Journal of Zhejiang University

ENGINEERING Information Technology & Electronic Engineering

Accepted manuscript available online (unedited version)

Decoupled prompt-guided mixture-of-experts dynamic distillation for multimodal recommendation

Author(s): Lili WU¹, Renmin ZHANG^1*, Bin ZHANG¹, Jincheng ZHANG³, Xi CHEN², Yingjing QIAN⁴
Affiliation(s): ¹School of Communication and Electronic Engineering, Jishou University, Jishou 416000, China ²Tencent Inc., Shenzhen 518057, China ³Ant Group Co., Ltd., Changsha 410000, China ⁴School of Electronic and Information Engineering, Huaihua University, Huaihua 418000, China
Corresponding email(s): Renmin ZHANG, rzhang1981@163.com
Key Words: Multimodal recommendation, Knowledge distillation (KD), Prompt tuning, Mixture-of-experts (MoE), Dynamic temperature scheduling

\|Next Paper >>>

Lili WU¹, Renmin ZHANG^1*, Bin ZHANG¹, Jincheng ZHANG³, Xi CHEN², Yingjing QIAN⁴. Decoupled prompt-guided mixture-of-experts dynamic distillation for multimodal recommendation[J]. Journal of Zhejiang University Science ,in press.Frontiers of Information Technology & Electronic Engineering,in press.https://doi.org/10.1631/ENG.ITEE.2026.0160

@article{title="Decoupled prompt-guided mixture-of-experts dynamic distillation for multimodal recommendation",
author="Lili WU¹, Renmin ZHANG^1*, Bin ZHANG¹, Jincheng ZHANG³, Xi CHEN², Yingjing QIAN⁴",
journal="Journal of Zhejiang University Science ",
year="in press",
publisher="Zhejiang University Press & Springer",
doi="https://doi.org/10.1631/ENG.ITEE.2026.0160"
}

%0 Journal Article
%T Decoupled prompt-guided mixture-of-experts dynamic distillation for multimodal recommendation
%A Lili WU¹
%A Renmin ZHANG^1*
%A Bin ZHANG¹
%A Jincheng ZHANG³
%A Xi CHEN²
%A Yingjing QIAN⁴
%J Journal of Zhejiang University SCIENCE
%P
%@ 2095-9184
%D in press
%I Zhejiang University Press & Springer
doi="https://doi.org/10.1631/ENG.ITEE.2026.0160"

TY - JOUR
T1 - Decoupled prompt-guided mixture-of-experts dynamic distillation for multimodal recommendation
A1 - Lili WU¹
A1 - Renmin ZHANG^1*
A1 - Bin ZHANG¹
A1 - Jincheng ZHANG³
A1 - Xi CHEN²
A1 - Yingjing QIAN⁴
J0 - Journal of Zhejiang University Science
SP -
EP -
%@ 2095-9184
Y1 - in press
PB - Zhejiang University Press & Springer
ER -
doi="https://doi.org/10.1631/ENG.ITEE.2026.0160"

Abstract
Chinese Summary
Academic Network
Reviewer Comment

Abstract: Multimodal recommendation aims to enrich preference modeling by leveraging visual and textual features. However, integrating high-dimensional pretrained features introduces substantial computational overhead. While knowledge distillation provides an effective compression strategy, existing frameworks face three intertwined challenges: rank bottlenecks caused by low-dimensional projections, cross-modal interference induced by shared fusion spaces, and optimization instability under static distillation temperatures. To address these issues, we propose ProMoE-DTS, a decoupled prompt-guided mixture-of-experts framework with dynamic temperature scheduling. Using an asymmetric teacher–student architecture, the teacher model leverages modality-aware soft prompts as semantic anchors to route heterogeneous features into parameter-disjoint expert networks, thereby alleviating cross-modal conflicts and resolving the rank bottlenecks. To ensure stable knowledge transfer, a feedback-driven dynamic temperature scheduler adaptively regulates the distillation intensity based on epoch-wise signals. This asymmetric design confines intensive multimodal operations to the offline teacher, leaving the online student model with a highly efficient, pure identifier-based structure. Extensive experiments on three benchmark datasets demonstrate that ProMoE-DTS improves Recall@20 by 2.24%–3.96% over state-of-the-art baselines, while requiring only 3.28%–3.55% of the teacher’s parameters.

Darkslateblue:Affiliate; Royal Blue:Author; Turquoise:Article

Reference

Open peer comments: Debate/Discuss/Question/Opinion

<1>

Full Text: <55>

CLC number:

On-line Access: 2026-09-29

Received: 2026-05-26

Revision Accepted: 2026-08-04

Crosschecked: 0000-00-00

Cited: 0

Clicked: 34

Citations: Bibtex RefMan EndNote GB/T7714

Darkslateblue:Affiliate; Royal Blue:Author; Turquoise:Article

Reference

- Go to