Research Engineer, Gemini Multi-turn

Job description

Scope and drive research efforts to improve complex frontier Gemini capabilities, such as multi-turn, factuality and tool-use. Review the latest literature to guide research and experimental directions. Curate and generate data to evaluate and improve Gemini capabilities. Design and implement both human and automated evaluation strategies. Design and conduct supervised fine-tuning and reinforcement learning experiments to improve the performance of Gemini capabilities such as multi-turn and factuality. Collaborate with partners and product functions to deliver new model capabilities to production. Minimum Qualifications: Bachelor's degree in Computer Science, Artificial Intelligence, a related technical field, or equivalent practical experience. 5 years of software engineering experience using Python. 1 year of experience working on large language model post-training (e.g., SFT, RLHF, DPO, PPO), model alignment, or core generative model development in an industry AI lab, research institute, or frontier AI organization. Experience working with deep learning frameworks (e.g., JAX/Flax) for distributed systems and accelerator clusters. One or more first-author papers accepted at or published in an LLM-related conference such as NeurIPS, ICML, ICLR, ACL, or EMNLP. Preferred Qualifications: PhD degree in Computer Science, Artificial Intelligence, Machine Learning, or a related technical field. 5 years of experience in industry AI research labs or premier research institutes developing, post-training, and deploying frontier large models. Experience designing and scaling automated evaluation frameworks and data curation pipelines for generative models. Experience in complex multi-turn conversational modeling, long-context reasoning, tool-use/function calling, or search grounding. Experience in optimizing distributed training performance, memory efficiency, and throughput on massive accelerator topologies (e.g., model/pipeline/tensor parallelism, FSDP). Strong publication record with multiple first-author or core-contributor papers in premier machine learning or NLP conferences.