Reproduce “Multi-Modal Semantic Expansion with Constrained LLM Reranking for Conversational Music Recommendation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Naman Garg, Sarika Jain, George Fazekas
Researchers from the National Institute of Technology, Kurukshetra, and Queen Mary University of London engineered a multi-modal conversational music recommender system for the ACM RecSys 2026 TalkPlayData Challenge. This system, which achieved a composite score of 0.3213 (37th place), integrated seven distinct embedding spaces with lexical retrieval and optimized weighted Reciprocal Rank Fusion, while also identifying a 'Conservative Injection Principle' for beneficial, albeit narrow, LLM-guided reranking.
Key insights
- Optimized RRF weights showed that behavioral signals (user collaborative filtering) and explicit artist substring matches were substantially more influential than content-based signals in music recommendation.
- LLM-guided reranking interventions are most effective within a narrow, conservative operating range (e.g., when 0-5 target artist tracks are initially present), as aggressive application can lead to significant performance degradation, potentially due to LLM position bias.
- Successfully balancing a multi-objective evaluation metric required careful component design and trade-offs between retrieval accuracy, response quality, and diversity, particularly when deploying resource-constrained models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.