MMQ Advances Multimodal Fusion and Aligns Behavior for Large-Scale Recommendation Models
The paper introduces MMQ, a multimodal mixture‑of‑quantization semantic‑ID framework that compresses item multimodal features via shared‑specific expert networks and behavior‑aware fine‑tuning, achieving lower reconstruction loss, superior recall and ranking performance, and online gains of +1.29% REV, +4.33% CVR, +2.61% GMV, and +1.18% ROI.
