Skip to content
View Yaxin9Luo's full-sized avatar
34 followers · 60 following

Achievements

Achievement: Pair ExtraordinaireAchievement: Pull Sharkx2Achievement: Starstruckx2Achievement: QuickdrawAchievement: YOLO

Achievements

Achievement: Pair ExtraordinaireAchievement: Pull Sharkx2Achievement: Starstruckx2Achievement: QuickdrawAchievement: YOLO

Highlights

  • Pro

Organizations

@MetaAgentX

Block or report Yaxin9Luo

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Yaxin9Luo/README.md
Yaxin Luo. ML PhD student at MBZUAI VILA Lab, research intern at Meituan LongCat. Post-training agents for long-horizon expert work: the model and the harness around it, improved together.

Homepage Google Scholar CV Email

Now · Research intern at Meituan M17 LongCat. Open to research collaboration.

ML PhD student at MBZUAI (VILA Lab, advised by Prof. Zhiqiang Shen). I work on one question: how do we get agents to do long, expert work, currently focus on artifact design tasks, and how to keep agents getting better at it? My answer is to post-train the model and co-evolve its harness together.

后训练面向长程agentic任务（目前focus on artifact design）的智能体：模型与它周围的 harness，一起进化。

Research map. 01 Multimodal (2024-25): APL, γ-MoD, DViN. 02 Eval (2025-26): Open CaptchaWorld, Next-Gen CAPTCHAs, FigMirror. 03 Post-training (2026 onward): AutoDesign, followed by LongCat-2.5, where the model and its harness improve together.

Featured work

03 · Post-training

  • AutoDesign ★ 246 · arXiv 2026 · first author
    Meta-harness optimization for long-horizon agentic design: editable posters, slides, webpages and videos.
  • LongCat-2.5 · Meituan M17 · 1.6T-parameter model · post-training contributor
    Large-scale post-training for agentic design: harness design, trajectory collection, Mid-Training, SFT and RFT.

02 · Eval

  • Open CaptchaWorld ★ 94 · NeurIPS 2025 · first author
    Interactive web benchmark for multimodal LLM agents: 20 CAPTCHA types, 225 tasks.
  • Next-Gen CAPTCHAs ★ 27 · ICML 2026 · co-first author
    Scalable GUI-agent defense built on the cognitive gap between humans and agents.
  • FigMirror ★ 530 · open source · contributor
    Reference-driven plotting agents that turn your data into editable Matplotlib in a paper's figure style.

01 · Multimodal

  • γ-MoD ★ 46 · ICLR 2025 · first author
    Mixture-of-depth adaptation that skips redundant layers in multimodal LLMs.
All publications (8)

Abstracts, figures and co-authors on the homepage.

To My Website

Pinned Loading

  1. AutoDesign AutoDesign Public

    The open-source design agent and harness, better than Claude Design on academic communication artifacts production. This DesignHarness can also be used with any coding harness you like ( Codex/Clau…

    Python 248 11

  2. VILA-Lab/FigMirror VILA-Lab/FigMirror Public

    An Automated AI Agent Tool for Plotting Your Data in Any Paper's Figure Style.

    Python 532 39

  3. MetaAgentX/OpenCaptchaWorld MetaAgentX/OpenCaptchaWorld Public

    [NeurIPS 2025] The first web-based benchmark and platform to evaluate visual reasoning and interaction capabilities of MLLM powered agents through diverse and dynamic CAPTCHA puzzles.

    JavaScript 94 4

  4. Gamma-MOD Gamma-MOD Public

    [ICLR2025] γ -MOD: Mixture-of-Depth Adaptation for Multimodal Large Language Models

    Python 46 3

  5. MetaAgentX/NextGen-CAPTCHAs MetaAgentX/NextGen-CAPTCHAs Public

    [ICML 2026]A defense framework against MLLM-based web GUI agents. This repository provides both the generative CAPTCHA system and tools for evaluating agent resistance.

    Python 29

  6. spending-effort-with-jev spending-effort-with-jev Public

    Jev-powered /effort advisor for Claude Code: tells you when to switch effort, per prompt

    Python 7