Routing Discovery in Conversational Music Recommendation

Open-catalog ranking in conversational music recommendation tends to plateau well before a model looks exhausted, and this paper explains why. Studying the RecSys Challenge 2026 Music-CRS task with Compass—a retrieval-first system that scores an upper-quartile nDCG@20 of 0.4304 (composite 0.5475) on the official Blind-B split—we decompose the plateau by request type. Turns split into continuation (the listener extends an artist already in the session) and discovery (the listener wants a new artist), and the history-based continuity prior that dominates continuation actively suppresses the roughly 28% of turns that are discovery; that novel-artist slice sits near nDCG 0.01 and holds much of the remaining headroom. Our main claim is that this is a routing problem, not a capacity one: continuation-versus-discovery intent is predictable from the query embedding (AUC 0.788), and a learned gate that softly scales continuity by that intent recovers part of the novel slice in-domain (a significant + 0.008 to + 0.010 nDCG@20 on discovery turns), reallocating capacity between regimes rather than adding it—so aggregate nDCG is unchanged by design. The same levers do not move Blind-B nDCG, so the core of the ceiling is structural—a one-to-many labeling property of the data rather than a tuning gap. The reusable lesson is to treat personalization as a signal to route by predicted intent, not merely to strengthen.