Jobs at Pluralis Research

Develop end-to-end reinforcement learning post-training for large language models operating across decentralized, high-latency consumer hardware. Responsibilities include building rollout ingestion, reward computation, policy updates, weight synchronization, asynchronous and off-policy algorithms, evaluation systems, and public model releases. The role requires strong Python and PyTorch engineering, hands-on RLHF/RLVR experience, and research ability in distributed or asynchronous reinforcement learning.

Develop efficient algorithms and systems for verifying untrusted decentralized training and inference. Responsibilities include modeling threats such as poisoning, free-riding, data extraction, and reward manipulation; designing statistically calibrated tests with controlled error rates across heterogeneous hardware; and deploying verification services in production inference systems.

Build and optimize decentralized distributed pretraining systems for frontier-scale models across heterogeneous consumer GPUs and unreliable internet connections. Responsibilities include model, data, and pipeline parallelism; communication optimization; elasticity and fault tolerance; checkpointing and recovery; and monitoring throughput, bottlenecks, and model quality. The role requires hands-on PyTorch distributed training experience, strong production Python engineering, and demonstrated execution through shipped systems, research code, open-source work, or substantial projects.

Build and own a geo-distributed inference stack for reinforcement learning rollouts and future model serving. Responsibilities include pipeline-parallel execution, routing, transport, serving-engine development, failure handling, and designing algorithms for consumer hardware connected over unreliable public networks. The role requires hands-on experience shipping inference systems, research ability in distributed inference or adjacent fields, and expertise with low-bandwidth, high-latency networking.

Conduct novel research in Protocol Learning, focusing on decentralized large-model training and inference across consumer devices. Develop foundational methods addressing communication efficiency, fault tolerance, heterogeneous compute, and malicious participants, with the goal of publishing in tier-1 machine learning venues. Interns own a research problem, leverage significant compute, and receive mentorship from senior scientists during a six-month PhD internship.

Build an end-to-end decentralized RL post-training stack for large language models, including rollout ingestion, reward computation, policy updates, weight synchronization, and evaluation. Develop algorithms for asynchronous, high-latency, partially trusted environments and ship post-trained models as public artifacts. The role requires hands-on RLHF, RLVR, or reasoning-focused RL experience, strong Python and PyTorch engineering, and research capability. Experience with slow networks, decentralized systems, model serving, reward modeling, or P2P networking is advantageous.

Conduct foundational research on Protocol Learning, focusing on communication-efficient decentralized training, convergence under churn and staleness, heterogeneous systems, and malicious-participant robustness. Publish findings in top-tier conferences and collaborate with engineering teams to deploy methods in live frontier-model training runs. The role requires machine learning research excellence, distributed training experience, strong PyTorch skills, and alignment with decentralized AI.

Build and own a geo-distributed inference stack for decentralized reinforcement learning and future model serving. Responsibilities include pipeline-parallel execution, node placement and routing, network transport, serving-engine internals, algorithm development, and failure handling across consumer hardware and unreliable public-internet connections. The role requires hands-on experience shipping large-scale inference systems, research capability in distributed inference or related fields, and expertise in low-bandwidth, high-latency networking.

Build and optimize decentralized distributed pretraining systems for frontier-scale models on heterogeneous consumer hardware. Responsibilities include model-parallel training, communication efficiency, elasticity, fault tolerance, checkpointing, state synchronization, recovery, and monitoring across unreliable internet-connected devices. The role requires hands-on PyTorch distributed training experience, strong production Python skills, concurrency and profiling expertise, and evidence of shipping research or engineering systems.

Develop threat models, statistical verification algorithms, calibrated tests, and production verification services for decentralized training and inference on untrusted heterogeneous hardware. The role focuses on detecting incorrect model outputs, malicious training behavior, free-riding, poisoning, data extraction, and reward manipulation while controlling false positives and negatives. Candidates need deep statistics and probability expertise, experience shipping or publishing calibrated decision systems, and familiarity with statistical testing, re-execution, cryptographic proofs, and trusted hardware.