Original title: Encoder Language Models for Zero-Shot Recommender Systems: Cross-Domain and Cross-Lingual Evaluation
Patrik Müller, Bogdan Walek, Radim Farana · September 2026
Recommender systems often struggle to suggest items to new users who have little history, an issue known as the cold-start problemThe difficulty of making recommendations when there is no prior user history or data.. Modern generative language models can help, but they are expensive and complex. To fix this, researchers tested over 20 top encoder language modelsA type of language model that reads and understands text by converting it into a list of numbers; it focuses on understanding meaning rather than generating new text. to build a simpler, zero-shotThe system can make recommendations for items or users it has never seen before, without any extra training. system without extra training. The system converts item descriptions into embeddings, which are lists of numbers that capture meaning. It blends these item embeddings into weighted user profiles. Then, it compares each profile to all possible item candidates using cosine similarityA way to measure how similar two lists of numbers are by calculating the angle between them; closer to 1 means very similar. to find the best match. The team tested models by revealing user interactions step by step over time. They found that using these frozenThe model's parameters are not changed during the recommendation process—it stays exactly as it was after initial training. embeddings is a viable choice with a strong cost-performance ratio. However, teams cannot rely on general benchmark scores or model size. They must test models directly on their specific domain tasks.
You walk into a bookstore and ask for a recommendation, but the clerk has never seen you before. Without knowing your reading history, they might guess based on bestsellers, which often misses the mark. In this paper, the system instead reads the words of the books you just looked at, builds a quick summary of those words, and finds new books with similar word patterns. This works without ever having seen your past purchases, just like the clerk using the books in your hands right now.
This research shows that you can build a recommendation system without training a custom model, using pre-frozen language models instead. That saves vast computing resources and works even for new users or items. It also works across different domains and languages, making it more flexible and cost-effective than traditional approaches or large generative models.
This study does not claim that frozen embedding recommenders always beat trained models. The comparison with large language models was limited to one experiment. Results may not generalize across all domains; selection must be tuned per task. The temporal evaluation simulates cold-start but may not cover all real-world scenarios. The approach is evaluated on only three datasets, so broader validation is needed.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Encoder Language Models for Zero-Shot Recommender Systems: Cross-Domain and Cross-Lingual Evaluation — Patrik Müller, Bogdan Walek, Radim Farana — Computers — doi:10.3390/computers15090639
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.