Skip to content

About ​

Humanize is part of NVIDIA RSI effort

We build the flow around the agents: the flows, the runtime that runs them, the benchmark that compares them, and the projects that test them on hard problems. These are the people who do it.

Founding member

Sihao Liu @SihaoLiu

NVIDIA Research

Holds Humanize to the CLIs it actually drives: he aligned the drivers for Claude Code, Kimi Code, Qwen Code, pi, opencode, grok, Cursor and ZCode with what each release really takes, and the sign-in ways behind them. A Research Scientist in the Architecture Research Group at NVIDIA Research, designing chips with agent systems like Humanize.

Built the team

Ligeng Zhu @Lyken17

NVIDIA Research

Assembled Humanfia. Everyone here is here because he went and found them — across NVIDIA, MIT, UCLA, Tsinghua and a dozen other places — and pointed our flows at problems with an outside grader. He scales both the team and agents, and keeps improving the efficiency & productivity of token maxing at Nvidia Research. He works on every project.

Humanize

The runtime that drives thirteen coding-agent CLIs through a flow and records every turn.

  • Zijian Zhang Lead

    @futrimeTsinghua · NVIDIA Research

    Lead. The runtime itself — the anchor, the tracing, the backends and the CLI.

  • Dongyun Zou

    @DongyunZouTsinghua University

    Backends, and the kernel workloads the runtime has to run.

  • Changye Li

    @antoinegg1Tsinghua University

    Flows, the flowverse, and running Humanize across Kaggle competitions.

  • Sihao Liu

    @SihaoLiuNVIDIA Research

    Harness alignment: every driver held to what its CLI really does.

  • Dong Zhou

    @dongz9Carnegie Mellon University

    Execution and the container path.

  • Zheng Du

    @crmsnduGeorgia Tech

    Long-horizon runs and their failure modes.

KDA

Kernel Design Agents. Agents that research, write and profile GPU kernels, judged by speed on real hardware.

  • Dongyun Zou Lead

    @DongyunZouTsinghua University

    Lead. The workflow, and the MLSys 2026 FlashInfer contest entries.

  • Yixin Dong

    @ubospicaCarnegie Mellon University

    Established the baseline and evaluations, and continuously improves KDA generalization.

  • Junxian Guo

    @JerryGJXMIT · NVIDIA Research

    Working on the self-evolving KernelWiki.

  • Song Bian

    @WaterpineNVIDIA Research

    Working on the self-evolving KernelWiki.

  • Xiaoyu Zhang

    @BBufRadixArk

    KDA-Pilot, and the forty-plus operators merged into SGLang.

  • Yahui Cui

    @smoothsmoothNVIDIA

    Bootstrapped CuteDSL-related primitives into KDA.

HOA

Humanfia Olympiad Agents. Mathematics, physics and quantum information, checked by Lean 4.

  • Zhengyang Zhang Lead

    @ZhengyangZhang06Tsinghua University

    Lead. PutnamBench, IMO 2026 and both Lean-Eval runs.

  • Jing Xiong

    @menik1126

    Physics and quantum information, formalized end to end.

  • Hongzhou Lin

    @hongzhoulin89Amazon AGI

    Research-level proofs, and how they are checked in Lean.

  • Jui-Hui Chung

    @unixtomatoPrinceton University

    Set up the PutnamBench baseline and proposed recursive lemma proving for HOA.

  • Mark Kim

    @mark_sfIndependent Researcher

    Discuss the math autoformalization and help setup the 1stproof contest.

FlowBench

Scores flows against each other on day-long tasks; the winner becomes the default.

  • Zijian Zhang Lead

    @futrimeTsinghua · NVIDIA Research

    Lead. The harness, the task format and the scoring.

  • Zheng Du

    @crmsnduGeorgia Tech

    Tasks drawn from real engineering work.

  • Junxian Guo

    @JerryGJXMIT · NVIDIA Research

    Kernel and systems tasks.

HMA

Humanize MLE Agents. Two native coding agents take turns over one workspace, and medal more often than either alone.

  • Changye Li Co-lead

    @antoinegg1Tsinghua University

    Co-lead. The method, the paper experiments, and the Kaggle entries and audit before them.

  • Zijian Zhang Co-lead

    @futrimeTsinghua · NVIDIA Research

    Co-lead. The runtime the agents run on, and the Kaggle leaderboard.

  • Menghan Li

    @zgdlltTsinghua University

    Kaggle submissions, across most of the competitions HMA grew out of.

  • Yitong Liu

    @apostle715Tsinghua University

    Kaggle submissions and reproduction.

We think the method around the agent outlasts the model ​

A better model writes a better function; it does not, on its own, work for eleven hours, hold on to a constraint from round one, and stop when the job is done. That is the loop around the agent, and we think the loop is what lasts. Models are replaced every few months; a method can be written down, measured and improved, and it gets better with each new model. So we open-source our flows, benchmark them against other flows, and drop ours when they lose.

How we work ​

  1. The person decides what done means. People own the intent, the tradeoffs and the definition of done. Agents do the work within those decisions.
  2. A separate agent reviews the work. Every serious flow has a second agent that arrives with no memory of the work and reads the repository instead.
  3. Reward hacking is the normal failure. Long runs fail by looking finished — a narrowed test, a special case, a stub under a confident summary. Our reviewers hunt for exactly that.
  4. We only claim what we have measured. FlowBench tests our flows against others, and we remove the ones that lose.
  5. Open source. The runtime, the flows and the reproduction code are open source. Every published number comes with what is needed to check it.

Credit ​

Each result is credited at the top of the post that reports it, not folded into a team byline. Everything has been built with many more people than are named above — NVIDIA Research, MIT HAN Lab, UCLA PolyArch, Tsinghua, Nunchux, and a long tail of community contributors. The full list is in each repository's contributor graph: Humanize · the flowverse · this site.

Get in touch ​