Research intern (PhD)

This line is co-supervised with a lab at HKUST and does not feed the product lines. The goal is papers, with no near-term business output — we do not expect revenue from it this quarter, and we will not push it for any.

Three directions; pick one.

The three directions

1. Evaluation benchmarks

Build our own evaluation standard for commerce operations, and run three baselines against it: the ceiling of general-purpose agents, vertical competitors, and human operators. The last is the hardest and the most valuable — to measure how well an agent makes operating decisions, you first have to know how well a person does.

Output: a benchmark paper.

2. Self-evolving agents

The direction is Self-Evolving Agents with Economic Ground Truth.

Most self-evolution research runs on synthetic benchmarks — an agent improves itself inside an environment it was given, and what improved is hard to state. Our reward is operating profit in real money, from a real business. The signal is dirtier, but it is real.

Output: a paper.

3. Model post-training

At 7B / 14B scale, post-train for tool-call sequences and skill scheduling, and measure the gain on our own evaluation benchmark.

Output: a paper.

What we look for

• A PhD candidate working on agentic systems
• The ability to carry a direction on your own. The direction you take is yours to move forward; nobody will decide your next step for you
• Fluency with tools like Claude Code. A requirement, not a bonus point
• Willingness to talk to the product side. Your data comes from a real business, and those people sit on another line