Paper Overview
Field: Machine Learning Authors: He Ma, Lan You, Qiyang Wu, Chen Liu Published: 2026-10-22 arXiv: 2610.26948
Summary
Conversational recommender systems have recently adopted vision-language models to match user dialogue intents with multimodal product information. However, extending these methods to many product domains faces a dilemma: deploying a separately adapted VLM per domain scales cost linearly with the number of domains and requires routing each dialogue to the right model, while naively merging all domains into a single VLM degrades accuracy because product characteristics vary widely across categories.
The authors propose LightCRS, a lightweight, query-conditioned framework for unified multi-domain multimodal conversational recommendation:
- Offline: multimodal product representations are precomputed once using frozen encoders.
- Online: a compact 5.3M-parameter module fuses this evidence for each dialogue through query-conditioned cross-attention, with parameters shared across all domains, and jointly compares candidate items via cross-candidate transformers.
- Benchmark: to enable rigorous evaluation, the authors construct MM-R2A-25, a 25-domain benchmark pairing natural recommendation-seeking conversations from Reddit with Amazon product listings.
- State-of-the-art on average HR@1, NDCG@5, and MRR.
- HR@1 improves 35.4% over the strongest baseline.
- Per-query latency reduced by 266x.
- Joint multi-domain training yields 15.2% higher HR@1 than individually trained single-domain models.
Results
Original Abstract
Conversational recommender systems have recently adopted vision-language models to match user dialogue intents with multimodal product information. However, extending these methods to many product domains faces a dilemma: deploying a separately adapted VLM per domain scales cost linearly with the number of domains and requires routing each dialogue to the right model, while naively merging all domains into a single VLM degrades accuracy because product characteristics vary widely across categories. We propose LightCRS, a lightweight, query-conditioned framework for unified multi-domain multimodal conversational recommendation. LightCRS precomputes multimodal product representations once offline with frozen encoders. Online, a compact 5.3M-parameter module fuses this evidence for each dialo...
---
*Auto-collected on 2026-10-22*