Research Scientist Intern, FM Quality ML (PhD)

Job description

Meta was built to help people connect and share, and over the last decade our tools have played a critical part in changing how people around the world communicate with one another. With over a billion people using the service and more than fifty offices around the globe, a career at Meta offers countless ways to make an impact in a fast growing organization. Meta is seeking Research Interns to join the FM Quality ML team within Facebook Ads Ecosystem. FM Quality sits at the intersection of machine learning and system optimization across Meta's family of apps. We answer three fundamental questions: how many ads to show, which ads are best for each person, and how to deliver them, balancing revenue growth with user experience. Our team applies state-of-the-art ML, Generative AI, Reinforcement Learning, and large-scale systems to personalize ad quality across Feed, Videos, Reels, and Stories. We own 0→1 initiatives end-to-end and collaborate broadly across Facebook and Instagram to build the foundation for sustainable business growth. Our interns will have the opportunity to frame practical, high-impact ads quality problems as rigorous machine learning problems, and bring their own expertise to build innovative solutions under the guidance of our research scientists and software engineers. The central research challenge is to build a quality system that understands each person and each ad deeply enough to reason about long-term user value, and to show people the ads they will like and trust. People tell us about their ad experience through many quality signals, such as survey ratings, hiding or reporting an ad, and their experience after a click. Each signal captures only part of the picture, and their effects add up across a session and over weeks. We are post-training large language models into a quality foundation model that reasons over a person's ad and organic interaction history together with the ad's content, and returns both an explanation grounded in that history and calibrated predictions for many quality signals at once. Open problems include post-training recipes and reward design that make reasoning measurably improve prediction and calibration, representing ads and user behavior in a shared semantic vocabulary the model can reason over, evaluating reasoning against people's own judgments, and distilling the model's reasoning into efficient production models that serve within strict latency and compute constraints. Offline evaluation and online experiments help us connect model improvements to better ad experiences and long-term value for people and advertisers. If you are excited about using LLM reasoning to help billions of people see ads they like and trust, come join us! Our internships are twelve (12) to sixteen (16), or twenty-four (24) weeks long and we have various start dates throughout the year.