Dixing Xu

Dex Hunter, also known as Dixing Xu, is a London-based Member of Technical Staff at Weco AI and a co-author of AIDE and AIDE². His work on automated research spans agents that search for better machine learning programs, model-compression improvements in OpenAI’s Parameter Golf challenge, and systems that experiment on the research agent’s own code.

From portfolio models to distributed training

Dex’s collaboration with Weco co-founder Zhengyao Jiang dates to their 2017 reinforcement-learning framework for portfolio management, co-authored with Jinjun Liang. The system learned how to redistribute a portfolio while accounting for previous allocations and transaction costs. The team evaluated it through cryptocurrency-market backtests; the reported results were not live-trading performance.

His subsequent work combined research with infrastructure engineering. After a 2018 internship, Dex became a maintainer of the Hyperledger Fabric Python SDK and later mentored contributors. Research appointments at Nanyang Technological University and Zhejiang University involved privacy-preserving machine learning; at Hong Kong Baptist University, he worked on distributed training. His career also includes backend engineering in Go and Rust at Hex Trust. He joined Weco as an AI research consultant in 2024 and became a Member of Technical Staff in May 2026.

Two contributions explain the technical range of that earlier work:

• ObfNet: transforming inputs before they leave a device. Dex and Mengyao Zheng contributed equally to ObfNet, a collaborative project that placed a small neural network on an edge device to disguise data before sending it to a remote classifier. Case studies involving spoken digits, handwritten digits and sign-language images tested whether transformed inputs could conceal their recognizable form while preserving classification accuracy.

• Custom communication in distributed training. Dex’s Horovod tutorial implemented a custom hierarchical allreduce operation—the communication step that combines results across GPUs during training. It addressed how to change the training framework’s communication behavior, extending his work from models to the infrastructure that runs them.

Searching programs, then improving the searcher

The jointly authored AIDE research treats machine learning engineering as a search through executable programs. An agent drafts code, runs it, repairs failures and branches from promising solutions. Its experiment tree preserves alternatives, allowing the agent to return to useful earlier work rather than make every subsequent change to a single evolving script.

Parameter Golf put automated experimentation under concrete constraints: a 16 MB model-artifact limit and a ten-minute training budget. Dex built a research agent that produced record-setting entries in the challenge. One submission highlighted by OpenAI extended Raahil Shah’s earlier quantization work with full-Hessian GPTQ, which accounts for interactions among weights when compressing a model. A later accepted entry changed quantization clipping to preserve useful outlier information in the model’s MLP weights. These were cumulative improvements in a shared competition, building on other entrants’ techniques.

Dex then co-authored AIDE², which makes the research agent itself the object of experimentation. An outer agent proposes changes to the inner agent’s code, evaluates the resulting agents and retains improvements. This extends the search from candidate machine learning programs to the machinery that chooses, runs and remembers experiments.

In an eight-day autonomous run, the AIDE² team reported seven successive improvements, including changes to search and memory management. The strongest discovered agent matched or exceeded a human-engineered production agent on four held-out benchmarks. The result demonstrates improvement within that experimental system and budget; it does not establish unlimited self-improvement.