Jacqueline Wood

Jacqueline Wood builds language-steerable generative recommenders that connect a listener’s words and listening history with real catalog items—and explain why those choices fit. In 2026, she was a Staff Machine Learning Engineer at Spotify, working on adapting open-weight language models to understand both natural-language intent and Spotify’s catalog.

From data science to generative recommendation

Wood’s career at Spotify developed from data science into machine learning engineering. She took on project leadership, onboarded colleagues, and helped other engineers build their first ML pipelines. Her professional history describes a technical transition accompanied by mentoring and helping colleagues put machine learning into practice.

By November 2025, Wood had contributed to Spotify’s semantic-ID personalization effort. The approach gives language models a vocabulary for catalog entities, allowing them to work with ordinary language and structured references to content in the same response. A fluent recommendation must also identify something a listener can actually play.

In 2026, she co-authored NEO, a framework that brings search, recommendation, and item-grounded text generation into one model, and GLIDE, a production podcast-discovery system. Both are collaborative Spotify projects. NEO supplies the general modeling framework; GLIDE applies it to podcast discovery, where recommendations must work within production cost and latency constraints.

Teaching models the catalog without losing language

In “Teaching LLMs to Speak Spotify,” presented with Yves Raimond, Wood explains the modeling choices that make language-steerable recommendation possible:

• Semantic IDs as a catalog vocabulary: Content embeddings are converted into sequences of discrete tokens and added to an open-weight model’s vocabulary. A request for a podcast about morality, combined with listening history represented by those tokens, can produce both an episode identifier and an explanation of its relevance. The model learns to generate language and references to catalog items together.

• Preserving language ability during adaptation: NEO separates learning the catalog vocabulary from learning recommendation tasks. During initial grounding, the pretrained backbone and original embeddings remain frozen while the new token embeddings learn their relationship to language. The model is then tuned across several tasks, with optional further post-training. Wood emphasizes this sequence because catalog adaptation must preserve the language understanding that makes user control possible. Multitask training also allows learning to transfer between content types, including help for cold-start audiobook recommendations.

• Balancing accuracy and latency: Wood describes the team’s choice of beam search after top-p sampling reduced recommendation accuracy in their comparisons. Constrained decoding adds latency but can help restrict results to particular content, such as new items. These choices connect the model’s ability to generate catalog identifiers with the practical requirements of serving useful recommendations.

Discovery and grounded evaluation

GLIDE combines recent listening and context with compact representations of longer-term preferences to help listeners find unfamiliar episodes and shows. In large online experiments, the collaborative system increased non-habitual podcast streaming by up to 5.4% and new-show discovery by up to 14.3%, while meeting production constraints. Its aim is to make discovery useful beyond a listener’s established favorites. GLIDE research

Wood also emphasizes grounded recommendation evaluation. Listening metrics can record an interaction, but they cannot by themselves establish whether an explanation is accurate or a recommendation matches someone’s intent. Her approach to evaluating generative recommenders gives LLM judges relevant context: summaries of listening history for podcast recommendations, and behavioral evidence from similar searches for search evaluation. The evaluator needs enough context to judge what fits this listener, especially when a request is ambiguous.