Post-Training Researcher
You will design reward functions and then try to break them. This is the work the labs say they are short of: writing a grader, attacking it with a model that is actively trying to cheat, and confirming afterwards that it measured what you intended.
- up to $220 / hr
- Remote
- Contract · flexible hours
- Engineering & ML
- Feeds SimReal-MLBench
What we are looking for
Frontier lab or equivalent
- Post-training experience — RLHF, RLVR, or reward-model work at a lab or serious research group
- Adversarial mindset — You reach for the shortcut before you reach for the solution
- Systems judgment — Understanding of how a weak reward propagates through a training run
- Rigour — Comfort with variance, pass^k, and contamination checks
Work you might do
- Designing invariant-based reward functions for multi-step tasks
- Running degenerate policies against graders and documenting the bypasses
- Reviewing reward specifications for exploitable slack
- Formalising what a robust verifier looks like in a new domain
How it works
Apply now, or join the list.
- 01
Apply or join the waitlist
Apply for this role in a minute, or leave your email and field below.
- 02
Get invited when your world opens
We invite experts in waves, as environments in each field go live.
- 03
Define what correct means
Paid, remote contract work: set the rules agents are graded against and review their runs.
Not ready to apply? Join the expert waitlist and we will write when a world in your field opens.