About
A year ago the post-training stack was SFT on GUI tasks. Today it runs SFT plus online and offline RL across GUI, API, MCP and coding domains, on tens of thousands of tasks. Getting RL to run is no longer the problem. Getting it to work is: sparse rewards, long-horizon credit assignment, and gains that survive outside the benchmark.
This is a well-funded frontier AI company building computer-use agents and the models behind them. The Models team owns the training pipeline from mid-training and SFT through offline and online RL and reward modelling. Dedicated MLOps, infrastructure and inference teams sit alongside.
Several seats are open, mid to senior, leaning senior. The team is hybrid, three days a week in the office, with a few days together in the other office every couple of months. Releases land on public leaderboards regularly.
What you'll do
- Train and improve LLMs and VLMs for agent use: instruction following, tool use and long-horizon computer use
- Design and run online and offline RL setups, including reward models, for multi-turn agent tasks
- Build the data pipelines and distributed training systems behind them, on PyTorch, veRL and Prime-RL
- Run ablations that separate real capability gains from benchmark noise
- Turn agent failures from research and product into new training signal and reward changes
- Move models from experiment into regular releases alongside the agent and inference teams
What you'll need
- Hands-on experience training LLMs or VLMs with SFT, DPO, RLHF/RLVR, reward modelling or offline RL
- Practical RL experience, online or offline. Exceptional SFT-only backgrounds are considered case by case
- Strong Python and PyTorch, with software engineering fundamentals
- Experience with large-scale distributed training (FSDP or similar)
- Ability to work across the stack, from training backends and data ingestion to RL algorithm design
- Low ego and comfort in a fast-moving environment with many team dependencies
Optional Bonus
- Publications at top-tier venues, or a PhD or MSc in ML, NLP or CV
- Experience training models for computer use or other multi-turn, multimodal agent settings
- RL with sparse rewards, curriculum learning, data mixture design or distillation
Shortlisted candidates will be contacted within 48 hours.
