ML Researcher for the Early-binding Architectures Team

We are Yandex's R&D team that transfers approaches from the LLM world to recommender systems. Our goal is to create next-generation personalization based on transformer architectures.

We build fundamental models from scratch, experiment with training paradigms, and implement SOTA architectures into key services handling hundreds of thousands of RPS. There are no rigid frameworks or ready-made answers here — we are the ones forging the path. We have expertise, a huge volume of unique behavioral data, and a dedicated GPU cluster for large-scale ML experiments.

What tasks await you

Models and Training We represent a user as an ordered set of events — a history — which is encoded using transformer models into a compressed representation and used for candidate generation and ranking. In the field of recommender systems, there is no established approach, so we actively experiment with task formulation, training recipes, and architectures.

Researching Early-binding Models Early binding of user and candidate in transformer models is the model's ability to 'see' the history and the candidate via the attention mechanism. Such models are computationally expensive at runtime, complex to train, but provide a significant quality boost.

Adapting Models for Production An important challenge for us is making models work at runtime under high load. We explore architectural optimizations and use specialized inference frameworks, sometimes even writing our own CUDA kernels in Triton.

Publishing Results at International Conferences We encourage writing papers and traveling to top conferences. Last year we presented Yambda at RecSys 2025; the next goal is publishing new approaches to ranking.

Opportunity for Broad Development As an R&D team, we are not limited to one product or one technology. If desired, you can try different approaches in recommendations or dive into other services.

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

We expect you to

• Write maintainable code in Python and have excellent command of PyTorch
• Not just use ready-made architectures, but understand the underlying mathematics of transformers, optimizers, and distributed training (DDP/FSDP)
• Be able to quickly reproduce ideas from papers and bring hypotheses to final conclusions
• Have engineering savvy: understand why a model trains slowly, how to utilize GPU at 100%, and how to speed up inference

Will be a plus

• Follow trends in RecSys, LLM, or VLM and regularly read papers
• Have submitted publications to A/A*-conferences: NeurIPS, ICML, ICLR, RecSys, KDD, ACL
• Have been involved in competitive programming, participated in ML competitions, or hackathons