Research Scientist Intern - ML Innovation for Products

The ML Innovation for Products (MLIP) team is an applied machine learning and research team within Meta's Ranking & Foundational AI org, focused on the state-of-the-art (SOTA) 0-to-1 ML that powers retrieval, ranking, and auction across Meta's ads systems. This is a very unique team that functions like an applied research lab with a production deploy button: we imagine, evaluate, prototype, and ship 10x ideas into one of the largest recommender systems in the world, a system that serves billion-scale users every single day and is the engine behind Meta's revenue. As a horizontal team, we have the unique leverage to develop SOTA techniques alongside expert engineers and researchers, then scale them into the head models powering Facebook, Instagram, and beyond. What you would work on Because the team works at the boundary of research and production, our research and its applications are intertwined, and much of the work is published at major ML and recommender systems conferences. Some of the work the team is involved in includes: Memora, a new scaling paradigm for ads recommendation inspired by memory research in LLMs. Learned representations in large recommenders turn out to be highly concentrated, so much of the input signal is effectively ignored and further architectural scaling yields flattening returns. We build memory modules that store and retrieve historical model-generated representations and inject them back as differentiating signals, improving model quality without proportionally scaling compute. Two workstreams: Learning from Memory (LFM) and Enriched Ad Sequences (EAS). Graph Learning at billion-node scale. We build heterogeneous graph representations of users, ads, and their interactions, including user journey graphs, and we push the modeling frontier: k-hop induced subgraph methods, temporal GNNs, graph transformers, and deeper architectures with residual and spectral rewiring to counter over-smoothing and over-squashing. A parallel track turns this into a self-serve platform so that experiment throughput, not infrastructure, is the bottleneck. LLM and Multimodal Fusion. Ad-side and user-side content understanding are both crowded research areas, but almost no one combines them with graph-structured relational signals. That gap is the opening. Our SELF workstream, for Semantic-Engagement LLM Fusion, contrastively trains ad and user encoders so that engagement structure from the interaction graph and semantic content from multimodal LLMs land in one shared embedding space that retrieval can search directly. Agentic AI. We are building a shared agentic layer that abstracts planning, tool orchestration, memory, and execution into reusable services. The flagship is IDEA, which has already taken techniques from paper to validated production model improvements end to end, with domain overlays that specialize it for memory, sequences, graph learning, retrieval policies, and foundation models. The research questions it raises about autonomous experimentation are wide open. Reinforcement Learning. Large parts of the ads retrieval flow are still tuned by hand: how many candidates to keep at each stage, how to mix retrieval routes, where to set recall cutoffs. These decisions have no clean ground-truth label to supervise against, which is exactly the gap RL is built to close. We replace static global constants with per-request policies learned from logged traffic that optimize the realized system outcome directly rather than a low-fidelity offline proxy. Our first production policy personalizes a retrieval keep ratio per request on Facebook Mobile Feed. The longer-term direction is joint multi-knob optimization across the flow on one shared reward and off-policy evaluation foundation. Who we are: We are a team of 15 ML engineers and research scientists spanning Sunnyvale and the Bay Area, supported by senior technical leads with deep experience taking research from idea to production. We partner closely with product groups across Meta, so the surface area available to an intern is much wider than any single pod. The team values knowledge sharing and continuous learning, and runs regular technical deep dives and AI best-practice workshops alongside a buddy system for onboarding.