Skip to content

Details

This meetup session will be about LLM's understanding vs. reliance on statistics in search and recommendation. Our speaker will be Aleksandr Petrov from Spotify, who will present his works from RecSys 2026 in form of an interactive tutorial.
Location: Science Park 904 - Room C3.161
Date & Time: Friday, November 27, 16:00–17:00 (CET)
Zoom link: https://uva-live.zoom.us/j/67382138261
We’ll also post our future events on GatherEasy.

Details below:
Speaker: Aleksandr V. Petrov
Title: Counting vs. Understanding: When Do Language Models Really Help in Search and Recommendation?
Abstract:
Language models are rapidly becoming the default tool for search and recommendation, from Transformer-based sequential recommenders to LLM rerankers and generative systems. But how much of their success comes from understanding, and how much from the behavioural statistics that simpler methods have always exploited? This talk is based on three of our papers from Spotify, all presented at RecSys 2026.

First, I show that on widely used sequential recommendation benchmarks, simple recency-weighted pairwise transition models match or outperform strong Transformer baselines. This suggests that many benchmarks cannot distinguish genuine sequence modelling from co-occurrence counting. Second, I turn to production search, where behavioural signals are powerful enough to become a shortcut. Injecting them into an LLM reranker improves ranking quality, but it can crowd out the semantic and user-context patterns needed for sparse or unseen queries. I describe a simple training strategy that keeps the gains while remaining robust when those signals are missing. Finally, I look at the long tail of user taste, where behavioural data runs out. There, LLM-generated natural-language hypotheses can act as a planning layer for personalised shelves on Spotify Home.

Together, these results suggest a practical principle: rely on behaviour where it is dense, on semantics where it is sparse, and always test against strong simple baselines before crediting a model with understanding.
The talk is based on:

Bio:
Aleksandr V. Petrov is a Research Scientist at Spotify, where he focuses on making generative language models work effectively for search and recommendation. He completed his PhD in Machine Learning at the University of Glasgow, supervised by Prof. Craig Macdonald and Prof. Iadh Ounis. His doctoral research on Transformer models for sequential recommendation received the Best Paper Award at RecSys 2023 and a Best Student Paper nomination at RecSys 2022. Before his PhD, he was a Senior Software Development Engineer at Amazon, where he led large-scale ML projects for advertising and personalisation. He also co-founded E-Contenta, a startup providing white-label recommender systems for media platforms. His research has been published at SIGIR, RecSys, WSDM, ECIR and in ACM TORS. He co-organises the Unified Search and Recommendation Workshop (USRW) at RecSys and serves as a PC member for SIGIR, RecSys, WWW and ECIR.

You may also like