Zhengyao Jiang is the co-founder and chief executive of Weco AI, where he builds autonomous research agents that write machine-learning code, run experiments, evaluate results, and improve their own research methods. His central proposition is that scientific discovery can be accelerated by machines executing systematic experimentation while people define the questions, constraints, and standards that make the results worthwhile.
Jiang’s earlier research explored how software agents learn to make decisions under practical constraints. His 2017 work on financial portfolio management produced PGPortfolio, an open-source implementation of deep reinforcement learning for investment allocation. In 2019, he and Shan Luo introduced Neural Logic Reinforcement Learning, representing policies with first-order logic to improve interpretability and generalization.
He completed doctoral research in reinforcement learning at University College London under Tim Rocktäschel and Edward Grefenstette, concentrating on interaction-efficient reinforcement learning. Subsequent projects included planning in compact latent action spaces, published at ICLR 2023, and work on generalist planning for humanoid control.
With co-founder Yuxiang Wu, Jiang applied those ideas to automated machine-learning research: agents that generate hypotheses, modify code, execute experiments, and retain improvements. Their open-source system AIDE approaches machine-learning engineering as a search through candidate programs, using execution results to debug failures and guide further exploration. OpenAI’s MLE-bench evaluation reported its strongest tested configuration combined an OpenAI model with AIDE scaffolding. In July 2025, Jiang and Wu announced Weco’s $8 million seed round.
- Evaluation-driven research. Jiang treats evaluation as the objective function for autonomous experimentation: poorly chosen metrics reward misleading solutions. Software architecture also shapes what agents can discover; separating training and test data through stricter interfaces, for example, reduces opportunities for data leakage.
- Research that other engineers can use. Weco’s experimental agent Aiden ran approximately 1,300 experiments over 22 days in OpenAI’s Parameter Golf challenge, producing seven accepted leaderboard records. Jiang considers community reuse a more meaningful test than benchmark gains alone. One successful submission combined published gated-attention research, quantization to satisfy a 16-megabyte limit, and another participant’s tokenizer improvement—illustrating how agents can recombine human ideas and resolve implementation obstacles.
- Bounded recursive self-improvement. Weco’s AIDE² uses one research agent to improve another agent’s research harness. Jiang reported an eight-day experiment in which the resulting system surpassed a hand-tuned baseline on held-out benchmarks. The claim concerns measurable improvements to a specific research workflow, not unlimited self-accelerating intelligence.
For Jiang, stronger research automation raises the value of human experimental design: researchers still choose the problems, define trustworthy evaluations, and build the abstractions that determine where autonomous systems search.
