Gustavo Penha

We wrote a post summarizing our #RecSys2024 paper on bridging search and recommendation with generative retrieval 🧵 (1/N) research.atspotify.com/2024/10/bridgi… w. @AliVardasbi, @denadai2, @enricopalumbo91, Hugues Bouchard

For anyone worried their LLM might be making stuff up, we made a budget‐friendly truth serum (semantic entropy + Bayesian). See for yourself: youtube.com/watch?v=x_8ORG… Paper: arxiv.org/pdf/2504.03579

I doubt to what extent improvements on these datasets would translate to improvements in today's real-world recommendation settings. Reference: arxiv.org/abs/2508.19399…

@svakulenk0 Thanks! But I won’t go this time :( my co-authors will present this time

@_Guz_ enjoy Prague ;)

Happy to share our #recsys25 paper: “Evaluating Podcast Recommendations with Profile-Aware LLM-as-a-Judge”. 🧠 90 days of listening → natural-language user profiles → LLM judges alignment 📊 Aligns with human eval. With amazing Spotify co-authors. 📄 arxiv.org/abs/2508.08777

@svakulenk0 Yes! We calculate the agreement between the llm judge and users in the paper

@_Guz_ did you ask users to evaluate?

📄 arxiv.org/abs/2508.10478

Excited to share our paper “Semantic IDs for Joint Generative Search & Recommendation” @ RecSys'25 🧠 Jointly fine-tuning embeddings for both tasks → shared Semantic IDs that work for search and recs ⚖️ 📦 No more task-specific trade-offs!

Semantic IDs for Joint Generative Search and Recommendation @_Guz_ et al. at Spotify introduce a bi-encoder model fine-tuned on both search and recommendation tasks to obtain item embeddings, followed by construction of unified Semantic ID space. 📝arxiv.org/abs/2508.10478

Evaluating Podcast Recommendations with Profile-Aware LLM-as-a-Judge Spotify introduces a profile-aware LLM framework for evaluating personalized podcast recommendations using natural-language user profiles distilled from listening history. 📝arxiv.org/abs/2508.08777

Describe What You See with Multimodal Large Language Models to Enhance Video Recommendations @denadai2 et al. at Spotify use multimodal LLMs to generate natural-language descriptions of video content for better recommendations 📝arxiv.org/abs/2508.09789 👨🏽‍💻huggingface.co/datasets/marco…

What if we could use off-the-shelf Multimodal Large Language Model to enrich current video recommendation models? This is what we asked ourselves in our recent #recsys2025 paper arxiv.org/pdf/2508.09789 🧵

🔗 Blog Post: lnkd.in/dRkuSAx8 📚 Paper: lnkd.in/dQCCKDMj

🔎 LLM alignment techniques can enhance query expansion by eliminating the need for multiple generations followed by re-ranking/filtering steps. Check out this work led by @adam_x_yang during his internship with us at @SpotifyResearch w. @enricopalumbo91 and Hugues Bouchard⬇️

Adaptive Repetition for Mitigating Position Bias in LLM-Based Ranking Spotify introduces a dynamic early-stopping method that adaptively determines repetitions needed for each ranking instance, reducing LLM calls by 81% while preserving accuracy. 📝arxiv.org/abs/2507.17788

Aligned Query Expansion: Efficient Query Expansion for Information Retrieval through LLM Alignment @adam_x_yang et al. leverage LLM alignment techniques to fine-tune models for generating query expansions that directly optimize retrieval effectiveness. 📝arxiv.org/abs/2507.11042

Contextualizing Spotify's Audiobook List Recommendations with Descriptive Shelves Spotify introduces a pipeline that generates personalized audiobook recommendations with descriptive shelves to help users explore content based on their interests. 📝arxiv.org/abs/2504.13572

Blog post: research.atspotify.com/2025/04/text2t… Paper: arxiv.org/pdf/2503.24193

The best-performing ID strategy was to use collaborative-filtering embeddings as input to the discretization approach for semantic IDs

We just published this blog post about our research on music track search with generative retrieval. 🧵 With @enricopalumbo91 @adamianou @peputo Timothy Christopher, Alice Wang, Hugues Bouchard, @mounialalmas