Research Scientist Intern

Meta; Menlo Park (CA); On-site; USD 55,000 - 83,000; Full time

Job summary

Meta in Menlo Park, CA is seeking Research Interns to join our AI Search research team. You will work on advancing AI by improving how agents learn from experience, enhance search and reasoning workflows, and transfer gains to new tasks.

The internship spans 12 to 24 weeks with multiple start dates throughout the year. Interns will explore three directions (learning to improve workflows, persistent memory and continual learning, and learning from imperfect feedback) with mentor guidance, aiming

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 our AI Search research team. We are committed to advancing the field of artificial intelligence by making fundamental advances in how AI agents learn from their own experience — improving their search and reasoning workflows, retaining useful knowledge, and transferring those gains to unfamiliar tasks. A central research challenge is making repeated self-improvement reliable when feedback is sparse or noisy.

The intern will work in one of three directions, scoped closer to the start date together with the intern manager:

• Learning to improve research workflows — agents learn better search, tool-use and verification strategies from prior attempts by improving their harness: the workflow, memory, tools and feedback loop around the model. Topics include meta-learning across tasks, automated agent design, and learning from natural-language feedback.

• Persistent memory and continual learning — agents retain useful lessons and update knowledge across tasks, whether in context, in compressed memory, or in weights, without forgetting what they already know. Topics include long-context modeling, context compression, continual learning, and test-time training.

• Learning from imperfect feedback — turning noisy judgments and failed attempts into useful learning signals, and improving the strategy an agent uses to learn, including how it selects and evaluates its next improvement. Topics include reward learning, credit assignment, and learning to learn.

Across all three directions, the key test is whether improvements transfer to new tasks. The goal is to develop novel methods and experiments toward a first-author submission at a leading ML/NLP venue. Recent internships on this team have produced work such as REFRAG, a new architecture for retrieval-augmented generation.

Our internships are twelve (12) to twenty-four (24) weeks long and we have various start dates throughout the year.