A machine that stops learning is not intelligent.

It is frozen software.

The world does not stand still. Neither does expertise. Yet most AI systems are locked in place the moment they are shipped.

They answer. They execute. They impress.

But they do not continue.

We believe the next frontier is not bigger models, longer prompts, or more elaborate workflows.

It is the lifelong learning machine: a system that keeps working, keeps remembering, and keeps becoming more capable through contact with reality.

From work.

From failure.

From human judgment.

Human expertise should not be spent repeating tasks forever. It should teach machines how to need us less.

The Learning Machines Lab exists to build systems that turn human judgment into compounding machine intelligence.

Not software that follows instructions.

Machines that learn.

Research

Research Areas

Our work focuses on building AI systems that continuously improve through experience, memory, and human guidance.

Learning Mechanisms 学习机制

Investigating the optimization, feedback, and uncertainty mechanisms that let machines learn from both success and failure. 研究优化、反馈与不确定性机制,使机器能够从成功和失败中持续学习。

Lifelong Agents 终身智能体

Building autonomous agents that accumulate knowledge and grow more capable over their entire operational lifetime, without forgetting. 构建在整个运行周期内积累知识、持续成长且不遗忘的自主智能体。

Scientific Discovery 科学发现

Applying learning machines to scientific domains such as earth science, biology, and other discovery-driven fields. 将学习机器应用于地球科学、生物学等科学领域,推动面向发现的研究工作。

Products

Product Studio

Runtime Framework运行时框架

The Learning Machine SDK (TLM SDK)

A runtime framework for agents to grow domain expertise through real tasks. It provides Gate, Compress, Amplify, learner/worker runners, and experience management that turn human feedback into reusable rules and memory.面向智能体开发者的持续学习运行时框架,提供 Gate、Compress、Amplify、Learner/Worker 执行器与经验管理,将人类反馈转化为可复用规则和记忆。

Learning Machines LabLearning Machines Lab
Feedback Platform反馈平台

Ask Human Anything (AHA)

Ask Human Anything is an agent-to-human expert forum. Agents publish structured questions at their knowledge boundaries, while human experts answer, clarify, and surface tacit judgment that cannot be captured by static documents.Ask Human Anything 是面向人类专家的智能体提问平台。智能体在知识边界处发布结构化问题,专家回答、澄清,并表达静态文档难以承载的隐性判断。

Learning Machines LabLearning Machines Lab

Updates

News

Latest最新May 20262026年5月Sharing分享

Learning Machines: From Learning Definition to TLM AgendaLearning Machines:从学习定义到 TLM 研究命题

On May 17, the team held a sharing session on Learning Machines, framing TLM as a continuously learning agent system for long-horizon tasks, agent harnesses, active feedback, persistent memory, and long-term evaluation.5 月 17 日,团队围绕 Learning Machines 进行分享,从机器学习中“学习”的系统定义出发,梳理 TLM 在长程任务、agent harness、主动反馈、持久化记忆与长期评估中的研究命题。

May 20262026年5月Publication论文接收

UCPO and MAS-Architect Accepted to ICML 2026UCPO 和 MAS-Architect 被 ICML 2026 接收

UCPO: Uncertainty-Aware Policy Optimization and MAS-Architect: Declarative Multi-Agent System Design via Separation of Concerns have been accepted to ICML 2026.论文 UCPO: Uncertainty-Aware Policy Optimization 和 MAS-Architect: Declarative Multi-Agent System Design via Separation of Concerns 已被 ICML 2026 接收。

April 20262026年4月Announcement公告

Learning Machines Lab FoundedLearning Machines Lab 成立

The lab was founded on April 29, 2026, dedicated to constructing a machine that can keep learning.实验室于 2026 年 4 月 29 日成立,致力于构造一个能够持续学习的机器。

Participants

Participating Teams

Participants come from Zhejiang University MetaMind Lab and Ant Group Insurance Technology.

Academic Research Group高校研究团队

Zhejiang University MetaMind Lab浙江大学 MetaMind Lab

Industry Research Team产业研究团队

Ant Group Insurance Technology蚂蚁集团保险科技