Inactive
P

Preference Model

Reinforcement learning environments for frontier AI

Member of Technical Staff - Research Engineer, Post-training

AI ResearchPosted on Sep 12
Full-timeSan Francisco, CA, USAIn-person

About the Job

Requirements

  • Experience running end-to-end post-training pipelines for large language models at least 7 billion parameters in size.
  • Proficiency in Python and PyTorch or JAX.
  • Experience with at least one modern reinforcement learning training framework.
  • Experience building and operating machine learning infrastructure at scale.

Responsibilities

  • Train and evaluate models on proprietary reinforcement learning environments to validate data quality, surface task-coverage gaps, and connect environment design to model capability.
  • Architect and optimize reinforcement learning training infrastructure, including training abstractions and distributed experiment management, using frameworks such as Verl, OpenRLHF, or similar; scale systems for increasingly complex research workflows.
  • Design, implement, and test training environments, evaluations, and methodologies for reinforcement learning agents.
  • Profile and optimize training runs end to end, from data loading through reward computation, to maximize experiment throughput and shorten research iteration cycles.

Desired Qualifications

  • Experience evaluating model outputs and building reward or evaluation signals.
  • Stay current on post-training research and translate papers into running code.
  • Strong opinions, held loosely, about structuring reinforcement learning training code for reproducibility and fast iteration.
  • Ability to balance research exploration with engineering rigor.
  • Strong systems design and communication skills.

About the Company

Preference Model

Reinforcement learning environments for frontier AI

Preference Model builds reinforcement learning environments and harnesses to train frontier AI models, focusing on how reward signals shape model behavior. It designs robust RL harnesses and simulators that treat reward functions and graders as an attack surface, testing them across millions of training rollouts to harden against loopholes and gaming, with fail-safes for safety. Unlike others that only build models or tools, it targets the reward design and alignment bottlenecks, drawing talent from Anthropic, Stripe, Google DeepMind, and others, and backed by notable investors. Its goal is to broaden access to capable AI while preserving strong safeguards to keep frontier systems safe.

Company size
11-50
Company stage
Seed
Total funding
$16M
Headquarters
San Francisco, CA
Founded
2025

Recently Posted Jobs

Explore more opportunities at Preference Model

See more jobs