# MusiCRS: Benchmarking Audio-Centric Conversational Recommendation

arXiv: https://arxiv.org/abs/2509.19469 · HTML: https://mlsanity.com/paper/2509.19469/

## Key facts

- Cited by 2 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 23 Sep 2025.

## Built on by (most cited first, 2 of 2)

- [AMPS: Adaptive Modality Preference Steering via Functional Entropy](https://mlsanity.com/paper/2602.12533/) (2026)
- [Many Right Answers, One Training Example: Removing Responder Noise from GRPO for LLM Recommenders](https://mlsanity.com/paper/2610.17771/) (2026)

## Abstract (arXiv)

Conversational recommendation has advanced rapidly with large language models (LLMs), yet music remains a uniquely challenging domain in which effective recommendations require reasoning over audio content beyond what text or metadata can capture. We present MusiCRS, the first benchmark for audio-centric conversational recommendation that links authentic user conversations from Reddit with corresponding tracks. MusiCRS includes 477 high-quality conversations spanning diverse genres (classical, hip-hop, electronic, metal, pop, indie, jazz), with 3,589 unique musical entities and audio grounding via YouTube links. MusiCRS supports evaluation under three input modality configurations: audio-only, query-only, and audio+query, allowing systematic comparison of audio-LLMs, retrieval models, and traditional approaches. Our experiments reveal that current systems struggle with cross-modal integration, with optimal performance frequently occurring in single-modality settings rather than multimodal configurations. This highlights fundamental limitations in cross-modal knowledge integration, as models excel at dialogue semantics but struggle when grounding abstract musical concepts in audio. To facilitate progress, we release the MusiCRS dataset (https://huggingface.co/datasets/rohan2810/MusiCRS), evaluation code (https://github.com/rohan2810/musiCRS), and comprehensive baselines.