Research Scientist

Anyone AI Inc.; United States; Remote; USD 110,000 - 160,000; Full time

Job summary

Anyone AI Labs in a remote capacity across LatAm/US seeks a Research Scientist focused on LLM evaluations and benchmarking. You will design frontier-grade evaluation methods and build benchmarks across reasoning, coding, agents, and multi-modal capabilities.

You will collaborate with labs, defend methodologies under scrutiny, and contribute to public benchmarks and papers, targeting venues like NeurIPS Datasets & Benchmarks, ICLR, and ACL.

Qualifications

• Research background in ML evaluation or benchmarking.
• Expertise in LLM/frontier-model benchmarking.
• Understanding measurement constructs and psychometrics.
• Interest in safety and robustness of evaluation.
• Ability to lead a team and maintain high standards.
• Experience framing, running studies, and writing results.

Responsibilities

• Evaluation research: convert targets into benchmark designs.
• Benchmark development with expert-verified ground truth.
• Recruit and calibrate a pool of experts across coding, agents, and reasoning.
• Serve as the technical contact for labs with CEO support.
• Turn lab requests into pilots and publish benchmarks.

Skills

ML evaluation

LLM benchmarking

Measurement science

Safety in evaluation

Team leadership

Study design

Job description

Research Scientist, LLM Evaluations & Benchmarking

Anyone AI Labs
Reports to: CEO Remote / LatAm / US

The role Evaluation is one of the hardest open problems in AI: we still don't have reliable ways to measure what frontier models can and can't do, and the field mostly runs on benchmarks that are saturated, contaminated, or measuring the wrong thing. You'll own that problem at Anyone AI, measuring frontier model capability.

This is a research role at heart: you decide what a good evaluation is , design the benchmarks that prove it, and defend the methodology under lab scrutiny. You'll build frontier-grade evaluation packages across reasoning, coding, agents, tool use, and multi-modal — grounded in expert-verified truth, validated against multiple models, and QC'd to survive buyer-side review.

Responsibilities

• Evaluation research. Turn eval targets into original benchmark designs. Own the hard measurement questions: construct validity, item discrimination, headroom, reliability, contamination, and capability elicitation. Push toward evals that stay informative as models improve.
• Benchmark development. Build evaluation packages with subject-matter experts, each with expert-verified ground truth, multi-model headroom results, and rigorous QC (calibration layers, severity-weighted rubrics, deterministic verifiers).
• Experts. Recruit, calibrate, and review a pool across coding, agentic/tool-use, and STEM/reasoning. Be the final arbiter of correctness and frontier difficulty.
• Lab relationships. Be a technical point of contact for labs, with CEO support. Understand what they're trying to measure and translate it into an evaluation design.
• Delivery & dissemination. Turn lab requests into winning sample packages and own pilots end to end. Where the work generalizes, help turn it into public benchmarks and papers: we support publishing at venues like NeurIPS Datasets & Benchmarks, ICLR, and ACL.

What we're looking for

• Research background in ML evaluation or benchmarking (a track record of published or open benchmarks, eval/measurement research, or equivalent hands‑on work that labs have relied on).
• Deep LLM/frontier-model benchmarking expertise, with real strength in code-model and agentic evaluation.
• Fluency with the measurement problem itself: construct validity, psychometrics, rubrics, pass rates, headroom, contamination, and what makes a task genuinely discriminate a model.
• Interest in the safety side of evaluation, capability elicitation, robustness, and measuring the things that are hardest to measure honestly.
• Proven ability to hold a team or expert pool to a rigorous standard.
• Comfort with the full research loop: framing the question, running the study, and writing it up.