Releases·Tool Release
Distill the LLM, Don't Serve It: Search & Personalization at DoorDash — Raghav Saboo, DoorDash
Raghav Saboo, Staff ML Engineer for Search & Personalization at DoorDash, describes using LLMs offline to solve semantic understanding in marketplace discovery, then distilling that reasoning into small serving models. The approach supports DoorDash's expansion into grocery, retail, pets and gifting across billions of store-level items.
42 engagement1 sourceFri, Sep 25, 2026, 01:30 PM
AI Overview
- DoorDash Staff ML Engineer Raghav Saboo distills offline LLM reasoning into small serving models for search across grocery, retail, pets and gifting.
- LLM-generated relevance labels improved retrieval NDCG by 2.3%; learned semantic IDs raised ranking MRR 4-5% and power query reformulation.
- Consumer memory spans three timescales stored as text, vectors and graphs; LLM-generated personalized collections lifted pets order rate nearly 1%.
- Saboo argues marketplace discovery's real bottleneck is semantic understanding, not engagement, which can rank wrong items like regular spaghetti for gluten-free queries.
Related
Developments·Paper
Distill Globally, Adapt Locally: Reasoning Distillation and Product-Type Test-Time Training for Scalable Trade-Up Recommendation
An arxiv paper introduces a two-level framework for trade-up recommendation, distilling LLM reasoning into an efficient non-generative student model. The student's decision boundary is adapted per product type via test-time training, avoiding the impracticality of running LLMs directly over hundreds of millions of product pairs.
An arxiv paper introduces a two-level framework for trade-up recommendation, distilling LLM reasoning into an efficient non-generative student model. The student's decision boundary is adapted per product type via test-time training, avoiding the impracticality of running LLMs directly over hundreds of millions of product pairs.
0 engagement1 sourcearxivFri, Sep 4, 2026, 05:08 PM
Developments·Paper
Beyond Relevance: Structured Semantic Supervision for Product Search with LLM-Augmented Annotations
A new arXiv paper proposes augmenting query-product pairs with structured LLM-generated attributes plus human-validated relevance, explanations, and centrality judgments for e-commerce search. Evaluated with a dual-encoder retriever and MLP re-ranker on an augmented ESCI subset, synthetic signals approach human-oracle performance, with the largest gains on difficult, low-performing queries.
A new arXiv paper proposes augmenting query-product pairs with structured LLM-generated attributes plus human-validated relevance, explanations, and centrality judgments for e-commerce search. Evaluated with a dual-encoder retriever and MLP re-ranker on an augmented ESCI subset, synthetic signals approach human-oracle performance, with the largest gains on difficult, low-performing queries.
0 engagement1 sourcearxivSun, Sep 20, 2026, 01:49 PM
Developments·Paper
arXiv paper proposes intent-conditioned recall expansion for e-commerce search
A new arXiv paper presents a scalable system for discovery-augmented search in e-commerce, particularly grocery, that generates implicit user intents to expand candidate recall while maintaining relevance. The approach addresses the limitation of traditional search systems that prioritize precision over recall, improving discoverability of substitute, complementary, and thematically related items.
A new arXiv paper presents a scalable system for discovery-augmented search in e-commerce, particularly grocery, that generates implicit user intents to expand candidate recall while maintaining relevance. The approach addresses the limitation of traditional search systems that prioritize precision over recall, improving discoverability of substitute, complementary, and thematically related items.
0 engagement1 sourcearxivWed, Jul 29, 2026, 05:46 PM
Developments·Paper
Coupled Calibration and Learning: Mitigating Teacher Bias in LLM Distillation without Target-Domain Reward Feedback
Two September 2026 arXiv papers tackle distinct failure modes in LLM distillation. iSDFT treats the teacher as a budgeted information source during self-distillation fine-tuning, improving specialisation while limiting forgetting. CCL couples teacher calibration with student updates to mitigate teacher bias without target-domain reward feedback, with a polynomial convergence guarantee.
Two September 2026 arXiv papers tackle distinct failure modes in LLM distillation. iSDFT treats the teacher as a budgeted information source during self-distillation fine-tuning, improving specialisation while limiting forgetting. CCL couples teacher calibration with student updates to mitigate teacher bias without target-domain reward feedback, with a polynomial convergence guarantee.
0 engagement2 sourcesarxivTue, Sep 15, 2026, 05:15 PM
Social·Community
Teaching LLMs to Speak Spotify — Yves Raimond & Jacqueline Wood, Spotify
A YouTube talk by Spotify's Yves Raimond and Jacqueline Wood explains how Spotify rebuilt personalization around LLMs. The Large Taste Model now serves one in four US Premium subscribers daily, and Wood details NEO, a four-stage training recipe using semantic IDs on a frozen open-weight backbone.
A YouTube talk by Spotify's Yves Raimond and Jacqueline Wood explains how Spotify rebuilt personalization around LLMs. The Large Taste Model now serves one in four US Premium subscribers daily, and Wood details NEO, a four-stage training recipe using semantic IDs on a frozen open-weight backbone.
40 engagement1 sourceyoutubeFri, Sep 25, 2026, 02:30 PM