Research Scientist, Multi-Modal Human Understanding

Responsibilities

Design and implement novel multi-modal architectures that fuse vision, language, and temporal signals for holistic human understanding Develop and train Vision-Language Models (VLMs) for tasks including visual question answering, image-text reasoning, and grounded human-centric understanding Build video foundation models capable of temporal reasoning, action synthesis and long-form video synthesis with applications to human behavior synthesis Research generative synthesis techniques for human-centric content including video generation, motion synthesis, and multi-modal content creation Conduct rigorous experiments to evaluate model performance across diverse benchmarks, analyze failure modes, and iterate on architectures to improve accuracy and generalization Contribute to the full research lifecycle from problem formulation and dataset curation through model development and evaluation

Qualifications

Currently has, or is in the process of obtaining a Bachelor's degree in Computer Science, Computer Engineering, relevant technical field, or equivalent practical experience. Degree must be completed prior to joining Meta 2+ years of experience in multi-modal AI research, including hands-on work with Vision-Language Models, video understanding, or human-centric AI systems 2+ years of experience implementing and training large-scale neural networks using frameworks such as PyTorch, with experience on transformer-based architectures Experience designing and executing experiments to evaluate multi-modal model performance, including quantitative analysis across vision, language, and video benchmarks Experience writing production-quality or research-quality code in Python for multi-modal AI applications Experience developing or fine-tuning Vision-Language Models for human understanding tasks Experience with video foundation models, temporal transformers, or large-scale video pretraining Track record of contributing to published multi-modal AI research at venues such as CVPR, ICCV, or NeurIPS Experience with generative models for human synthesis including diffusion models, GANs, or autoregressive models for video or motion generation