Jinyu Xiang
CoFounder at Amplift · AI Agent Researcher · Musician
I am the CoFounder at Amplift, where I spearhead the algorithm team in advancing the core capabilities of General Agents. Previously, I was a Researcher at MetaGPT (DeepWisdom), and initiated OpenManus with the team. My research focuses on LLM-based Agents, particularly Agent Self-Evolution and Foundation Agents.
I am also the Research Lead at Agent Universe, leading team research on AI content reformulation and redistribution.
Besides my work in AI, I am an Independent Musician on Tencent Music, creating instrumental music, Traditional Chinese style (Gufeng), and Electronic music, with over 2 million plays and 1k+ followers.
🔥 News
- 2025.05 CoFounded Amplift, focusing on General Agents and Marketing Agent applications.
- 2025.04 Participated in the Foundation Agent Survey V1, exploring cross-environment agents.
- 2025.03 Proposed and built OpenManus with the team, achieving 50K+ Stars!
- 2025.02 Self-Supervised Prompt Optimization was accepted by EMNLP 2025.
- 2024.10 AFLOW was accepted by ICLR 2025 as an Oral Presentation (Top 1.8%).
- 2024.08 Joined DeepWisdom as a Researcher, focusing on Agent Self-Optimization and Foundation Agent.
- 2024.06 MathAI won second place in the 2024 Alibaba Global Mathematics Competition AI Challenge (Top 1%, 2/563).
- 2024.02 Started leading research at Agent Universe on AI content reformulation.
📝 Selected Publications
ICLR 2025 Oral AFlow: Automating Agentic Workflow Generation
AFlow introduces an automated framework that reformulates workflow optimization as a search problem over code-represented workflows, using Monte Carlo Tree Search to efficiently explore and refine workflows through code modification and execution feedback.
EMNLP 2025 Self-Supervised Prompt Optimization
We propose a self-supervised framework for prompt optimization that enables language models to automatically refine prompts through iterative feedback and learning.
50K+ Stars OpenManus: An Open-Source Framework for Building General AI Agents
OpenManus is a minimalist Agent framework that articulates the minimal structural units for Agent design. We completed this project in just three hours, which garnered 35k stars within its first week of release.
Arxiv 2025 Advances And Challenges In Foundation Agents
A comprehensive survey on Foundation Agents with universal understanding, cognition, and action capabilities that can operate in any environments and collaborate to form collective intelligence.
ReCode: Unify Plan and Action for Universal Granularity Control
ReCode proposes a unified framework that integrates plan and action for universal granularity control in code generation tasks.
A Survey of Self-Evolving Agents: On Path to Artificial Super Intelligence
A comprehensive survey on self-evolving agents, exploring the path toward artificial super intelligence through autonomous agent evolution and adaptation.
MedAgentsBench: Benchmarking Thinking Models and Agent Frameworks for Complex Medical Reasoning
MedAgentsBench introduces a benchmark for evaluating thinking models and agent frameworks on complex medical reasoning tasks.
AOrchestra: Automating Sub-Agent Creation for Agentic Orchestration
AOrchestra presents an approach for automating the creation of sub-agents for agentic orchestration, enabling more flexible and scalable multi-agent systems.
AutoEnv: Automated Environments for Measuring Cross-Environment Agent Learning
AutoEnv proposes an automated framework for creating environments to measure and evaluate cross-environment agent learning capabilities.
💻 Job Experience
- 2025.06 - Present CoFounder, Amplift, China.
- 2024.08 - 2025.05 Researcher, DeepWisdom, China.
- 2024.02 - Present Research Lead, Agent Universe, China.
- 2019.02 - Present Independent Musician, Tencent Music, China.
🎯 Competitions and Side Projects
- 2025.03 OpenManus, Github 50K+ stars, Founder. Proposed and assembled a team to recreate Manus functionality, completing the project in just three hours, which garnered 35k stars within its first week.
- 2024.07 Zhihu AI Social Experiment, Author. Conducted a social experiment by creating a fully autonomous AI account on Zhihu that posted content for a month without human intervention or detection.
- 2024.06 2024 Alibaba Global Mathematics Competition AI Challenge, Top 1% (2/563), Author. Developed the MathAI multi-agent system to tackle Olympiad level math problems, reaching the finals and placing second.
- 2024.05 AI Research Weekly, Author. Led a team in developing an automated system that crawls Hugging Face for new AI research papers weekly, utilizing LLMs to automatically categorize papers by domain and generate comprehensive summaries.
🔈 Invited Talks
- 2025.10 Weibo AI Deep Talk #27: Is AI Entrepreneurship a Traffic Password?
- 2025.09 INCLUSION·Conference on the Bund Roundtable: Industry Observations from a Gen-Z Agent Application Entrepreneur
- 2025.07 Tencent: The Second Half of AI - Defining Problems is More Important Than Solving Them
- 2025.04 Tencent: The Explosive Rise of Agents - Unlimited Application Space?
📚 Recent Blog Posts
- DEC 31, 2025
End-of-2025 Reflections: Thoughts on Agents
An end-of-year reflection centered on Agent Learning/Evolving, methodologies for AI products, industry competition and long-term moats, and how people rewire “rewards” amid uncertainty.