Algorithm Engineer II @ Huawei 2012 Laboratory
Agent systems · RLHF post-training · Recursive self-improvement
HUGOXK LAB · Academic Homepage · LinkedIn · Email
I build systems where agents improve themselves. At Huawei 2012 Laboratory I work across the stack: large-scale RLHF post-training on Ascend NPU clusters, multi-agent scheduling for scientific discovery, and recursive self-improvement (RSI) loops where agents evolve their own skill libraries under strict train/validation/test discipline.
- RLHF at cluster scale — core modules of GRPO and DAPO, long chain-of-thought sequence packing, and stability validation on Ascend NPUs.
- Agent systems — a pluggable DAG-based multi-agent scheduler with configurable node types and dependencies for scientific discovery.
- Recursive self-improvement — rollout, failure-driven skill induction, and top-K frontier selection over versioned, auditable skill libraries; exploring island populations, MAP-Elites archives, and tree search.
- Self-generated evaluation — letting agents drive their own curriculum.
hugoxk.online is my personal research lab: one open-source framework, three online services, and a shared console. The site publishes project pages, release notes, and research results.
Current open-source work is published here and in the upstream repositories. The RSI framework and agent-software work below is part of my work at Huawei 2012 Laboratory:
| Project | What it is | Role | Links |
|---|---|---|---|
| AgentDescent | Parallel, asynchronous framework for self-evolving agents where diffs are gradients and the aggregator is the optimizer | Core contributor | GitHub · Docs · Paper |
| FlyBot | A local-first team of long-lived AI Bots that run in your browser and delegate work down a tree | Contributor | GitHub · Service |
| ScienceDiscovery | An AI research workspace with a fail-closed sandbox and end-to-end provenance for code, environment, logs, and citations | Contributor | GitHub |
| Bot Social | Owner-governed social network for AI assistants via MCP, with signed messages and an auditable approval chain | Contributor | GitHub · Service |
The lab's research page collects the AgentDescent paper, held-out results, and reproduction entry points. I'm especially interested in collaborations on agent infrastructure, self-evolving agents, and evaluation methods for open-ended improvement.
- M.S. in Computer Science, UC San Diego — Machine Learning concentration, 2024
- B.S. in Electrical Engineering (minor in Computer Science), UIUC, dual-degree program with Zhejiang University, 2023
- B.Eng. in Electrical Engineering and Automation, Zhejiang University, 2023
- Previously: Algorithm Engineer Intern at Ant Group (LLM applications for anti-money laundering) and AI Engineer Intern at State Street (AI-based reconciliation)
Selected honors: Huawei 2012 Star Award (2026), STDT Outstanding Project Delivery Award (2026), NSF REU Scholarship (2022), MCM/ICM Meritorious Winner (2021).
- Active learning for knowledge-graph question answering — advised by Prof. Hanghang Tong, UIUC (NSF REU)
- Exploration of Reflective ASMs for Security — advised by Prof. Klaus-Dieter Schewe, Zhejiang University
- AI-assisted psychological counseling chatbot — advised by Dr. Zhenzhong Lan, Westlake University (Xinhua coverage)
- hugoxk.online — HUGOXK LAB: projects, online services, research, and the shared console
- hugoxk.github.io — academic homepage with bio, experience, and travel map
- FlyBot Docs — documentation site for FlyBot
Older repositories on this account are coursework archives from Zhejiang University, UIUC, and UC San Diego, kept for reference. Current open-source work is in the repositories listed above.