The Centrality of Data in 

LLM-Based Recommendation:

Generation, Representation, and Evaluation

KEYNOTE

Abstract

Large language models are transforming recommender systems from pattern-matching engines into reasoning systems capable of understanding user intent and responding intelligently. Yet one hard-learned truth of the RecSys world remains: the most consequential design decisions are often not about the model, but about the data. In this talk, I'll explore the centrality of data in LLM-based recommendation at Spotify - touching on themes like synthetic data generation, item representation through Semantic IDs, and LLM-as-a-judge evaluation - and what this means for the next generation of recommender systems.

Building Semantic IDs with Multimodal Priors for Generative Recommendation

KEYNOTE

Abstract

The rapid development of large generative models has motivated a shift in recommender systems toward generative architectures built on compact token vocabularies. A fundamental operation in this paradigm is action tokenization: converting human-perceivable actions, such as user-item interactions, into machine-readable token sequences, often referred to as semantic IDs. In many early approaches, semantic IDs are constructed by leveraging multimodal item features as semantic priors. This talk will discuss several recent works that investigate what these multimodal semantic priors bring to generative recommendation. First, we will show how preserving semantics during action tokenization actually improves the performance of generative recommendation models by enabling better generalization. We will then discuss a less desirable consequence: when combined with autoregressive generation, semantic priors can also constrain the expressive power of semantic ID-based models. Finally, we will introduce Latte, a simple approach that alleviates these expressive limitations, with an interesting case in which Latte learns to adaptively prioritize different modalities for different users. Together, we hope these findings offer new perspectives on the understanding and modeling of semantic IDs for future generative recommender systems.

Multimodal LLMs in Action: Addressing Cold-Start Recommendation at Xiaohongshu, a Leading Chinese Social Media Platform

INVITED TALK

Abstract

Xiaohongshu, also known internationally as RedNote, is a leading Chinese content-driven social and e-commerce platform with more than 300 million users. The platform hosts a massive and continuously evolving stream of posts, images, and short videos spanning a wide range of topics, making large-scale recommender systems central to its operation.

Like many content-driven platforms, Xiaohongshu relies on click-through rate (CTR) prediction models to rank and recommend content. These models, however, depend heavily on item ID embeddings and typically struggle when newly uploaded items have little or no interaction history. This talk presents IDProxy, a production-scale system designed at Xiaohongshu to address this cold-start challenge.

IDProxy leverages multimodal large language models (LLMs) to generate proxy embeddings from rich multimodal content signals, enabling CTR prediction for new items in the absence of usage data. Through a lightweight coarse-to-fine mechanism, these proxies are aligned with the ID embedding space and trained end-to-end with the ranking model, allowing seamless integration into production-facing pipelines.

The talk discusses the motivations behind IDProxy, its design and evaluation at scale, and the industry lessons learned from bringing multimodal LLMs into a large-scale recommender system. IDProxy is now deployed across Xiaohongshu's Content Feed and Display Ads features, where it supports recommendations reaching hundreds of millions of users every day.

This talk is based on work selected for an oral presentation in the Industry Track at RecSys 2026.