Research Scientist Intern, AI Core Machine Learning (PhD)

Meta•Sunnyvale, CA

11h•$7,650 - $12,667

About The Position

Meta is seeking a Research Scientist Intern to join Monetization Ranking & AI Foundations and investigate whether AI research agents can improve through learning from their own experiments. The project explores recursive self-improvement for autoresearch: agents form hypotheses, run experiments, analyze results, and use accumulated experience to improve future research. The team develops an agentic machine-learning research system that proposes, trains, and evaluates model changes at production scale, accumulating research trajectories and outcomes for the system to learn from. The intern will help shape research on which parts of an agent should learn from experience, how to create reliable reward signals for slow and expensive experiments while preventing reward gaming, and whether self-improvement compounds, plateaus, or collapses over successive rounds. The intern will collaborate with researchers and engineers within the organization and across Meta, including FAIR, and will have latitude to define a research question. The goal is to produce a full paper for ICLR, ICML, or NeurIPS and contribute research that can be applied to Meta product development. Internships last 12 to 16 weeks, or 24 weeks, with various start dates throughout the year.

Requirements

Currently has or is in the process of obtaining a Ph.D. in Machine Learning, Artificial Intelligence, Computer Science, Information or Multimedia Retrieval, Computer Vision, Natural Language Processing, Reinforcement Learning, Optimization, Computational Statistics, Applied Mathematics, or a related technical field.
Must obtain work authorization in the country of employment at the time of hire and maintain ongoing work authorization during employment.
Experience with deep learning frameworks such as PyTorch or TensorFlow.
Experience with Python, C++, C, Java, or other related languages.
Experience with research and building systems based on machine-learning and/or deep-learning methods.

Nice To Haves

Intent to return to a degree program after completion of the internship or co-op.
Research experience in one or more of: LLM agents, reinforcement learning for LLMs, automated ML research, AutoML, recommender systems, or evaluation of agentic systems.
A proven track record of significant research results, demonstrated by grants, fellowships, patents, or first-authored publications at leading workshops or conferences such as NeurIPS, ICML, ICLR, AAAI, KDD, IJCAI, CVPR, ICCV, ACL, NAACL, ICASSP, or similar.
Demonstrated software engineering experience through an internship, work experience, coding competitions, or widely used open-source contributions.
Demonstrated experience and self-driven motivation in solving analytical problems using quantitative approaches.
Experience manipulating and analyzing complex, high-volume, high-dimensionality data from varying sources for large-scale training.
Experience working and communicating cross-functionally in a team environment.

Responsibilities

Conduct state-of-the-art research on recursive self-improvement, autoresearch, and self-improving AI agents.
Design and evaluate agents that form hypotheses, use tools, run experiments, analyze outcomes, and learn from prior research trajectories.
Develop methods to improve the underlying LLM through reinforcement-learning post-training and/or optimize the surrounding harness, including prompts, tools, memory, retrieval, and orchestration.
Build reliable evaluation and reward signals for long-running, expensive ML experiments, including safeguards against reward hacking and regressions.
Study whether gains compound across multiple self-improvement rounds and rigorously attribute improvements to model learning versus harness changes.
Collaborate with researchers and engineers across varied disciplines, including communicating research plans, progress, and results.
Publish research results at a top-tier machine learning conference and contribute research that can be applied to Meta product development.