Humanfia

Appearance

Team ​

Humanize 2 is part of NVIDIA RSI effort

Long-horizon work is done by people, and a result reported without saying whose it is has left out the part that is accountable for it.

Proposed Humanize 1

Sihao Liu @SihaoLiu

NVIDIA Research

Wrote Humanize 1 at UCLA PolyArch, out of his own GAAC project, and with it the rule the rest of this site is built on: the agent that wrote the code is not the agent that says whether it is any good. Now a Research Scientist in the Architecture Research Group at NVIDIA Research, designing chips with agent systems like Humanize 2.

sihaoliu.github.io

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 half a dozen other places — and pointed the loop at problems somebody else keeps the scoreboard for. He works on every project.

lzhu.me

Humanize 2: Agent Flow System

The runtime: ten coding-agent CLIs, one flow, and the whole run written down.

  • 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 work the runtime had to survive.

  • Changye Li

    @antoinegg1Tsinghua University

    Flows, the flowverse, and running the thing at Kaggle scale.

  • Dong Zhou

    @dongz9Carnegie Mellon University

    Execution and the container path.

  • Zheng Du

    @crmsnduGeorgia Tech

    Long-horizon runs, and what breaks on them.

Humanize 1: RLCR Loop

Humanize 1, where it all started: Claude implements, Codex reviews, the criteria end the run.

  • Sihao Liu Lead

    @SihaoLiuNVIDIA Research

    Lead. Proposed it, wrote it, and still maintains the plugin.

  • Shinan Liu

    @shinan6University of Chicago

    The plugin, and the monitor you watch a loop through.

  • Zijian Zhang

    @futrimeTsinghua · NVIDIA Research

    The port: the three commands, as flows.

FlowBench

The referee. Scores flows against each other on work that takes a day, not a minute.

  • Zijian Zhang Lead

    @futrimeTsinghua · NVIDIA Research

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

  • Zheng Du

    @crmsnduGeorgia Tech

    Tasks, drawn from work that actually had to be finished.

  • Junxian Guo

    @JerryGJXMIT · NVIDIA Research

    Kernel and systems tasks.

HOA

Humanize Olympic 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 what a Lean proof is worth.

  • Jui-Hui Chung

    @unixtomatoPrinceton University

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

KDA

Kernel Design Agents. Faster on real hardware, or it does not count.

  • 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.

HKA

Humanize Kaggle Agent. Real competitions, real leaderboards, and an audit that refuses to flatter itself.

  • Changye Li Co-lead

    @antoinegg1Tsinghua University

    Co-lead. The workflows, the entries and the audit.

  • Zijian Zhang Co-lead

    @futrimeTsinghua · NVIDIA Research

    Co-lead. The leaderboard, and the runtime underneath the runs.

  • Menghan Li

    @zgdlltTsinghua University

    Submissions, across most of the nineteen.

  • Yitong Liu

    @apostle715Tsinghua University

    Submissions and reproduction.

Credit ​

Everybody named has shipped something you can go and read. Each post on the blog names the people who ran it.

Humanize 1 was proposed at UCLA PolyArch, and everything since has been built with many more people than are named here — NVIDIA Research, MIT HAN Lab, UCLA, Tsinghua, and a long tail of community contributors. The full list is the one git keeps: Humanize 2 · Humanize 1 · the flowverse.

Working with us ​

  • A question, a bug, or a flow that beats ours: open an issue or a pull request on the repository it belongs to — Humanize 2, Humanize 1, or the flowverse.
  • Working on long-horizon agent systems: we would like to compare notes. Open a discussion on Humanize 2.
  • Everything we ship: github.com/humanfia. Apache-2.0.