Talk Less, Rank Better: A Decoupled Pipeline for Conversational Music Recommendation
TL;DR
This paper describes the solution submitted by team Hallucinated to the ACM RecSys Challenge 2026, based on the TalkPlayData conversational music recommendation dataset, which addresses the ranking task with a classic multi-stage pipeline that combines a diverse pool of candidate generators, Reciprocal Rank Fusion, and a learned XGBoost reranker.
Abstract
This paper describes the solution submitted by team Hallucinated to the ACM RecSys Challenge 2026, based on the TalkPlayData conversational music recommendation dataset. The Challenge poses a dual task: at every turn of a multi-turn dialogue, a system must both (i) rank the most relevant music tracks from the catalogue and (ii) generate a natural language assistant response justifying the recommendation. We address the ranking task with a classic multi-stage pipeline that combines a diverse pool of candidate generators, Reciprocal Rank Fusion, and a learned XGBoost reranker, with validation subsets shaped to resemble the Blind data and tuning objectives that favour stable performance under distribution shift. The response generation task is handled by a multi-stage local Large Language Model pipeline, optimised for lexical diversity. The complete code and data are available at https://github.com/remaplab/recsys-challenge-2026-music-crs for reproducibility.