Careers at PhAI Labs
Run the experiment, train the model
PhAI Labs does model training and real research at the same time. We are looking for researchers and engineers who want to bring AI into the process of discovery: not to push a benchmark score, but to have a model tested against real questions, real data and real experiments.
Why PhAI
The problems here are real. The model you train goes into a scientist’s research; the environment you build runs a real scientific codebase; the evaluation you design gets checked against an experimental result. You see the whole chain in one place, from scientific definition and model training to real execution and continuous improvement, rather than one segment of it.
Research and engineering directions
- 01
Foundation models
Training and post-training foundation models for scientific tasks: reasoning, long-horizon planning, multimodal understanding.
- 02
Reinforcement learning
Post-training where the reward is a real outcome: turning experimental results, evaluation feedback and expert judgement into training signals.
- 03
Research agents
Recursive research agents and the Science IDE: exploring, executing, evaluating and improving long-running research.
- 04
Scientific tools and environments
Packaging real scientific codebases into runnable, measurable, reproducible environments, with defined tasks and checkable success criteria.
- 05
AI for life science and biomedicine
Omics and single-cell analysis, disease mechanisms and biomarkers, molecules and proteins, experiment design and testable hypotheses.
- 06
Systems and infrastructure
Training and inference systems, data pipelines, and the scheduling and reproduction of experimental environments.
Who fits
- Genuinely curious about a scientific question, not only about the model
- Able to narrow a vague problem into a plan that can be executed and verified
- Comfortable reading a paper and a lab notebook, writing training code and debugging an environment
- Honest about results: a failed experiment and a failed hypothesis are both worth recording
- Used to working with people from other backgrounds: scientists, engineers, clinicians
How we work
We organize work around real research problems, not around functions. A question comes in from a scientist, passes through problem formulation, model training, tool use and experimental verification, and goes back to the scientist; every step in between has an owner and can be checked. We write more than we say. What can be verified matters more than what can be demoed. We collaborate with scientists in universities, institutes and laboratories over the long term, and team members take part in those collaborations directly.
Open positions1
-
Research Intern · Post-training and self-evolution
For master's and doctoral students, and strong undergraduates: post-training and self-evolution research on LLMs and agents. Interns who do well have the chance to publish as first author or core contributor. CNY 1,000-5,000 a day, negotiable for exceptional candidates.
What you will do
- RL post-training: reinforcement learning for reasoning and agent tasks, including GRPO and RLHF
- Reward and verifier modelling: reward models, verifiable rewards, self-correction
- Agentic post-training: skill, memory and tool-use architecture for agents
- Self-evolving harness: training and evaluation systems that keep improving
What we look for
- Master's or doctoral student in CS, AI, mathematics or physics; strong undergraduates welcome
- Research or project experience in reinforcement learning, large models or agents
- Genuine research drive and the ability to work through problems independently
- At least four days a week on site, for at least three consecutive months
- A plus: top-venue papers, substantial open-source work, or familiarity with verl / Megatron / DeepSpeed
How to apply
Beyond the openings listed above, if one of the directions is something you want to work on, write to us directly: say what problem you want to work on and the most relevant work you have done, and attach a CV or profile page, a code repository or papers.
New openings are listed here as they come up. If your direction fits but no position matches, write to us anyway.