From Raw IDs to Semantic Planning: How Recommender Systems Utilize Information at Scale

Changhong Jin, Shiqiu Yang, Roger Zhe Li, Yingjie Niu, Aghiles Salah, Mete Sertkan, Zheng Ju, Xingsheng Guo, Huifeng Guo, Ruihai Dong, Barry SmytharXiv 2026 · arXiv

TL;DR

Recommender systems are undergoing a fundamental paradigm shift from simple item-matching engines into active, multi-stakeholder decision platforms. This paper traces the evolution from opaque "Raw IDs" to structured "Semantic IDs," and introduces a new frontier called Semantic Planning—a framework where a system first determines the strategic purpose of the next exposure before deciding how to instantiate it.

What Problem Does This Paper Address?

For the past two decades, industrial recommender systems have relied heavily on "Raw IDs"—discrete, globally unique, but semantically blank database keys. While Raw IDs are excellent for exact database lookups and memorizing specific item behaviors, they create massive technical debt. They struggle with cold-start items, lead to bloated embedding tables, and decouple an item's actual content (text, images) from its identity layer.

Crucially, traditional retrieval architectures only answer one question: Which existing item should be shown next? This formulation forces the system to compress complex, often conflicting multi-stakeholder objectives (user satisfaction, platform revenue, advertiser goals) into a single, implicit item score.

Methods

As a conceptual vision paper, the authors do not introduce a new algorithmic implementation, but rather propose a structural paradigm shift: Semantic Planning.

Instead of mapping user context directly to an item identifier, Semantic Planning splits the recommendation process into two distinct steps:

  1. Predicting the Semantic Target: The system uses structured Semantic IDs to determine an intermediate decision variable—what the next exposure should accomplish (e.g., "reassure a hesitant traveler by emphasizing flexible cancellation options").
  2. The Instantiation Layer: Once the target is set, the system decides how to fulfill it. This could mean retrieving a matching item from the catalog, generating a personalized ad creative, or flagging a signal of "unmet demand" if no suitable item exists.
Figure 1: Paradigm shift from semantic retrieval to semantic plan- ning. (a) Semantic Retrieval maps the context directly to a semantic ID to fetch an existing catalog item; (b) Semantic Planning separates the semantic target from the instantiation.

Datasets

Not discussed in the paper. (As a position paper, it focuses on architectural paradigms rather than empirical benchmarking.)

Key Insights and Ideas

  • Identity is Moving to Semantics: Modern systems are increasingly moving semantic structure directly into the identity layer itself (tokenization). This allows models to share a unified vocabulary across different domains, modalities (text/images), and tasks (search/recommendation).
  • Decoupling Intent from Inventory: By separating what the system wants to achieve from the items available, platforms can explicitly detect gaps in their inventory. If a semantic target cannot be satisfied, it becomes a valuable business signal of unmet demand.
  • Architectural Hybridity: The authors argue that the future isn't a total abandonment of the past. Industrial systems will likely adopt a hybrid approach: Raw IDs will remain essential for auditability and exact memorization of high-traffic "head" items, while Semantic IDs will drive planning, generalization, and cold-start handling.

Key Results

Not discussed in the paper.

Limitations and Future Work

The authors outline several major open challenges that the research community must solve to make semantic planning viable in production:

  • ID Stability: How to design semantic IDs that remain stable and consistent even as dynamic product catalogs constantly update.
  • Target Grounding: Ensuring the intermediate planning layer doesn't generate highly coherent but completely unsatisfiable targets that the current inventory or infrastructure cannot fulfill.
  • Objective Observability: Finding ways to supervise and evaluate multi-stakeholder planning when platform and provider objectives are only partially observable or change dynamically.
  • Rethinking Evaluation: Moving away from traditional, static evaluation metrics (like hit rates or NDCG) toward "planning resonance"—measuring how well a trajectory of decisions navigates a user toward long-term intent fulfillment using advanced simulation agents.

Summary updated

Abstract

The evolution of recommender systems can be explored by asking how they utilize information at scale. Throughout most of the historical period under consideration during the past two decades, industrial systems have relied on raw IDs, which are discrete, globally unique, and semantically opaque identifiers that enable exact lookup, logging, and item-specific memorization at scale. Over time, however, recommender systems have sought to utilize richer sources of information, including item content, context, multimodal signals, and cross-domain structure. This development has led to a new stage in which part of such information is no longer used solely as auxiliary features around item identity, but is increasingly encapsulated in semantic IDs that provide a more structured, model-facing form of identity. We argue that this shift goes beyond the rise of generative recommendation over traditional methods. Indeed, it reflects a broader evolution in how recommender systems utilize information under industrial-scale constraints. This paper looks at the past, present, and future to examine three connected questions: why raw IDs dominated the early development of recommender systems, why semantic information is increasingly being encapsulated in IDs today, and what may come next once recommendations move beyond semantic retrieval. In particular, we introduce semantic planning as a possible future direction in which the system first predicts the semantic target of the next exposure, and only then instantiates that target as a specific item or generated creative. We further argue that such a shift may require changes not only in model design but also in evaluation and in the way recommender systems coordinate the objectives of users, platforms, and providers.
BibTeX
@misc{jin2026from,
  author = {Changhong Jin and Shiqiu Yang and Roger Zhe Li and Yingjie Niu and Aghiles Salah and Mete Sertkan and Zheng Ju and Xingsheng Guo and Huifeng Guo and Ruihai Dong and Barry Smyth},
  title = {{From Raw IDs to Semantic Planning: How Recommender Systems Utilize Information at Scale}},
  year = {2026},
  eprint = {2607.09540},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2607.09540},
}
Figure