YandexGPT Reasoning Team Lead

YandexGPT Reasoning Team Lead

Our team trains the YandexGPT reasoning model. It is used in the 'Reason' mode in Alice and internal products. We influence the entire training process — from pre-training and SFT to reinforcement learning.

We strive to make YandexGPT a leader in solving complex tasks: from math and programming olympiads to handling demanding user queries. Our goal is to create a language model that sets standards in the field of reasoning LLMs.

Join us to influence the development of YandexGPT's intelligence, solve unique technological challenges in the field of AI, and participate in shaping the future of language models.

What tasks await you

Developing the Reasoning Model Your main task is to create the strongest reasoning model. You will improve dataset collection pipelines for all training stages, implement algorithmic innovations in RL, identify growth points, plan experiments, and test new approaches.

Research at the Frontier of Scientific Knowledge You will be responsible for defining the scientific and technical agenda and enhancing the team's expertise. You will oversee the entire research cycle: from analyzing SOTA and formulating hypotheses to planning experiments, evaluating quality, and interpreting results.

Leading a Team of Researchers You will lead a team of senior researchers: set ambitious goals, develop expertise and scientific culture, and establish collaboration with DL, analytics, and infrastructure teams. Another important task is to find and hire the best specialists, as projects are becoming increasingly complex and large-scale.

We expect you to

• Have led an LLM team or want to grow in this direction
• Have excellent knowledge of classical ML/DL and NLP
• Have practical or research experience in reinforcement learning
• Follow trends in language models and can quickly understand and evaluate papers

Nice to have

• Have trained reasoning language models
• Are actively interested in papers on reasoning models and agents
• Have scientific publications and have presented at conferences