Research Scientist, Post-Training

True Talent Solutions Partners•San Francisco, CA

2h•Onsite

About The Position

Join a post-training research team at an applied AI data company focused on improving foundation models. The role involves designing controlled experiments to measure how training data influences model capabilities, generalization, and alignment, and translating findings into practical recommendations. The scientist will contribute to evaluation benchmarks, data types, and shared evaluation infrastructure, publish research and technical communications, and use evidence to improve data quality and production. The role also involves collaboration with external research partners to communicate data-driven improvements and the conditions behind them.

Requirements

Zero to four years of experience building or delivering work in LLM training, post-training, or model evaluation.
Prior experience at a frontier AI lab.
Strong familiarity with LLM training and evaluation methodologies, including SFT and RL-based post-training.
Experience independently designing, implementing, and analyzing controlled experiments, and extracting useful conclusions from imperfect results.
Software engineering skills to build or modify experimental infrastructure, such as evaluation harnesses, training pipelines, or benchmarks.
Ability to work across domains such as finance, software engineering, and policy, with a practical, builder-oriented approach.

Nice To Haves

Research experience in model evaluation, post-training, or data curation.

Responsibilities

Run controlled supervised fine-tuning and reinforcement learning experiments to measure the impact of datasets on model performance.
Help develop evaluation benchmarks and new data types that advance model training and assessment.
Build shared evaluation infrastructure to make findings reproducible across models, capabilities, and domains.
Publish external research, technical reports, and blog posts based on the work.
Use experimental results to help improve data quality and production.
Collaborate with external research partners to communicate clear evidence about data-driven improvements and the conditions behind them.