Farewell to Item IDs: Unlocking the Scaling Potential of Large Ranking Models via Semantic Tokens

TL;DR

This work addresses the instability and scalability limits of item ID embeddings in large ranking models by introducing semantic tokens and a token-based framework (TRM). TRM combines collaborative-aware multimodal item representations, a hybrid tokenization scheme that separates generalization (gen-tokens) from memorization (mem-tokens), and a joint discriminative–generative training objective to exploit token sequence structure. Empirically, TRM outperforms both ID-based and prior token-based baselines, achieving notable offline AUC gains and substantial reductions in sparse parameter size, while demonstrating favorable scaling laws as model capacity and compute grow. The approach yields practical benefits in production, improving user-active days and reducing the change-query ratio in online experiments, and offers a viable path toward removing item IDs in large-scale ranking systems. The work thereby provides theoretical and empirical support for semantic-token based scaling in LRMs and highlights the importance of balancing generalization and memorization through hybrid tokenization and joint optimization.

Abstract

Recent studies on scaling up ranking models have achieved substantial improvement for recommendation systems and search engines. However, most large-scale ranking systems rely on item IDs, where each item is treated as an independent categorical symbol and mapped to a learned embedding. As items rapidly appear and disappear, these embeddings become difficult to train and maintain. This instability impedes effective learning of neural network parameters and limits the scalability of ranking models. In this paper, we show that semantic tokens possess greater scaling potential compared to item IDs. Our proposed framework TRM improves the token generation and application pipeline, leading to 33% reduction in sparse storage while achieving 0.85% AUC increase. Extensive experiments further show that TRM could consistently outperform state-of-the-art models when model capacity scales. Finally, TRM has been successfully deployed on large-scale personalized search engines, yielding 0.26% and 0.75% improvement on user active days and change query ratio respectively through A/B test.

Figures (5)

• Figure 1: Illustration of different impacts of item IDs and semantic tokens to the scaling up of ranking models. Semantic tokens demonstrate more stable distribution as the number of parameter scales.
• Figure 2: Framework of TRM.
• Figure 3: AUC Gain compared to ID baseline by substituting different types of semantic tokens for item IDs. We report results divided by item life time when the query request happened.
• Figure 4: Scaling behavior under dense-parameter and compute budgets. We plot CTR qAUC gain over the 7M MLP baseline as a function of dense parameters (left) and training FLOPs per batch with bs=4096 (right). TRM-RankMixer (token-based) consistently dominates other token/ID baselines across scales and the qAUC gain margin further enlarges with increasing model capacity, demonstrating a steeper and more favorable scaling trend.
• Figure 5: The change of AUC gain for items of different life time when varying the number of mem-tokens. We report evaluation results of CTR AUC compared to using only gen-tokens.

Theorems & Definitions (1)

• Proposition 1.1: Quantization floor shift under CTR loss