Research Scientist - Post-Training
Pluralis Research · Remote · San Francisco · Australia ·
COMPENSATION
€•• – ••k
Listed on the employer's job board (Ashby)
- Fully remote (global)
- RL + Distributed Systems
- Visa sponsorship available
- Senior research role
- Python + PyTorch
You will develop the end-to-end RL training loop for Protocol Learning, enabling post-training of large models across distributed consumer devices. The work requires inventing algorithms that handle staleness and off-policy corrections in untrusted, geo-distributed networks.
The team has already pretrained an 8B model via Agora and now focuses on making it useful for agentic use cases. You will define the technical direction, build evaluation frameworks, and release the first decentralized post-trained models.
What you'll do
- Build the end-to-end RL training loop including rollout ingestion and policy updates
- Adapt standard RL recipes for asynchronous, high-latency, and partially trusted hardware
- Develop communication-efficient policy updates and staleness tolerance mechanisms
- Create evaluation frameworks to demonstrate model improvement over time
- Manage the release of the first decentralized post-trained models as public artifacts
What you'll need
- Hands-on experience running RL post-training on large language models
- Deep understanding of systems layers like rollout generation and weight synchronization
- Production-quality engineering skills in Python and PyTorch
- Ability to handle concurrency, failure modes, and performance profiling
- Professional-level written and spoken English proficiency
- Research background in RL post-training or distributed RL fields
Nice to have
- Experience training models over slow networks or in federated setups
- Familiarity with serving engine internals such as vLLM or SGLang
- Background in reward modeling or building verifiable-reward datasets
- Knowledge of P2P networking and NAT traversal techniques
- Prior experience at proprietary or open-source AI laboratories
Stack
- Python
- PyTorch
- vLLM
- SGLang
How you'll work
- Fully remote with team members distributed globally
- Main teams located in Australia and North America
- Requires comfort working across multiple time zones
- Flexible work environment
What you get
- Equity-heavy compensation package with significant ownership
- High base salary for key technical contributors
- Optional full visa sponsorship for Australia or the US
- Relocation support provided if moving to sponsored locations