Method

DPO / preference optimization

Preference tuning without RL loop (DPO, CPO, IPO).

6 cases from 6 companies apply DPO / preference optimization. 1 was checked in an online A/B test; 1 report results in production.

Typical tasks: Analyze user behavior, Label data with LLMs, Evaluate models, Recommend from what the user just did. Companies: Pinterest, Spotify, Netflix, Wildberries, Yandex and 1 more. Often combined with: Supervised fine-tuning, LLM-as-a-judge, Knowledge distillation, LoRA fine-tuning, Synthetic data generation (LLM) and 21 more.