Research Scientist - Post-Training

Pluralis Research • San Francisco, CA

11d • Remote

About The Position

Pluralis Research develops Protocol Learning: training and serving large models in a fully decentralized way across small consumer-grade devices connected over the internet. Its Agora project pretrained an 8B model from scratch on consumer GPUs distributed across the internet, without any participant holding the full weights. This role focuses on making that pretrained model useful for agentic use cases through decentralized reinforcement-learning post-training. Unlike conventional post-training stacks that rely on datacenters, synchronous rollouts, fast interconnects, and trusted workers, the work must accommodate consumer GPUs and Macs, geographically distributed inference, high-latency rollouts, and weights that are never held by a single participant.

The Research Scientist will lead the end-to-end algorithms and systems for RL post-training, including rollout ingestion, reward computation, policy updates, and distribution of updated weights. The role also involves adapting RL methods to asynchronous, high-latency, partially trusted generation, developing evaluations, and taking the first decentralized post-trained model from a training run to a public release. The team works remotely across Australia and North America, and this role requires comfort working across time zones and professional-level written and spoken English.

Requirements

• Hands-on experience running RL post-training on large language models, such as RLHF, RLVR, or reasoning-focused RL.
• Experience working directly on the systems layer, including rollout generation, asynchronous training loops, or weight synchronization; experience limited to launching jobs on another team's stack is not sufficient.
• Production-quality Python and PyTorch engineering, including concurrency, failure handling, and profiling before optimizing.
• Professional-level written and spoken English proficiency.
• Comfort working across time zones.

Nice To Haves

• Publications in RL post-training, asynchronous or distributed RL, or nearby fields; detailed, defensible unpublished work is also a strong signal.
• Experience training over slow networks or with decentralized or federated setups.
• Familiarity with serving-engine internals such as vLLM or SGLang.
• Experience with reward modeling or building verifiable-reward datasets.
• Experience with peer-to-peer networking and NAT traversal.
• Experience at proprietary, open-weight, and open-source AI labs.

Responsibilities

• Build the end-to-end RL post-training loop, including rollout ingestion from the geo-distributed inference pipeline, reward computation, policy updates, and distributing updated weights to the network.
• Set direction and deliver the post-training stack.
• Adapt RL algorithms to asynchronous, high-latency, partially trusted generation, including staleness tolerance, off-policy corrections, and communication-efficient policy updates.
• Build evaluations to demonstrate model improvement.
• Deliver the first decentralized post-trained model from training run to public artifact.

Benefits

• Significant equity ownership for key technical contributors, in addition to a high base salary.
• Flexible work environment with team members distributed globally.
• Optional full visa sponsorship and relocation support to Australia or the United States.

What This Job Offers

• Job Type: Full-time
• Career Level: Senior
• Education Level: No Education Listed