Research Intern

Mercor•San Francisco, CA

2d•$80 - $80•Onsite

About The Position

As a Research Scientist Intern at Mercor, you’ll work on research at the frontier of post-training, reinforcement learning with verifiable rewards (RLVR), data generation, and model evaluation. You’ll investigate how datasets, rewards, and training methods affect the capabilities and behavior of large language models. This may include designing controlled experiments, developing new evaluation methodologies, conducting systematic failure analysis, and testing approaches to improve tool use, agentic behavior, and real-world reasoning. You’ll work closely with research scientists, research engineers, and domain experts to turn open-ended questions into rigorous experiments. Your work will contribute to Mercor’s research agenda and may support external publications, benchmark releases, and the development of frontier AI systems.

Requirements

Currently pursuing a master’s or PhD in computer science, machine learning, statistics, mathematics, or another relevant field.
Demonstrated ability to formulate research questions, design experiments, and draw sound conclusions from empirical results.
Demonstrated experience in at least one of the following: Training, fine-tuning, or evaluating language models; Agentic AI system, RL environments; Developing benchmarks, evaluation methodologies, or data-quality measures.
At least one publication or open source project.
Strong programming skills, particularly in Python, and the ability to write reliable research code.
Familiarity with machine learning fundamentals, experimental design, and statistical analysis.
Intellectual curiosity.
Comfort operating in a fast-paced research environment with rapid iteration and a high degree of ownership.

Nice To Haves

Previous research experience in language models, reinforcement learning, model evaluation, or post-training.
Experience training, fine-tuning, or evaluating language models.
Familiarity with RLVR techniques, reward modeling, or agentic AI systems.
Experience developing benchmarks, evaluation methodologies, or data-quality measures.
Research publications or submissions at competitive CS conferences such as ACL, NeurIPS, ICML, ICLR, or EMNLP.
Research papers, technical reports, open-source projects, or other work samples demonstrating relevant skills.

Responsibilities

Develop and investigate research questions related to post-training, RLVR, data quality, and model evaluation.
Design and run controlled experiments to understand how datasets, rewards, and training strategies affect model performance.
Study reward-shaping and post-training methods, including approaches such as GRPO and DAPO.
Develop methods for measuring data quality, usability, and performance uplift on key benchmarks.
Design and evaluate datasets, rubrics, evaluators, and scoring frameworks for complex model capabilities.
Conduct systematic error analysis to identify model failure modes and opportunities for improvement.
Analyze experimental results and communicate findings through clear reports, research artifacts, and presentations.
Build the research tooling and data pipelines needed to conduct experiments at scale.
Collaborate with research scientists, research engineers, applied AI teams, and domain experts producing training and evaluation data.
Contribute to research publications, benchmark releases, and other public research outputs where appropriate.