Research Engineer - Post-Training
Pluralis Research2 months ago
Sydney, AustraliaSenior
Responsibilities
- Build the end-to-end RL post-training stack, including geo-distributed rollout ingestion, reward computation, policy updates, and distributing updated weights across the network.
- Adapt RL algorithms to asynchronous, high-latency, partially trusted generation through staleness tolerance, off-policy corrections, and communication-efficient policy updates.
- Build evaluations to measure model improvement and ship the first decentralized post-trained model as a public artifact.
- Set technical direction and drive implementation of the post-training system and algorithms.
Requirements
- Hands-on experience running RL post-training on large language models using RLHF, RLVR, or reasoning-focused RL.
- Direct systems experience with rollout generation, asynchronous training loops, and weight synchronization.
- Production-quality Python and PyTorch engineering experience, including concurrency, failure handling, and profiling.
- Research ability demonstrated through publications or detailed, defensible unpublished work in RL post-training, asynchronous or distributed RL, or nearby fields.
- Professional-level written and spoken English proficiency.
- Ability to work comfortably across global time zones.
Benefits
- Equity-heavy package with significant ownership for key technical contributors in addition to a high base salary.
- Remote-first work environment with globally distributed team members.
- Optional full visa sponsorship and relocation support to Australia or the United States.
- Opportunity to work on largely unpublished problems in decentralized training and serving of frontier models.
- The role is remote across the world, with main teams in Australia and North America.
Categories
AI ResearchML Engineering
About Pluralis Research
Pluralis Research builds a protocol and tooling for collaborative, decentralized training and ownership of foundation models, enabling training and serving across heterogeneous, consumer-grade devices over the internet. The team demonstrated Agora, a permissionless run that pretrained an 8B model from scratch without any single participant holding the full weights. The privately held company is backed by Union Square Ventures and focuses on developer-facing infrastructure for large-scale model training and serving.