[Under Review 2026] IDProxy: Bridging the Semantic-Collaborative Gap for Cold-Start Ads at Xiaohongshu

TL;DR

Xiaohongshu (RedNote) engineers have solved the notorious "Cold-Start" problem in CTR prediction by using IDProxy. This system uses Multimodal LLMs (MLLMs) to generate "proxy" embeddings for new items that behave exactly like well-trained ID embeddings. By aligning MLLM hidden states with existing recommendation architectures, they achieved a +1.93% increase in Advertiser Value and a 2x AUC improvement for new content compared to global traffic.

The Semantic Mismatch: Why Cold-Starts are Hard

In industrial systems, CTR models are "ID-centric." They learn specific embeddings for every post or ad based on user interactions. When a new post is uploaded, it has no ID history—it is invisible to the collaborative filtering logic.

The researchers at Xiaohongshu pointed out a critical "Insight": Academic benchmarks (like MovieLens) are misleading. In those datasets, ID embeddings form neat clusters based on genres. In a real-world system like Xiaohongshu, ID distributions are irregular and sparse (see Figure 1), making simple MLP mappings from text/image to IDs fail.

Figure 1: Comparison between academic (Left) and industrial (Right) ID distributions. The irregular distribution in production demands a more sophisticated alignment than simple projection.

Methodology: Coarse-to-Fine Alignment

IDProxy operates in two distinct stages to bridge the gap between multimodal semantics and ranking structures.

Stage 1: Coarse Proxy Generation

The system feeds the post's image and text into an MLLM (InternVL). Instead of just taking the text output, it uses a special [EMB] token. Through Contrastive Learning, the MLLM is trained to pull the content representation toward the actual ID embeddings of high-frequency (well-trained) items.

Stage 2: CTR-Aware Fine-Graining

A key contribution of this paper is the Structural Reuse. Instead of treating the MLLM output as a "side feature," IDProxy injects it into the Atomic ID Slots of the ranking model.

• Hierarchical Partitioning: They extract hidden states from shallow, middle, and deep layers of the MLLM (using K-means to group layers).
• Lightweight Adaptor: A small MLP fuses these multi-granularity features.
• Residual Gating: This ensures the model only uses the new fine-grained info that isn't already captured by the coarse proxy.

Figure 2: The IDProxy workflow. Content features flow through an MLLM and a gated adaptor to eventually reside in the same feature interaction space as traditional ID embeddings.

Experimental Results: Real-World Impact

The results demonstrate that "plugging in" MLLM knowledge into an existing ranker is far more effective than training a new multimodal model from scratch.

• Ablation Success: Replacing static MLP mappings with the full IDProxy (Stage 1+2) increased AUC by +0.14% (which is massive for a system serving 300M+ users).
• Cold-Start Efficiency: For "New Notes" (published <24 hours), the AUC lift was 0.32%, double the gains seen on global traffic.

Critical Insight & Conclusion

The core value of IDProxy lies in its Inductive Bias. Most researchers try to make CTR models more like LLMs. Xiaohongshu did the opposite: they forced a powerful MLLM to "speak the language" of their existing CTR ranker.

By reusing the structural priors (the target attention and sequence modules) of the production model, they avoided the "cold-start" of the ranking architecture itself. This work proves that we don't need to replace our mature recommendation pipelines to benefit from the LLM revolution; we just need a smarter proxy.

Limitations

While effective, the system requires real-time inference or pre-computation of MLLM embeddings for every new post. For platforms with millions of daily uploads, the computational overhead of the MLLM encoding stage remains a bottleneck that requires careful engineering of the "IDProxy generation service."