MiniMaestro: Resource-Conscious Conversational Music Recommendation with a Single Open-Weight 8B Model
TL;DR
The RecSys Challenge 2026 Music-CRS (TalkPlay) task formalizes this as two coupled sub-problems: given dialogue history and user context, retrieve a ranked list of the top-20 tracks from the full, unrestricted catalog, and generate a response that justifies the recommendation while sustaining conversational coherence.
Abstract
Conversational interfaces are replacing static recommendation lists in music discovery, requiring systems that reason jointly over what to recommend and how to justify it in natural language. The RecSys Challenge 2026 Music-CRS (TalkPlay) task formalizes this as two coupled sub-problems: given dialogue history and user context, retrieve a ranked list of the top-20 tracks from the full, unrestricted catalog, and generate a response that justifies the recommendation while sustaining conversational coherence. Team Overfit & Chill’s system, MiniMaestro, selectively rewrites the dialogue into a structured query for lexical retrieval, pools candidates from ten heterogeneous retrieval channels, and reranks them with a LightGBM LambdaRank model that learns each source’s reliability from per-candidate rank, presence, and context features. A single open-weight Qwen3-8B model performs both query planning and response generation, using no proprietary LLM APIs and no multi-stage large-model generation. On the official Blind B leaderboard, MiniMaestro reaches an nDCG@20 of 0.33 and an LLM-as-a-Judge score of 4.15/5, for a composite of 0.48.
