AI Research Scientist: Multimodal Foundation Models - Architecture, Pre-Training & Distillation

What you'll do

• Develop and evaluate new model architectures, pre-training objectives, and data strategies for multimodal foundation models.
• Design teacher-student learning techniques to transfer capabilities from large models to smaller, efficient versions.
• Conduct large-scale pre-training experiments using PyTorch or JAX across various modalities like text, image, video, and audio.
• Optimize model architectures specifically for memory, latency, and energy constraints on Apple silicon.
• Research and implement advanced distillation methods including offline, on-policy, and representation transfer techniques.
• Conduct research on compute-optimal scaling, including data mixtures, curricula, and training objectives.
• Translate successful research findings into production-ready technologies for integration across the Apple ecosystem.

What we're looking for

• Master's degree or equivalent practical experience in machine learning, computer science, or a related technical field.
• Hands-on experience designing, implementing, and running large-scale pre-training experiments for large language models.
• Experience with LLM pre-training topics including architecture, training objectives, data mixtures, tokenization, scaling, and optimization.
• Proficiency with modern deep learning frameworks such as PyTorch or JAX and distributed training systems.
• Strong understanding of transformer-based architectures and current approaches to efficient or scalable foundation-model training.
• Experience evaluating pre-trained models across language understanding, reasoning, instruction following, or multimodal capabilities.
• Preferred experience contributing to major foundation-model pre-training efforts or leading architecture experiments for large training runs.
• Preferred experience with knowledge distillation techniques and designing teacher-student training pipelines.