We use RL on computer-use tasks to train models for long-horizon work. The interface is streaming visual observation and low-level actions. Rewards come from task completion or rubric-based generative reward models.
You will design training procedures, build evals for failure modes, run experiments, and publish the results.
Training infrastructure.
Environment selection.
Algorithms and capability retention.
Evals and stability at long horizons.
#J-18808-Ljbffr
