Research Scientist - Decentralized Training and Inference Verification

Pluralis Research • San Francisco, CA

24d • Remote

About The Position

Pluralis Research is developing Protocol Learning: training and serving large models in a decentralized manner on 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 creating efficient algorithms and systems to verify work in a trustless training and inference network. The verifier must establish that inference tokens came from the claimed model and sampling parameters, and that training contributions are genuine and accurately represented.

The role owns the threat model across pre-training, post-training, and inference; develops and calibrates statistical verification tests; and builds and operates a verification service in the inference path. The work addresses challenges including training disruption, free-riding, poisoned updates and backdoors, private-data extraction, reward manipulation, and workers using the wrong model or sampling parameters. The position is remote-first, with teams distributed globally, and requires comfort working across time zones.

Requirements

• Experience building a calibrated statistical decision system with stated error rates and accountability for its mistakes; relevant publications in inference or training verification, or experience in fraud detection, anti-cheat, or experimentation platforms, are applicable.
• Deep expertise in statistics and probability, including designing experiments, calibrating decision thresholds, and defending claimed error rates.
• Knowledge of approaches to verifying untrusted compute, including statistical testing, re-execution, cryptographic proofs, and trusted hardware, and the ability to assess their suitability for a permissionless network.
• Belief that Protocol Learning is a viable path for collective, trustless, and sovereign AI.
• Professional-level written and spoken English proficiency.
• Comfort working across time zones.

Nice To Haves

• Familiarity with large-scale pre-training and RL post-training.
• Familiarity with decentralized ML security and adversarial threat models, such as poisoning, Sybil attacks, collusion, and replay.
• Experience at proprietary, open-weight, or open-source AI labs.

Responsibilities

• Enumerate and maintain the threat model for malicious or careless workers across pre-training, post-training, and inference, including disruption and denial-of-service, free-riding, model poisoning and backdoors, data extraction, and reputation or reward manipulation.
• Design statistical verification methods with stated error rates, benchmark and calibrate them rigorously, and control false positives and false negatives across heterogeneous hardware, including GPUs and Macs.
• Build and operate the verification service in the inference path, and take responsibility for its performance and errors.

Benefits

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

What This Job Offers

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