Research Scientist, APEX Benchmarks

Mercor•San Francisco, CA

40d•$200,000 - $400,000•Onsite

About The Position

Mercor is seeking a Research Scientist to lead the design of the next generation of APEX benchmarks and the expert-built datasets behind them. APEX is the AI Productivity Index, a family of benchmarks measuring the economic value of professional work by frontier models across various fields like investment banking, law, consulting, medicine, accounting, and software engineering. These benchmarks are created and graded by practicing experts, with results published as papers, open datasets, and public leaderboards. This role is highly visible, sitting at the intersection of research, company strategy, and go-to-market. The scientist will determine future measurement areas based on frontier model failures, design benchmarks and scoring, collaborate with partners, and scale production with internal teams. The role also involves being a credible external technical voice. Mercor is looking for someone to build its research methodology, tooling, and publishing practices from the ground up.

Requirements

Research background in evaluation: Strong applied or academic research background in LLM evaluation, benchmarking, NLP or a related field, with a track record of rigorous experimental design.
Judgment about what to measure: You can look at a frontier model’s behavior and identify the measurement that would actually be informative, rather than the one that is easiest to build.
Statistical rigor: You reason carefully about sampling, variance, contamination and grader reliability, and you don’t ship a number you can’t defend.
Hands-on: Strong coding skills. You can build an eval harness, run the experiment and analyze the results yourself.
Exceptional communication: You can present complex technical findings clearly to frontier lab researchers and to non-technical audiences, in writing and in person.
Comfort with ambiguity: You’ve operated in fast-moving, cross-functional environments where the problem space is not yet defined.
Interest in the commercial side: Genuine curiosity about GTM strategy, startup dynamics and the business of AI data — this role sits close to all three.
In-person: You are excited to work from our San Francisco office five days a week in a high-intensity, high-ownership environment.

Nice To Haves

Ph.D. in machine learning, NLP or a related field; equivalent industry or frontier lab research experience considered.
Publications at top-tier venues (NeurIPS, ICML, ACL, ICLR), especially in evaluation, benchmarking or data-centric AI.
Experience authoring a widely-adopted public benchmark or dataset.
Industry experience on an evaluation, benchmarking or post-training team at a frontier lab.
Domain depth in one of APEX’s professional verticals — finance, law, consulting, accounting, medicine or software engineering.
Experience designing rubrics, model-as-judge pipelines, or human annotation programs at scale.

Responsibilities

Benchmark design: Decide what the next APEX benchmark should measure, based on frontier model performance, saturation of existing evals, and where economically valuable work is still out of reach. Own the task taxonomy, difficulty calibration, contamination controls and statistical design.
Dataset design: Design expert-built datasets and grading rubrics at scale — deciding what makes a task hard, what makes a grade defensible, and how to hold quality while thousands of experts produce work in parallel.
Measurement rigor: Set the standard for how we report results: confidence intervals, inter-rater agreement, human vs. model-as-judge calibration, held-out splits, and the failure analysis that explains why a model scored the way it did.
Partnerships: Work with academic collaborators and industry partners (as we did with Cognition on APEX-SWE) to co-design benchmarks and get them adopted.
External research voice: Publish — arXiv papers, open datasets, blog posts, conference talks, leaderboard releases — and represent Mercor’s research in conversations with frontier labs, customers and the press.
Translate results into narrative: Turn benchmark findings into clear arguments about the ROI of expert-curated data, for technical reports, customer conversations and go-to-market material.
Cross-functional work: Partner with data operations, engineering, product and strategy to take a benchmark from design to production, and to surface research findings that shape the company roadmap.
Stay at the frontier: Track the LLM evaluation literature and bring what’s good into how Mercor builds benchmarks.