Member of Technical Staff (Research, Post-Training & RL)

In this role, you’ll lead our efforts to make AI reasoning and generation verified. Today’s models are incredibly fast but fuzzy in how they reason (think System 1), whereas formal methods are slow, deliberate and exact (System 2). Our hypothesis, and it is a bet, is that the real leap comes from truly interleaving the two. Nobody knows yet whether that’s right. Proofs are what make this tractable for us: a machine can check whether a proof holds, which gives us a far more precise signal to train against than most. You’ll build our post-training efforts from the ground up, figuring out how we train models to write specifications, proofs and correct code, and designing the rewards and environments that get them there. You won’t be doing this alone; you’ll work alongside a top-notch team of researchers and engineers in formal methods and AI. Honestly, we don’t know yet exactly what will work, and we might be wrong. We’re looking for someone who finds this thesis compelling and genuinely wants to find out whether it’s true, not someone out to prove us right.

Problems you’d own

• Design rewards from proofs. Turn verifiers, checkers and test results into reward signals that are exact and hard to game, and work out what to do between “proved” and “failed”.

• Build training environments. Create environments where models learn to specify, prove and fix real code, at a difficulty that keeps them learning.

• Generate the data. Build synthetic data and curricula from code, proofs and agent runs, and clean what goes in.

• Measure what matters. Build evals that track real gains in reasoning, not overfitting to benchmarks.

• Build the infrastructure, with the team. Work with a team of excellent engineers to build the training and data infrastructure our post-training efforts need, so experiments run fast and reliably.

• Work toward self-improving systems. Explore how models, verifiers and agents can improve one another over time.

You don’t need to have done all of this. We care about how you’d approach it.

You might be a fit if

You’ve done serious work in post-training, reinforcement learning or reasoning: training runs you designed, or papers and results that moved something forward. You form hypotheses, test them cheaply and update quickly. You’re drawn to problems without a known solution, and you’re honest about what the data shows.

You don’t need

• A background in formal methods. We’ll teach you, and you’ll teach us RL.

• Experience at a frontier lab.

Bonus points

• Verifiable rewards: you’ve worked on RL with verifiable rewards, theorem proving with language models, or code generation models.

• Infrastructure: you’ve built large-scale training infrastructure.

• Research: you’ve published research in these areas.

Who we hire

We hire at every level, and we mean it. If you’re senior, you’ll come in as a hands-on engineer and grow into leading and building a team. If you’re early in your career, fresh out of school or a few years in, you’ll get real ownership fast and as much scope as you can take on. What matters is the same bar for everyone: mastery, the drive to grow, and the instinct to help others grow too.

Compensation and logistics

• Benefits: unlimited PTO, health, dental and vision insurance, and a 401(k).

How we work

We work in roughly six-week cycles: five intense weeks, then a week for reflection, polish, feedback and mentorship. We bring opinionated ideas backed by evidence, bias for action, fix the class of problem rather than the instance, and close the loop with the people we serve. Read more in Our Culture.

Aretta comes from aretē, the Greek word for excellence and virtue: becoming the best version of yourself in everything you do. We value excellence for its own sake, and we believe it’s built across a whole life, not just the hours at a desk.