Research background in post-training LLMs with applicable work examples
Product sense to discern valuable techniques for users
History of delivering research usable by others
Self-motivated to take projects from ideation to realization
Willingness to work on-site in NYC or San Francisco
Responsibilities
Own post-training research projects such as async RL and on-policy distillation.
Collaborate with Forward Deployed Engineers to enhance model training based on real-world customer feedback.
Foster partnerships with external research labs for innovative projects like DFlash.
Work with engineering teams to implement new post-training techniques into products.
Help define and influence the ongoing research agenda.
Benefits
Collaborative environment with cross-functional teams
Exposure to cutting-edge research and real-world applications
Opportunity to shape the company's research direction
Access to resources like GPU sandboxes for experimentation
Engagement with leading external research labs
Full Job Description
The Role:
We're building a platform that covers the whole life of an LLM: training it, deploying it, and observing it in production. We already run multi-node training, elastic inference, sandboxes, and distributed volumes, and we control the infrastructure underneath. We're looking for research depth in post-training to sit alongside our systems and product work.
You will do hands-on post-training research at Modal, working with the research lead to pick high-impact bets and owning them end to end. The work that pays off fastest is tied to production workloads -- we're already experts at training speculators for deployed models, and there are open research questions like distilling a target model from its own production traffic. There is also room to prove what the platform makes possible, where training AI scientists or kernel engineers is a natural fit given our GPU sandboxes.
What you'll do:
Own end-to-end post-training research bets: async and agentic RL, on-policy distillation, long-context RL, small routing models, and whatever else the research agenda calls for.
Work directly with customers alongside our Forward Deployed Engineers to train models and bring what you learn back into the research.
Carry and expand collaborations with outside research labs. For example, our work with ZLab on DFlash, a speculator design built on KV injection and blockwise parallel drafting.
Work with engineering to turn frontier post-training techniques into products: an opinionated post-training framework, distributed-training approaches (DiLoCo, evolutionary strategies), online training for deployed models, and more.
Help shape the research agenda. None of the above is prescriptive; your work will help guide our future.
Requirements:
A research-leaning background in post-training LLMs, with work you can point to.
Enough product sense to tell which frontier techniques matter to users and which stay academic.
A record of shipping research that other people build on, whether in a lab or in industry.
The drive to take a research bet from idea to result without much hand-holding, working in the open with the rest of the team.
Ability to work in-person, in our NYC or San Francisco office.
