IDProxy: CTR Prediction with Multimodal LLMs for Cold-Start Recommendation at Xiaohongshu

Multimodal information can improve the accuracy of click-through rate (CTR) prediction and effectively alleviate item cold-start and long-tail problems. Recent studies commonly discretize pretrained multimodal embeddings into semantic identifiers (SIDs), allowing the model to learn task-specific semantic representations for recommendation. However, existing methods still provide limited gains due to two major limitations. First, codebook assignment fails to preserve semantic relevance and discards fine-grained continuous signals in the original embedding space. Second, the residual code paths are highly dependent on prefix codes, which limits the effective representational scalability of hierarchical identifiers. To address these issues, we propose PaletteID (PID), a prototype-based semantic identifier. Inspired by palette-based color composition, PID uses a compact set of representative prototype items as semantic anchors to bridge pretrained multimodal content space and recommendation models. Specifically, we first construct a prototype palette with Semantic Quality-Aware Determinantal Point Process (SQ-DPP), which jointly considers local content density and global semantic diversity. Then, for each target item, PID retrieves a sequence of semantically related prototypes and aggregates them into an informative PID representation, enabling rich and complementary semantic modeling. Extensive experiments on two public datasets demonstrate that PID consistently improves CTR prediction and yields larger gains for long-tail items. PID also produces more robust identifier assignments and provides more interpretable token semantics than existing residual SID methods.

Cold-start item recommendation remains a persistent challenge in real-world systems due to the absence of interaction histories. While prior models attempt to bridge this gap using item content features, they universally suffer from the \textbf{seesaw dilemma}: enhancing performance for cold items inevitably degrades performance for warm items, and vice versa. We identify that this dilemma stems from a fundamental \textbf{distributional disparity}: warm item embeddings occupy a complex behavioral manifold" shaped by rich interaction signals, whereas cold item embeddings are constrained to a semantic manifold" derived solely from auxiliary content. Existing methods often force a rigid mapping between these inconsistent spaces, causing the model to sacrifice the precision of warm representations to accommodate cold ones. To address this, we propose \textbf{DiffCold}, a diffusion-based generative model that unifies warm and cold representations. Unlike GANs or VAEs, DiffCold leverages conditional diffusion to reconstruct warm item embeddings from content, preserving the underlying manifold structure without degradation. We further tailor this paradigm with two specific designs: a \textbf{Retrieval-enhanced Aggregator} that initializes generation using semantically similar warm items to bypass inefficient noise, and a \textbf{Simulation-based Representation Alignment} module that enforces distribution consistency between generated and real embeddings via contrastive learning. Experiments on three benchmarks confirm that DiffCold resolves the seesaw dilemma, consistently outperforming state-of-the-art methods across all metrics.

Early ranking stages in recommendation systems precompute item embeddings and cache them in-model for scoring within strict latency constraints. Because this cache exists only at serving time, outside the training loop, training and serving use different item representations, a structural discrepancy that limits quality and adds operational fragility. We show that co-designing the training and serving paths removes this representation discrepancy at its source. We introduce the memory layer, an in-model key-value embedding cache co-trained with the model: the item tower writes embeddings during training and the model reads them at serving, one source of truth for item representations by construction. Always-on embeddings cover items not yet cached, so every item receives a prediction, and the design consolidates three separate trainer-to-predictor update paths into a single self-contained pipeline. Deployed in production on Instagram Reels, the memory layer raises prediction coverage from 96% to 100%, improves embedding freshness from $O(5\text{ min})$ to $O(20\text{ s})$, and narrows the training-serving Normalized Entropy (NE) gap by up to 86%, yielding over $2\times$ recall for the freshest content and a 5-6% cold start engagement lift. Because embeddings are produced during training, the system needs no separate bulk-evaluation or publish-time recomputation, cutting training-and-publish computational cost by 30% at neutral serving computational cost.