Method
Constrained decoding
Force LLM output to a grammar or schema via logit masking.
6 cases from 6 companies apply Constrained decoding. 2 were checked in an online A/B test; 2 report results in production.
Typical tasks: Recommend to new users and new items, Balance relevance and diversity, Recommend from what the user just did, Rank search results and 2 more. Companies: Netflix, VTB, Instacart, Raiffeisenbank, Alfa-Bank and 1 more. Often combined with: LLM-as-a-judge, Gradient boosting, Embedding similarity matching, LLM labeling, Label noise detection and 29 more.
| GenPage: Towards End-to-End Generative Homepage Construction at Netflix | media-streaming | 2026-07 | |
Fox One: Video Content Personalization Powered by GenAI![]() | ![]() | media-streaming | 2026-06 |
| An Automated Request Handling Service for SMBs | banking | 2026-05 | |
| Our Early Journey to Transform Instacart’s Discovery Recommendations with LLMs | e-commerce | 2026-02 | |
| Improving a Modular Chatbot with LLMs: Practical Experience | banking | 2025-05 | |
| History of the Unified Classifier Development | banking | 2023-05 |
