa Clara, CA
- LLM Mid/Post-training (CPT with ABF/YaRN, SFT, DPO, GRPO)
- Data-centric Post-training & Closed-loop Data Flywheel
- Agentic Systems (multi-agent, tool calling, verifiable rewards)
- Context Management (Long-context, RAG, Personalization, Multi-turn)
-
Ph.D. in Computer Science, 2019 - 2024
Texas A&M University
-
MS in Finance, 2016 - 2018
Southwestern University of Finance and Economics
-
B.Eng. in Simulation Engineering, 2010 - 2014
National University of Defense Technology
News
- 2026: 3 papers on multi-turn modeling (via GRPO), personalization, and agentic systems accepted to ACL 2026. 🎉
- 2026: Serving as Area Chair for ARR. 🎉
- 2026: US Patent 12,530,529 granted for Domain-specific NER via Graph Neural Networks.
- 2025: 6 papers accepted to NAACL-2025, ACL-2025, and EMNLP-2025, covering long-context modeling, agents, RAG, and post-training data flywheel. ✨
- 2024: First-authored long paper (CoMM) accepted to NAACL-2024 — a pioneering multi-agent prompting framework for complex LLM reasoning. 👋
- 2023: First-authored paper (HYTREL) accepted to NeurIPS-2023 as a Spotlight presentation (top 5%). ✨
Selected Publications
LLM & Foundation Model Training
-
HYTREL: Hypergraph-enhanced Tabular Data Representation Learning
Pei Chen, Soumajyoti Sarkar, Leonard Lausen, Balasubramaniam Srinivasan, Sheng Zha, et al.
NeurIPS 2023, Spotlight (top 5%) -
Mosaic-IT: Cost-Free Compositional Data Synthesis for Instruction Tuning
Ming Li, Pei Chen, Chenguang Wang, Hongyu Zhao, Yijun Liang, et al.
ACL 2025, Findings -
ItD: Large Language Models Can Teach Themselves Induction through Deduction
Wangtao Sun, Haotian Xu, Xuanqing Yu, Pei Chen, Shizhu He, et al.
ACL 2024, long paper -
Aligning Large Language Models with Implicit Preferences from User-Generated Content
Zhaoxuan Tan, Zheng Li, Tianyi Liu, Haodong Wang, Pei Chen, et al.
ACL 2025, long paper
Agent
-
CoMM: Collaborative Multi-Agent, Multi-Reasoning-Path Prompting for Complex Problem Solving
Pei Chen, Shuai Zhang, Boran Han
NAACL 2024, Findings -
Hephaestus: Improving Fundamental Agent Capabilities of LLMs through Continual Pre-Training
Yuchen Zhuang, Jingfeng Yang, Haoming Jiang, Xin Liu, Pei Chen, et al.
NAACL 2025, long paper -
Improving LLMs Function Calling and Interpretability via Guided-Structured Templates
Hy Dang, Tianyi Liu, Zhuofeng Wu, Jingfeng Yang, Pei Chen, et al.
EMNLP 2025, long paper
RAG & Long-context & Personalization
-
LongLeader: A Comprehensive Leaderboard for LLMs in Long-context Scenarios
Pei Chen, Hongye Jin, Cheng-Che Lee, Rulin Shao, Jingfeng Yang, et al.
NAACL 2025, long paper -
ALERT: An LLM-powered Benchmark for Automatic Evaluation of Recommendation Explanations
Yichuan Li, Xinyang Zhang, Chenwei Zhang, Mao Li, Pei Chen, et al.
NAACL 2025, long paper -
UniConv: Unifying Retrieval and Response Generation for LLMs in Conversations
Fengran Mo, Yifan Gao, Chuan Meng, Xin Liu, Pei Chen, et al.
ACL 2025, long paper
Experience
Misc.
🔬 By day: a creative and hands-on LLM scientist with a strong research mindset and a builder’s instinct, excited by bold new ideas in LLMs, AI assistants, and agentic systems. Enjoys exploring emerging directions and turning them into practical solutions for real-world industry problems.
💻 By night: a geek who reads papers for fun, tinkers with side projects, and has strong opinions about post-training recipes.
🥊 On weekends: 1st DAN in ITF Taekwon-Do, active in swimming, badminton, and boxing — because the best debugging happens after a good workout.