Conversational recommender systems have recently incorporated vision‑language models (VLMs) to align user dialogue intents with multimodal product information. Extending this to many product domains faces a dilemma: deploying a separately adapted VLM per domain scales cost linearly and requires routing each dialogue to the correct model, while merging all domains into a single VLM hurts accuracy because product characteristics vary widely.
LightCRS proposes a query‑conditioned lightweight framework for unified multi‑domain multimodal conversational recommendation. Offline, frozen encoders pre‑compute multimodal product embeddings. Online, a compact 5.3M‑parameter module fuses these embeddings for each dialogue via query‑conditioned cross‑attention, sharing parameters across all domains; an inter‑candidate Transformer then jointly compares candidates.
To enable rigorous evaluation, the authors build MM‑R2A‑25, a benchmark covering 25 domains that pairs natural Reddit recommendation‑seeking conversations with Amazon product listings. LightCRS achieves state‑of‑the‑art results on HR@1, NDCG@5, and MRR, improving HR@1 by 35.4% over the strongest baseline and reducing per‑query latency by 266×. Joint multi‑domain training further boosts HR@1 by 15.2% compared to separately trained single‑domain models.
Review