Reproduce “GateDiffInt: Gate-Mediated Controllable Diffusion and Multi-Intent LLM Distillation for User Behavior Modeling” on OpenResearch

About this paper

Jialong Duan, Zichen Zhang, Zirui Tu, Zheng Zhang, Zepeng Li, Qingyao Cui, Qinwen Wang, Yudan Liu and 2 more

Researchers from Xiaohongshu Inc. and Fudan University developed GateDiffInt, a framework addressing Noise–Intent Coupling in user behavior sequences for conversion rate prediction by integrating controllable diffusion-based denoising with multi-intent LLM distillation. The system achieved superior CVR prediction performance on public and industrial datasets and a +1.13% Gross Merchandise Volume uplift in online A/B tests on a major industrial platform.

Key insights

• Noise and intent in user behavior sequences are mutually reinforcing (Noise–Intent Coupling), requiring a unified, task-grounded framework for joint denoising and intent extraction rather than isolated approaches.
• Controllable diffusion, featuring behavior-importance-differentiated masking and dual gating, can effectively denoise sequences while preserving weak but informative signals critical for robust intent understanding.
• Large Language Models (LLMs) can provide strong semantic priors for explicit, structured intent modeling, and their knowledge can be efficiently distilled into lightweight student models for real-time industrial deployment using per-intent LoRA.

Summary by alphaXiv.