Research Scientist, Multimodal

Join the founding team at Reka to push the boundaries of large-scale multimodal foundation models that reason across text, image, video, and audio. You will contribute to the development of our frontier models, such as Reka Core, Flash, and Edge, by designing novel architectures and training paradigms. This role offers the unique opportunity to work on the full lifecycle of model development, from data ingestion to massive-scale distributed training and reinforcement learning.

- Design and implement state-of-the-art multimodal architectures for processing interleaved text, image, and video data. - Scale training runs across thousands of GPUs, optimizing for efficiency and numerical stability. - Develop innovative data curation pipelines to extract high-quality knowledge from massive, unstructured datasets. - Research and apply advanced RLHF and DPO techniques to align multimodal models with human intent. - Collaborate with the systems team to optimize inference kernels and model compression for real-world deployment. - Stay at the forefront of AI research, incorporating recent breakthroughs into the Reka production pipeline. - Contribute to the technical roadmap and long-term vision of Reka’s universal intelligence models.

- Ph.D. in Computer Science, Mathematics, or a related field, with a focus on Deep Learning or Computer Vision. - Proven track record of publishing at top-tier venues such as NeurIPS, ICML, CVPR, or ICLR. - Strong proficiency in PyTorch and experience with large-scale distributed training frameworks (e.g., Megatron-LM, DeepSpeed). - Demonstrated experience training models with billions of parameters on high-performance compute clusters. - Solid understanding of multimodal alignment techniques, including cross-attention mechanisms and contrastive learning. - Ability to work independently in a fast-paced, remote-first research environment across global time zones. - Fluent in Python and C++ with a commitment to writing high-quality, reproducible research code.