Vision & Goals
Win-Win Community in the Agent Era
True AGI is still a decade away. Agent is the essential path. Seize the wave to stay relevant.
Through deep practice, knowledge sharing, and project collaboration, break down technical barriers and transform them into personal advantages and real value.
Empower everyone, win the Agent decade.
What We Offer
- · Systematic learning paths and practice manuals
- · One-on-one guidance from technical mentors and industry experts
- · Paper notes, cutting-edge talks, case co-creation
- · Portfolio polishing, resume & interview coaching, referrals
Advanced Play
Deep Co-creation · Papers/Projects/Career Full Chain
If you want to get started with LLM Agents
- •Learning paths and open-source repo recommendations
- •Intro courses + hands-on projects
- •Like-minded exchange community
If you want further collaboration / papers / career
- •Paper collaboration and experiment co-building
- •Industry landing project cooperation
- •Big tech job referrals and interview coaching
If you seek partnerships
- •Co-build community brand
- •Joint promotion and events
- •AI product and training guidance
Learning & Exploration
Carefully curated learning resources to help you quickly master Agent core technologies
Advanced Path
Community Projects / Papers
Idea2Story
An Agent framework that automatically generates top-tier conference-level paper narratives from ideas. Trained on tens of thousands of top conference papers and their review data, teaching AI to master "Scientific Storytelling".
Talks & Roundtables · Paper Reading
就从来没人质疑过 OPD 的 reward 设计吗？我们发现了 log 本身就是问题，也许就不该用！
📄 arXiv：https://arxiv.org/abs/2606.17199 作者机构：东方理工、香港理工大学、上海交通大学、University of Waterloo` On-Policy Distillation（OPD）正在成为大模型后训练的标准组件。和传统 SFT 只在 teach
深入理解 Agentic RL 中的行为崩塌现象
作者：Jason 原文：https://zhuanlan.zhihu.com/p/2050794185909876253 过去一段时间里，我一直在思考一个问题：Agentic RL 到底和传统 RLVR 有什么本质区别？ 最容易想到的答案当然是工具调用。传统 RLVR 往往围绕单轮回答展开，模型生成
为什么 Pretrain Loss相同，但下游任务表现却天差地别？
作者：虚无 原文：https://zhuanlan.zhihu.com/p/2048769202819949265 TL, DR LLM Pretraining是模型能力的主要来源。模型的能力一般会随着pretraining loss的降低而单调提升。 那么，pretraining loss相同，模
深度对话！2025 "青稞" AI 嘉年华，与 20+ 位青年科学家一起探讨AI 技术瞬间
本次活动专为青年科学家打造，旨在搭建一场 AI 技术的深度对话，来自学术和工业界的 20+ 青年科学家，将与大家一起回顾 2025，展望 2026！
TRPO重生：大模型时代的信任域策略优化
在大型语言模型的强化学习阶段，特别是RLHF中，我们追求策略的持续优化。本次分享深入探讨TRPO在LLM时代的应用。
从 π_0 到 π_RL：面向流匹配 VLA 的强化学习后训练框架
深入解析流匹配VLA的强化学习后训练框架π_RL，探索具身智能的前沿技术。
RLinf：面向具身智能的"渲训推一体化"开源强化训练框架
开源强化训练框架RLinf，实现渲染、训练、推理一体化，加速具身智能研发。
RLinf-VLA 实践：从零上手 VLA（OpenVLA）强化学习
手把手教你使用RLinf-VLA框架进行OpenVLA强化学习实践，入门具身智能开发。
深度对话！2025 "青稞" AI 嘉年华，与 20+ 位青年科学家一起探讨AI 技术瞬间
本次活动专为青年科学家打造，旨在搭建一场 AI 技术的深度对话，来自学术和工业界的 20+ 青年科学家，将与大家一起回顾 2025，展望 2026！
TRPO重生：大模型时代的信任域策略优化
在大型语言模型的强化学习阶段，特别是RLHF中，我们追求策略的持续优化。本次分享深入探讨TRPO在LLM时代的应用。
从 π_0 到 π_RL：面向流匹配 VLA 的强化学习后训练框架
深入解析流匹配VLA的强化学习后训练框架π_RL，探索具身智能的前沿技术。
RLinf：面向具身智能的"渲训推一体化"开源强化训练框架
开源强化训练框架RLinf，实现渲染、训练、推理一体化，加速具身智能研发。
RLinf-VLA 实践：从零上手 VLA（OpenVLA）强化学习
手把手教你使用RLinf-VLA框架进行OpenVLA强化学习实践，入门具身智能开发。
Join Us
Foundation Mastery
LLM/Multimodal fundamentals, code skills enhancement and engineering standards
Agent Architecture
Planning/Memory/Tool calling and evaluation, real business case breakdown
Project Co-creation
Hands-on project teaming, mentor Q&A and code review
Career Leap
Portfolio polishing, interview workshops, mentor recommendations and referrals
Deep Practice + Mentor Q&A + Project Co-creation
Portfolio polishing, code review, weekly retrospectives, referral recommendations. Limited spots per cohort to ensure interaction quality.
Partnership & Consultation
WeChat Official Account: AgentAlpha
Co-build community / Promotion partnerships / AI products / Training guidance, or need paper, project, career support, scan QR code to follow official account for more info.
