PhD Student · Information Retrieval
David Vos
I am a PhD student at the Information Retrieval Lab Amsterdam, under the supervision of Maarten de Rijke and Harrie Oosterhuis. I study representation learning, generative modeling, and ranking for personalized recommendations and search. I collaborate with the Responsible Media Lab to apply my research at DPG Media. In 2026, I was a research intern in Bloomberg's AI Group in New York, where I worked on retrieval with instructions beyond topical relevance.
I am actively seeking research internships in information retrieval. Open to research collaborations.
Selected Research
Prior-Data Fitted Networks as Tabular Foundation Models for Ranking in Low-Data Settings
We propose to use prior-data fitted networks as tabular foundation models for ranking in low-data settings. We achieve state-of-the-art results in settings with limited training data.
LSNR: Learned Sparse News Recommendations with Attentive User Modeling
We introduce LSNR, which encodes each news article independently and combines the sparse representations into a user vector with learned attention. LSNR outperforms STAR, outperforms dense recommenders for users with short histories, and serves from an index an order of magnitude smaller.
A Study of Normative Diversity Metrics in News Recommendations
We analyze normative diversity metrics for news recommendations, quantifying their dependence on attribute representation pipelines and editorial context. We also quantify trade-offs between normative diversity and recommendation accuracy.
Towards Parameter-Efficient Automation of Data Wrangling Tasks with Prefix-Tuning
We show that prefix-tuning LLMs can automate data wrangling tasks while requiring only 0.9% of the parameters of full fine-tuning.
Recognised as best paper runner-up.
Feature-based Propensities for Counterfactual Learning to Rank
We estimate propensities from item and context features to correct for position bias in counterfactual learning to rank.
Efficient Optimization of Hierarchical Identifiers for Generative Recommendation
We show that generative models with hierarchical identifiers lead to superior recommendation performance when compared to strong baselines, but are bottlenecked in training time by the tree construction. We propose a greedy clustering algorithm that reduces the tree construction time drastically, without sacrificing performance.
Revisiting Language Models in Neural News Recommender Systems
We evaluate how different language models perform as the backbone of neural news recommenders, finding that smaller models can outperform larger ones when properly adapted.
Beyond Topical Relevance: Retrieval with Sorting Instructions over Latent Document Properties
We study how retrieval models can follow sorting instructions over latent document properties, extending relevance beyond topical similarity.
Education
-
PhD in Information Retrieval
University of Amsterdam
2023 - present · IRLab · with DPG Media
-
Master in Artificial Intelligence
University of Amsterdam
2019 - 2022 · cum laude
-
Minor in Computer Science
University of Edinburgh
2018 · Erasmus exchange
-
Bachelor in Artificial Intelligence
University of Amsterdam
2016 - 2019 · cum laude · honours programme