Method
Reward model
Model scoring outputs to guide RL or filter data, incl. process rewards.
15 cases from 8 companies apply Reward model. 6 were checked in an online A/B test; 1 report results in production.
Typical tasks: Recommend from what the user just did, Label data with LLMs, Reduce ML cost, Build an assistant over internal docs and 4 more. Companies: Netflix, Yandex, Rabota.ru, Wildberries, Nextdoor, SberDevices and 2 more. Often combined with: Supervised fine-tuning, PPO, RLHF, LLM-as-a-judge, Knowledge distillation and 51 more.



