Researcher, Evaluations and Benchmarks

This is a fully remote position, open to applicants in New York.

📋 Description

• Deliver a benchmark every two to three weeks that assesses a previously unmeasured frontier risk.

• Collaborate with top AI laboratories and academic institutions on benchmarks and research publications.

• Work alongside in-house researchers responsible for specific harm areas and guide freelancers.

• Manage the taxonomy, evaluation harness, quality standards, and release process.

• Ensure that frontier labs can replicate the benchmark and validate the reported figures.

• Review evaluations, confirm taxonomy alignment, and encourage researchers to maintain high quality.

• Oversee the benchmark strategy and timeline, keeping researchers on track.

• Supervise two or three freelance subject matter experts as necessary.

• Meet approximately once a month with the CTO, pod, and research leads to establish the quarterly release roadmap.

• Align the release plan with target accounts.

• Monitor the AI safety and security landscape, review research, maintain relationships with labs, and collect feedback.

• Attend conferences and engage with AI lab contacts on a weekly basis.

⛳️ Requirements

• PhD or Master's degree in computer science, machine learning, or a related discipline, or equivalent experience from industry research.

• More than 3 years of experience in developing and executing safety or security evaluations for language models in production, within an AI lab, model provider, or a safety and security research organization.

• At least 5 pertinent research publications in AI safety and security, including being the lead author on a minimum of 2.

• Strong engineering capabilities, including experience with evaluation harnesses, distributed inference, vLLM, and the ability to read and troubleshoot a codebase.

• Skill in developing a taxonomy, not just scoring against one.

• Capability to direct a researcher and two freelancers without formal management authority.

• Proficient in English, both written and spoken.

• Genuine curiosity about harms and the ability to learn a new subject every three weeks.

• Experience with post-training methods such as SFT, DPO, or GRPO (preferably).

• Experience in designing rewards for subjective and safety-critical objectives (preferably).

• Background in agentic evaluation involving tool use, orchestration, permissions, and prompt injection (preferably).

• Publications presented at leading conferences (preferably).

• Willingness to present work during client calls; excellent verbal and written communication skills and the ability to present to large and/or senior audiences (preferably).

🏝️ Benefits

• Budget allocated for freelancers as needed on an ad hoc basis.

• Opportunity to travel to conferences a couple of times each year.

• Ideally, travel to conferences at least three times a year.