RESEARCH · RESEARCH · #2328
LightCRS: lightweight query-conditioned multi-domain multimodal conversational recommender
LightCRS is a lightweight framework that precomputes frozen multimodal product representations offline and uses a compact 5.3M-parameter query-conditioned cross-attention module plus an inter-candidate Transformer to fuse evidence online for unified multi-domain conversational recommendation. Evaluated on MM-R2A-25, a new 25-domain benchmark pairing Reddit recommendation-seeking conversations with Amazon product listings, LightCRS achieves state-of-the-art average HR@1, NDCG@5, and MRR, boosting HR@1 by 35.4% over the strongest baseline and reducing per-query latency by 266×.
KEY POINTS
- LightCRS is a lightweight framework that precomputes frozen multimodal product representations offline and uses a compact 5.3M-parameter query-conditioned cross-attention module plus an inter-candidate Transformer to fuse evidence online for unified multi-domain conversational recommendation.
- Evaluated on MM-R2A-25, a new 25-domain benchmark pairing Reddit recommendation-seeking conversations with Amazon product listings, LightCRS achieves state-of-the-art average HR@1, NDCG@5, and MRR, boosting HR@1 by 35.4% over the strongest baseline and reducing per-query latency by 266×.
- Provides a practical, low-cost method to scale multimodal conversational recommendation across many product domains while substantially improving accuracy and latency, avoiding per-domain VLM deployment.
WHY IT MATTERS
Provides a practical, low-cost method to scale multimodal conversational recommendation across many product domains while substantially improving accuracy and latency, avoiding per-domain VLM deployment.