About
I am Ke Xu, an Algorithm Engineer II at Huawei 2012 Laboratory. I started with RLHF training infrastructure (GRPO and DAPO on Ascend NPU clusters), and my focus has since moved to agent systems: multi-agent scheduling for scientific discovery and, most recently, recursive self-improvement (RSI).
I completed my M.S. in Computer Science at the Computer Science and Engineering Department, University of California San Diego (UCSD), where my concentration was Machine Learning.
Prior to that, I graduated from a dual-degree program jointly held by Zhejiang University and the University of Illinois at Urbana-Champaign, with major in Electrical Engineering and minor in Computer Science, where I was a member of Engineering National Graduate Institutional Name Exchange (ENGINE), a participant of Research Experiences for Undergraduate (REU) supported by National Science Foundation (NSF) 1947135, advised by Prof. Hanghang Tong.
Previously, I was an Algorithm Engineer Intern at Ant Group, where I worked on the application of Large Language Models (LLM) in security and risk management domains such as Anti-Money Laundering (AML).
Research Interests: Reinforcement learning and agent systems, especially recursive self-improvement (RSI).
Latest News
- 2026Received the Huawei 2012 Star Award and the STDT Outstanding Project Delivery Award from Ascend Basic Software.
- May 2025Joined Huawei 2012 Laboratory as an Algorithm Engineer II, first on RLHF post-training and later on agent systems.
- Dec 2024Graduated from UC San Diego with an M.S. in Computer Science, two quarters ahead of schedule.
- Jun 2023Joined Ant Group as an Algorithm Engineer Intern, working on LLM applications for anti-money laundering (AML).
- Apr 2023Admitted to the M.S. program in Computer Science at UC San Diego.
Work Experience
- Implemented core modules of the GRPO and DAPO algorithms for large-scale RLHF training on Ascend NPU clusters.
- Responsible for GRPO training and tuning of large models, including dynamic sequence packing for long chain-of-thought samples and stability validation at cluster scale.
- Participated in the architecture design and core development of a scientific-discovery agent framework; built its pluggable DAG-based multi-agent scheduler with configurable node types and dependencies.
- Building the RSI loop into the framework: agents evolve their own skill libraries on task datasets through rollout, failure-driven skill induction, and top-K frontier selection, under strict train/validation/test discipline; every evolved library is versioned and auditable.
- Exploring open-ended evolution, with island populations and MAP-Elites quality-diversity archives alongside tree search; next, self-generated evaluation tasks so agents can drive their own curriculum.
- Core contributor to AgentDescent, a parallel, asynchronous self-evolution framework: prompts, skills, and code are versioned artifacts; candidate diffs are evaluated on held-out tasks and only evidence-backed changes are committed. Work spans conflict-aware merging, acceptance gates, RSI metrics, and auditable skill libraries.
- Treats self-evolution as distributed training: artifacts are parameters, diffs are gradients, and merging is the optimizer step; focuses on fusing conflicting edits from parallel workers and reproducible held-out evidence.
- Contributor to FlyBot, RSI-oriented agent software: role-scoped bots split engineering work down a tree, with delivery gated by risk-tiered CI, review, and a merge queue. It is the long-lived software counterpart to the AgentDescent RSI framework.
- Built a financial agent prototype on AntGLM and LangChain: ReAct-style reasoning, hybrid BM25/embedding retrieval, and multi-turn tool calling.
- Fine-tuned AntGLM with LoRA for anti-money-laundering (AML) anomaly detection.
- Designed the reconciliation algorithms for an AI-based automatic reconciliation system.
Research Experience
Designed an active-learning strategy for KGQA that fuses graph centrality, information density, and node uncertainty.
Investigated reflective Abstract State Machines (ASMs) for security.
Built blacklist/whitelist safety classification and knowledge-graph-based intent and emotion recognition for counseling dialogue.
Education
- M.S. in Computer Science, University of California San Diego (UCSD) — Sep 2023 – Dec 2024
- B.S. in Electrical Engineering (minor in Computer Science), University of Illinois at Urbana-Champaign (UIUC), dual-degree program with ZJU — Sep 2021 – May 2023
- B.Eng. in Electrical Engineering and Automation, Zhejiang University (ZJU) — Sep 2019 – Jul 2023
Honors & Awards
- 2012 Star Award, Huawei — 2026
- STDT Outstanding Project Delivery Award, Ascend Basic Software — 2026
- NSF REU Scholarship, National Science Foundation — 2022
- MCM/ICM Meritorious Winner (Top 7% of 10,053 teams) — 2021
Travel
Away from the terminal, I love exploring the world. The map and checklist are updated live from the console.
Click a country to focus it. Scroll-wheel zoom is off by default.
Gallery
Photos from my travels, uploaded and managed from the console.
Photos are being curated. Check back soon.