Milkyway
Forecasting agentI contributed to Milkyway, which previously ranked first on the public FutureX leaderboard. My work focuses on search infrastructure, forecasting workflows, and learning from evidence gathered before outcomes resolve.
Chuyang Wei
Ph.D. Student in Computer Science and Technology
I’m a Ph.D. student at the University of Science and Technology of China, jointly trained at Beijing Zhongguancun Academy.
I study self-improving agents: learning from delayed feedback, acquiring reusable skills, and transferring experience from agent harnesses into model weights. My work uses open-world forecasting as a testbed.
Stay foolish.
Current focus
RSILoop Transformer
October 2026Research interests
- Recursive self-improvement
- Improving agent harnesses through experience and internalizing reusable strategies through on-policy training.
- Learning with delayed feedback
- Using temporal evidence before outcomes resolve for credit assignment, self-correction, and calibration.
- Loop Transformer
- Recurrent computation and iterative refinement for reasoning.
Selected projects
All repositories ↗FutureWorld Offline
BenchmarkVersioned forecasting questions with resolved outcomes, temporal boundaries, and reproducible scoring.
Futurecast
Agent infrastructureTools, prompts, and reusable skills for forecasting with existing agent runtimes.
Awesome Harness Self-Evolving
Reading listA research map of persistent changes to agent instructions, memory, skills, workflows, and executable harnesses.
Selected papers
Google Scholar ↗Agents, forecasting & self-improvement
Preprint
Harnessing Pre-Resolution Signals for Future Prediction Agents
Milkyway evolves a persistent forecasting harness using temporal evidence before outcomes resolve.
arXiv ↗PDF ↗Preprint
FutureWorld: A Live Reinforcement Learning Environment for Predictive Agents with Real-World Outcome Rewards
A live RL environment that learns from delayed real-world outcomes through stored prediction trajectories.
arXiv ↗PDF ↗EMNLP
DEEPRUBRIC: Evidence-Tree Rubric Supervision for Efficient Reinforcement Learning of Deep Research Agents
Evidence trees align queries and evaluation rubrics for efficient reinforcement learning of deep research agents.
arXiv ↗PDF ↗Preprint
ZGCM-1: A Fully Open and Extremely Efficient Foundation Model for Math and Agentic Search
An open 7B foundation model and training recipe for mathematical reasoning and agentic search.
arXiv ↗PDF ↗Time series & representation learning
AAAI
ICML
IEEE TGRS
Education
University of Science and Technology of China
Ph.D. in Computer Science and Technology
Joint training
with Beijing Zhongguancun Academy
University of Science and Technology of China
Master’s studies in Big Data Technology and Engineering
Lanzhou University
B.Eng. in Artificial Intelligence
Contact
I welcome conversations about self-improving AI, recurrent reasoning, and forecasting.
weichy2023@mail.ustc.edu.cn