Senior DL Developer for Neuro Team

Neuro is a multimodal product of the future, where the power of generative models is combined with various sources of external information, the list of which is constantly expanding: web search, image search, business information on Maps, etc. We have implemented such a system in Yandex Search and are now facing a new challenge: learning to solve complex scenarios that arise in chats with Alice.

We are developing an LLM evaluator and reward models — these are key components of the Neuro pipeline: their evaluations directly impact how Yandex's neural networks learn, generate, and analyze. Our LLM assessor not only detects errors but also explains them, bringing us closer to creating a system that can think, analyze, and improve. It is we who are steering Neuro towards the generative product of the future.

Join us to compete with international IT giants and build the product of the future in the present!

What tasks await you

Improving Neuro in Alice You will refine the Neuro alignment process using reward models and the LLM evaluator, as well as solve related tasks connected to alignment.

Research in the field of LLM-as-a-judge You will conduct experiments with test-time scaling approaches for the LLM evaluator, which not only assigns scores but also explains them.

Improving the LLM evaluator You will need to improve the LLM evaluator at all stages of its training: from annealing to GRPO, and also develop a multimodal VLM evaluator: we aim to teach the LLM assessor to evaluate not only text but also other multimodal enrichments of the response.

More about Alice AI

More about ML at Yandex — in the Yandex for ML channel

We expect you to

• See the mathematics behind the PyTorch code: understand how LLMs work "under the hood"
• Know how to turn scientific papers into code: have implemented SOTA methods and algorithms
• Possess broad knowledge and competencies in NLP and DL

Will be a plus

• Have deep knowledge in the field of LLMs and RL
• Have experience with large models and distributed training