--- title: Hypothesis-Driven Shelf Generation for Personalised... description: Spotify Research describes a system that generates personalised shelf concepts for Spotify Home rather than picking from a fixed inventory of hand-designed... canonical: https://daily.dev/posts/hypothesis-driven-shelf-generation-for-personalised-recommendation-fxvfjt1er twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Hypothesis-Driven Shelf Generation for Personalised Recommendation | daily.dev og:description: Spotify Research describes a system that generates personalised shelf concepts for Spotify Home rather than picking from a fixed inventory of hand-designed... og:url: https://daily.dev/posts/hypothesis-driven-shelf-generation-for-personalised-recommendation-fxvfjt1er og:image: https://api.daily.dev/og/posts/fXvfjT1eR.png og:image:alt: Hypothesis-Driven Shelf Generation for Personalised Recommendation og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Hypothesis-Driven Shelf Generation for Personalised Recommendation **[Spotify Research](https://daily.dev/sources/spotify_research)** · 8 min read · 0 upvotes · 0 comments ## Summary Spotify Research describes a system that generates personalised shelf concepts for Spotify Home rather than picking from a fixed inventory of hand-designed templates. The architecture separates planning (an LLM-derived hypothesis describing a shelf concept from a listener's profile) from fulfilment (generative retrieval that grounds the hypothesis in real catalogue items via Semantic IDs), followed by an alignment stage that selects final items and rewrites the shelf title/subtitle for coherence. Everything runs offline with no LLM inference at serving time. Offline evaluation shows generative retrieval outperforming BM25 and dense/hybrid baselines, and alignment substantially improving shelf-title accuracy. Early online tests under randomised exposure show hypothesis-driven shelves are competitive for albums but underperform for podcasts and playlists, with further work planned on spoken-word content and tighter planning-fulfilment coordination. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How does Spotify generate personalised recommendation shelves instead of using fixed templates? Spotify's research separates shelf planning from fulfilment: a distilled open-source LLM generates a natural-language shelf hypothesis (a concept like 'glacial ambient post-rock with orchestral textures') from a listener's profile, then a generative retrieval model resolves that hypothesis into real catalogue items using Semantic IDs, followed by an alignment stage that selects final items and rewrites the title for coherence, all running offline. _Teams designing personalization pipelines can follow architecture writeups like this one on daily.dev._ ### How much better is generative retrieval than BM25 or dense embeddings for grounding shelf concepts in a catalogue? Generative retrieval scored 0.71 on a Hypothesis-to-Shelf Judge (0-2 scale) versus 0.56 for the strongest baseline among BM25, dense embedding retrieval, and a hybrid approach, about 27% higher, with the advantage holding across every reported dimension after Bonferroni correction. This helps when a shelf concept depends on indirect relationships not stated explicitly in item metadata. _Engineers comparing retrieval approaches for recommendation systems can track results like these on daily.dev._ ### Does adding an alignment step improve shelf title accuracy in recommendation systems? Yes, an alignment stage that reselects items and rewrites the shelf title and subtitle increased overall Hypothesis-to-Shelf Judge scores from 0.71 to 1.27, a 78% gain, with the largest improvement in title-promise fulfilment, which rose from 0.66 to 1.31, a 99% increase, addressing cases where individually relevant items fail to match what the displayed title promises. _Anyone building coherent, title-accurate recommendation UIs can follow findings like these on daily.dev._ ## Similar posts on daily.dev - [Bootstrapping Conversational Recommendation Agents at Spotify: Synthetic Data Generation and Self-Improvement Loops](https://daily.dev/posts/bootstrapping-conversational-recommendation-agents-at-spotify-synthetic-data-generation-and-self-im-przovhtlz) · Spotify Research · 0 upvotes · 0 comments - [Our Early Journey to Transform Instacart’s Discovery Recommendations with LLMs](https://daily.dev/posts/our-early-journey-to-transform-instacart-s-discovery-recommendations-with-llms-wpv1ryxzx) · Instacart · 1 upvotes · 0 comments - [How Generative Recommenders Are Redefining RecSys at Scale](https://daily.dev/posts/how-generative-recommenders-are-redefining-recsys-at-scale-hjq1asios) · NVIDIA Developer · 0 upvotes · 0 comments - [The feature we were afraid to talk about](https://daily.dev/posts/the-feature-we-were-afraid-to-talk-about-xinygsqr9) · dltHub · 1 upvotes · 0 comments --- Tags: [#machine-learning](https://daily.dev/tags/machine-learning), [#llm](https://daily.dev/tags/llm), [#spotify](https://daily.dev/tags/spotify) [View this post on daily.dev](https://daily.dev/posts/hypothesis-driven-shelf-generation-for-personalised-recommendation-fxvfjt1er) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Hypothesis-Driven Shelf Generation for Personalised Recommendation","url":"https://daily.dev/posts/hypothesis-driven-shelf-generation-for-personalised-recommendation-fxvfjt1er","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/hypothesis-driven-shelf-generation-for-personalised-recommendation-fxvfjt1er"},"datePublished":"2026-09-22T17:34:57.754Z","dateModified":"2026-09-22T17:35:39.386Z","description":"Spotify Research describes a system that generates personalised shelf concepts for Spotify Home rather than picking from a fixed inventory of hand-designed...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/ed4d5841eaa0b1dc3c9dc6e68fb92d81?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/ed4d5841eaa0b1dc3c9dc6e68fb92d81?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Spotify Research","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Spotify Research","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/a020ee10033c40eaa05cc420c7fb45af","url":"https://daily.dev/sources/spotify_research"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/hypothesis-driven-shelf-generation-for-personalised-recommendation-fxvfjt1er","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"machine-learning,llm,spotify","timeRequired":"PT8M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Spotify Research","item":"https://daily.dev/sources/spotify_research"},{"@type":"ListItem","position":3,"name":"Hypothesis-Driven Shelf Generation for Personalised Recommendation"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/hypothesis-driven-shelf-generation-for-personalised-recommendation-fxvfjt1er#faq","mainEntity":[{"@type":"Question","name":"How does Spotify generate personalised recommendation shelves instead of using fixed templates?","acceptedAnswer":{"@type":"Answer","text":"Spotify's research separates shelf planning from fulfilment: a distilled open-source LLM generates a natural-language shelf hypothesis (a concept like 'glacial ambient post-rock with orchestral textures') from a listener's profile, then a generative retrieval model resolves that hypothesis into real catalogue items using Semantic IDs, followed by an alignment stage that selects final items and rewrites the title for coherence, all running offline. Teams designing personalization pipelines can follow architecture writeups like this one on daily.dev."}},{"@type":"Question","name":"How much better is generative retrieval than BM25 or dense embeddings for grounding shelf concepts in a catalogue?","acceptedAnswer":{"@type":"Answer","text":"Generative retrieval scored 0.71 on a Hypothesis-to-Shelf Judge (0-2 scale) versus 0.56 for the strongest baseline among BM25, dense embedding retrieval, and a hybrid approach, about 27% higher, with the advantage holding across every reported dimension after Bonferroni correction. This helps when a shelf concept depends on indirect relationships not stated explicitly in item metadata. Engineers comparing retrieval approaches for recommendation systems can track results like these on daily.dev."}},{"@type":"Question","name":"Does adding an alignment step improve shelf title accuracy in recommendation systems?","acceptedAnswer":{"@type":"Answer","text":"Yes, an alignment stage that reselects items and rewrites the shelf title and subtitle increased overall Hypothesis-to-Shelf Judge scores from 0.71 to 1.27, a 78% gain, with the largest improvement in title-promise fulfilment, which rose from 0.66 to 1.31, a 99% increase, addressing cases where individually relevant items fail to match what the displayed title promises. Anyone building coherent, title-accurate recommendation UIs can follow findings like these on daily.dev."}}]} ```