AI Agent Research

Building the Future of
Autonomous Intelligence

CodeAgent · DeepResearch · Agentic RL · Self-Evolving Systems

34+
Publications
60+
Researchers
April 2025
Founded
Affiliated with PKU · THU · CMU · Stanford · HKUST · ZJU

Research Focus

Agent Intelligence

CodeAgent

End-to-end execution for real-world coding tasks, enabling agents to autonomously understand, modify, and run complex repositories.

DeepResearch

Multi-step reasoning agents for open-ended scientific and knowledge discovery, synthesizing insights across domains.

Agentic RL

Reasoning-centered reinforcement learning, bridging decision-making, planning, and adaptive intelligence.

Self-Evolving

Agents that continuously improve through self-evolution and collaborative learning, enabling emergent collective intelligence.

AI4Science

AI4Investment

Agent-driven automated investment: discovering optimal factors, strategies, and execution models for intelligent trading and portfolio management.

AI4Healthcare

Efficiency-first AI for healthcare: scalable foundation models, vision-language alignment, and agentic workflows for real-world clinical impact.

AI4Bioinformatics

Agent-driven discovery in genomics: automating bioinformatics analyses and enabling novel scientific insights in biological research.

Selected Work

Papers, Products, and Coverage

Switch between the strongest publications, commercialized systems, and media coverage without leaving the homepage.

QuantaAlpha framework diagram comparing ML, agent baseline, and QuantaAlpha trajectories
AI4Investment · Reported by QbitAI mirror

QuantaAlpha Alpha Mining Engine

Evolutionary LLM-driven alpha mining (QuantaAlpha): each mining run is a trajectory refined by trajectory-level mutation and crossover—revising weak segments and recombining high-reward paths—with semantic consistency across hypotheses, executable factor code, and evaluation under noisy, non-stationary markets.

View product
EpochX production network ecosystem illustration
Agent Economy · Live Product

EpochX Production Network

Credits-native marketplace infrastructure for human–agent networks: participants post and claim tasks, route work through decomposition and delivery with verification, and persist reusable ecosystem assets (skills, workflows, traces, distilled experience), while Credits lock bounties, settle rewards, and pay creators when assets are reused.

View product
RepoMaster repository search, analysis, and autonomous exploration workflow diagram
RepoMaster · Sema Code

RepoMaster/Code Agents

RepoMaster (NeurIPS 2025 Spotlight) autonomously searches, analyzes, and explores large GitHub repositories—using hierarchical code trees, module-dependency and function-call views, and pruned context to complete complex tasks with far lower token use. Sema Code decouples the agent engine as embeddable infrastructure: the same core powers VS Code and messaging-style clients with multi-tenant isolation, scheduling, and layered permissions.

View product
Idea2Paper research narrative product landing screenshot
Research Automation · Public Experience

Idea2Paper / Auto Research

Research-assistant stack that turns rough ideas into structured narratives and paper drafts: iterate across literature, planning, experiments, writing, and revision—surfaced through ARK at idea2paper.org and the PaperBuild experience at paperbuild.cn, with open reference implementation on GitHub.

View product