Jacqueline Wood

Jacqueline Wood builds personalized, language-steerable generative recommenders. Her 2026 AI Engineer conference biography described her as a Staff Machine Learning Engineer at Spotify, with applied research focused on adapting open-weight large language models using semantic IDs to connect natural-language intent with Spotify catalog entities.

In “Teaching LLMs to Speak Spotify,” presented with Yves Raimond, Wood explains Spotify’s NEO training approach: quantizing content embeddings into catalog tokens, grounding those tokens while freezing the model backbone to preserve language capabilities, then applying multitask instruction tuning and optional further post-training. She illustrates how a listener’s request and listening history can produce both a podcast recommendation and a natural-language explanation. She presents findings on transfer between recommendation tasks, including benefits for cold-start audiobooks, and discusses the accuracy and latency tradeoffs of beam search and constrained decoding. Her presentation also covers evaluating generative recommendations with LLM judges grounded in listening-history profiles and behavioral signals to better reflect human preferences and user intent.