UxSID: Semantic-Aware Sequence Modeling

• The paper introduces UxSID, a framework that leverages Semantic IDs and dual-level attention to model ultra-long user interactions with enhanced target specificity.
• It shifts heavy computation offline by creating semantic-group shared interest memory, balancing the trade-off between item-specific search and item-agnostic compression.
• Empirical evaluations show UxSID achieves superior AUC and latency improvements in large-scale industrial settings, underscoring its practical scalability and monetization benefits.

UxSID is an end-to-end semantic-aware user interest modeling framework for ultra-long user sequence modeling in recommendation and advertising. It addresses the efficiency–effectiveness trade-off that arises when user histories extend to thousands, and in some analyses up to $10{,}000$, past interactions by positioning itself between two established paradigms: item-specific search and item-agnostic compression. Its defining mechanism is the use of Semantic IDs (SIDs) and a dual-level attention strategy to construct semantic-group shared interest memory, so that target-aware user representations can be precomputed offline and served online through constant-time lookup (Zhang et al., 9 May 2026).

1. Problem setting and conceptual position

The UxSID paper frames ultra-long user sequence modeling as a systems problem in which heavy users may interact with up to $10{,}000$ items per week, making full target-conditioned attention over the entire history costly at serving time. It distinguishes two prior solution families. The first, item-specific search, includes methods such as SIM and TWIN, which retrieve a small target-relevant subset from the full history before applying a more expressive model. The second, item-agnostic compression, includes methods such as MIMN, LURM, and C-Former, which compress the full history into dense memories or user embeddings that are efficient online but less target-aware (Zhang et al., 9 May 2026).

UxSID is presented as a third path: semantic-group shared general user interest memory. Rather than building a distinct compressed memory for every target item, it builds memories shared by semantically similar items. This shifts the unit of conditioning from exact item identity to semantic group. The paper’s central claim is that this preserves target awareness more effectively than purely target-agnostic compression while avoiding the online cost and retrieval bias of item-specific search.

This positioning is important because the paper does not argue only for a better architecture. It also redefines the serving interface between long-sequence modeling and ranking. UxSID moves the expensive stages of sequence processing offline, then reintroduces target specificity online through SID-indexed retrieval. In that sense, it is both a modeling proposal and a deployment strategy.

2. Semantic IDs and the representational basis of the model

The key primitive in UxSID is the Semantic ID. A SID is a discrete semantic code assigned to an item, intended to reflect semantic similarity rather than arbitrary identity. The paper constructs SIDs by encoding multimodal item attributes with an MLLM encoder into a continuous embedding $z_i$, then quantizing that embedding with a hierarchical codebook (Zhang et al., 9 May 2026):

$$z_i = \mathrm{Enc}_{\mathrm{MLLM}}(\mathrm{Attributes}_i)$$

$$k_m = \arg \min_j \left\| r^{m-1} - C_{m,j} \right\|_2^2, \qquad r^m = z_i - \sum_{\ell=1}^{m} C_{\ell}(k_\ell), \qquad r^0 = z_i$$

This yields a hierarchical semantic code sequence $(k_1, k_2, \dots, k_M)$. In the deployed configuration described in the paper, the first-layer code $k_1$ is mainly used as the target SID because it offers a practical compromise between semantic resolution and serving cost. The result is a coarse-to-mid semantic cluster that is finer than broad categories or tags but not as specific as the item ID itself.

This design serves two functions. First, it provides a reusable query key for target-conditioned interest extraction. Second, it provides the indexing unit for offline embedding storage. The paper explicitly compares SIDs with category tags and reports that SIDs are more effective, which supports the claim that learned semantic tokenization is not interchangeable with conventional category metadata.

3. Architecture: compression, semantic probing, and final user representation

The model architecture consists of SID generation, Item-Agnostic Interest Compression (IAIC), hierarchical semantic probing, and integration into the final ranking model. The long user history is embedded as $$E = [e_1, e_2, \dots, e_L] \in \mathbb{R}^{L \times d}$$, while the target item contributes its SID-derived semantic query (Zhang et al., 9 May 2026).

IAIC compresses the ultra-long sequence into a small set of latent interests using learnable interest anchors $Q_{\text{anc}} \in \mathbb{R}^{K \times d}$. The cross-attention formulation is:

$$H = \mathrm{Softmax}\!\left( \frac{(Q_{\text{anc}} W_Q)(E W_K)^\top}{\sqrt{d}} \right)(E W_V)$$

Each anchor is then refined by an anchor-specific Per-token Feed-Forward Network, yielding $10{,}000$0. To prevent collapse of anchors onto the same dominant pattern, the model adds an orthogonality regularizer:

$10{,}000$1

The paper uses $10{,}000$2 IAIC anchors in the reported configuration.

The most distinctive component is hierarchical semantic probing, which uses the target SID in two stages. The first stage performs explicit semantic probing over the full raw sequence:

$10{,}000$3

The second stage uses a gate derived from the global response to refine the target semantic query:

$10{,}000$4

The refined query then attends over the compressed anchors:

$10{,}000$5

The final UxSID user representation concatenates the two levels:

$10{,}000$6

This construction is central to the paper’s claim that UxSID is neither a pure retrieval model nor a pure compression model. The explicit global probe preserves access to fine-grained signals in the raw history, while the local probe over compressed anchors provides a denoised latent-interest view.

4. Training, serving, and computational properties

The final ranking model consumes target item features $10{,}000$7, user profile features $10{,}000$8, context features $10{,}000$9, short-term behavior features $z_i$0, and the UxSID representation $z_i$1. The prediction layer is:

$z_i$2

Training is end-to-end with binary cross-entropy plus the orthogonality regularizer. Public experiments use Adam, batch size $z_i$3, learning rate $z_i$4, and NVIDIA L20 GPUs. For public datasets, the codebook shape is $z_i$5; in industrial deployment, the first-level codebook size is $z_i$6 and the embedding lookup dimension is $z_i$7 (Zhang et al., 9 May 2026).

The deployment mechanism is a critical part of the method. UxSID embeddings are generated offline and stored in an embedding server indexed by user and semantic group:

$z_i$8

The embedding server is updated once per week, and online serving uses point lookup by $z_i$9. The downstream model consumes compressed UxSID embeddings of shape $$z_i = \mathrm{Enc}_{\mathrm{MLLM}}(\mathrm{Attributes}_i)$$0 incrementally.

The paper emphasizes inference complexity. It reports $$z_i = \mathrm{Enc}_{\mathrm{MLLM}}(\mathrm{Attributes}_i)$$1 for UxSID, compared with $$z_i = \mathrm{Enc}_{\mathrm{MLLM}}(\mathrm{Attributes}_i)$$2 for C-Former and higher sequence-dependent terms for SIM-Hard, SIM-Soft, ETA, SDIM, MIRRN, and TWIN. The stated consequence is that UxSID’s online cost depends on compressed interest length $$z_i = \mathrm{Enc}_{\mathrm{MLLM}}(\mathrm{Attributes}_i)$$3 and hidden size $$z_i = \mathrm{Enc}_{\mathrm{MLLM}}(\mathrm{Attributes}_i)$$4, not on original sequence length $$z_i = \mathrm{Enc}_{\mathrm{MLLM}}(\mathrm{Attributes}_i)$$5. In the industrial analysis, the appendix reports about $$z_i = \mathrm{Enc}_{\mathrm{MLLM}}(\mathrm{Attributes}_i)$$6 unique SIDs per user on average in Kuaishou deployment and an embedding-server footprint of about $$z_i = \mathrm{Enc}_{\mathrm{MLLM}}(\mathrm{Attributes}_i)$$7 TB for $$z_i = \mathrm{Enc}_{\mathrm{MLLM}}(\mathrm{Attributes}_i)$$8 million active users. The same analysis states that a current online model with sequence length $$z_i = \mathrm{Enc}_{\mathrm{MLLM}}(\mathrm{Attributes}_i)$$9k occupies about $$k_m = \arg \min_j \left\| r^{m-1} - C_{m,j} \right\|_2^2, \qquad r^m = z_i - \sum_{\ell=1}^{m} C_{\ell}(k_\ell), \qquad r^0 = z_i$$0 A10 GPUs, naively scaling direct online ranking to $$k_m = \arg \min_j \left\| r^{m-1} - C_{m,j} \right\|_2^2, \qquad r^m = z_i - \sum_{\ell=1}^{m} C_{\ell}(k_\ell), \qquad r^0 = z_i$$1k would require about $$k_m = \arg \min_j \left\| r^{m-1} - C_{m,j} \right\|_2^2, \qquad r^m = z_i - \sum_{\ell=1}^{m} C_{\ell}(k_\ell), \qquad r^0 = z_i$$2 A10 GPUs, and using UxSID adds only $$k_m = \arg \min_j \left\| r^{m-1} - C_{m,j} \right\|_2^2, \qquad r^m = z_i - \sum_{\ell=1}^{m} C_{\ell}(k_\ell), \qquad r^0 = z_i$$3 ms latency over the baseline without UxSID embeddings (Zhang et al., 9 May 2026).

5. Evaluation and empirical performance

The empirical evaluation covers XLong, KuaiRec-Big, and a large-scale industrial advertising dataset from Kuaishou, with public benchmarks reporting AUC and industrial tasks reporting AUC, UAUC, and WUAUC. XLong contains $$k_m = \arg \min_j \left\| r^{m-1} - C_{m,j} \right\|_2^2, \qquad r^m = z_i - \sum_{\ell=1}^{m} C_{\ell}(k_\ell), \qquad r^0 = z_i$$4 users, $$k_m = \arg \min_j \left\| r^{m-1} - C_{m,j} \right\|_2^2, \qquad r^m = z_i - \sum_{\ell=1}^{m} C_{\ell}(k_\ell), \qquad r^0 = z_i$$5 items, $$k_m = \arg \min_j \left\| r^{m-1} - C_{m,j} \right\|_2^2, \qquad r^m = z_i - \sum_{\ell=1}^{m} C_{\ell}(k_\ell), \qquad r^0 = z_i$$6 interactions, average sequence length $$k_m = \arg \min_j \left\| r^{m-1} - C_{m,j} \right\|_2^2, \qquad r^m = z_i - \sum_{\ell=1}^{m} C_{\ell}(k_\ell), \qquad r^0 = z_i$$7, and max sequence length $$k_m = \arg \min_j \left\| r^{m-1} - C_{m,j} \right\|_2^2, \qquad r^m = z_i - \sum_{\ell=1}^{m} C_{\ell}(k_\ell), \qquad r^0 = z_i$$8. KuaiRec-Big contains $$k_m = \arg \min_j \left\| r^{m-1} - C_{m,j} \right\|_2^2, \qquad r^m = z_i - \sum_{\ell=1}^{m} C_{\ell}(k_\ell), \qquad r^0 = z_i$$9 users, $(k_1, k_2, \dots, k_M)$0 items, $(k_1, k_2, \dots, k_M)$1 interactions, average sequence length $(k_1, k_2, \dots, k_M)$2, and max sequence length $(k_1, k_2, \dots, k_M)$3. The industrial dataset is based on Kuaishou impression logs from April 1–7, 2026 and preserves history up to $(k_1, k_2, \dots, k_M)$4k (Zhang et al., 9 May 2026).

Setting	Reported result	Context
XLong	AUC 0.8408	Higher than DIN 0.7889, SIM-Hard 0.8201, TWIN 0.8135, C-Former 0.8154
KuaiRec-Big	AUC 0.8348	Reported as stronger than all listed baselines
Industrial CTR	AUC 0.8728, UAUC 0.6125, WUAUC 0.6161	Higher than SIM-Hard, SIM-Soft, and TWIN
Industrial CTCVR	AUC 0.8626, UAUC 0.6269, WUAUC 0.6350	Higher than SIM-Hard, SIM-Soft, and TWIN
Online A/B test	Exposure +0.111%, Cost +0.231%, Revenue +0.337%	Week-long deployment on Kuaishou short-video advertising platform

The paper treats the A/B test as especially consequential because the method is evaluated under production latency constraints. It also notes that the revenue improvement exceeds the exposure improvement, which it interprets as suggesting higher precision and better monetization rather than merely more impressions. The scaling-law analysis further reports that UxSID’s advantage widens at $(k_1, k_2, \dots, k_M)$5k sequence length (Zhang et al., 9 May 2026).

6. Ablations, interpretability, and limitations

The ablation results are organized around the model’s main claims. Replacing SIDs with category or tag features degrades performance and strongly reduces Int.R@50, which the paper uses to argue that SIDs provide denser and more discriminative semantic routing. Removing $(k_1, k_2, \dots, k_M)$6 harms performance, indicating that compressed anchors alone miss some fine-grained target-to-history alignments. Removing $(k_1, k_2, \dots, k_M)$7 also harms performance, showing that direct raw-sequence attention alone is insufficient. Removing the gating network reduces performance, which supports the use of the refined query $(k_1, k_2, \dots, k_M)$8. Removing $(k_1, k_2, \dots, k_M)$9 weakens results, supporting the claim that anchor diversity matters (Zhang et al., 9 May 2026).

The hyperparameter analysis reports that performance improves as the number of anchors $k_1$0 increases, with the best setting around $k_1$1. Too small a value, such as $k_1$2, produces an overly severe bottleneck; too large a value introduces redundancy and diffuse routing. The study also reports that moderate orthogonality regularization works best: too little allows anchors to collapse, while too much over-separates the latent space.

The interpretability analysis is qualitative rather than causal. Different target SIDs produce different attention patterns over the same user history, and positive samples tend to have much higher Int.R@50 than negative samples. A t-SNE analysis of IAIC anchors is reported to show diverse coverage of the behavior space. This suggests that the model is capturing multi-interest structure rather than merely amplifying frequent behaviors.

The paper does not foreground a dedicated limitations section, but several limits are either stated or strongly implied by the design. A plausible implication is that performance depends on SID quality, since target-aware routing is built around semantic tokenization. Another plausible implication is that the method is best suited to settings with content-rich items, because SID construction relies on multimodal item attributes. The offline storage and weekly refresh regime also indicate a structural trade-off: UxSID shifts cost from online computation to offline embedding generation and storage. The paper’s own deployment analysis treats this trade-off as acceptable in the reported industrial setting, but it remains part of the method’s systems profile (Zhang et al., 9 May 2026).

7. Broader interpretive context

The exact term UxSID refers, in the direct arXiv usage, to the ultra-long sequence framework described above. A broader reading of “UxSID” as “UX-oriented socio-technical and infrastructure design” (Editor’s term) is not the definition used in the recommendation paper, but adjacent research makes such a reading conceptually suggestive. In science gateways, UX is framed not as a late-stage usability check but as a “design-oriented perspective for reasoning about sustainability,” with recurring orientations described as ad hoc, project-based, and strategic (Parsons, 24 Oct 2025). In designer–developer collaboration research, UX–software integration is characterized as a coordination, communication, artifact, and human-relations problem, synthesized into three challenge categories, eleven subcategories, and six best-practice families (Zhang et al., 20 Jan 2025). In AI-enabled Human-in-the-Loop systems, UX is redefined as a multi-layer construct spanning backend performance, workflow integration, governance, and trust, with four sociotechnical metrics—Accuracy, Operational Latency, Adaptation Time, and Trust—proposed as an evaluative scaffold (Mafi et al., 4 Mar 2026).

Other adjacent work reinforces this broader socio-technical interpretation. Integrating UX practices into software processes has been argued to be difficult because UX is subjective and emergent, producing eleven recurring organizational challenges and at least five issues tied to the nature of UX itself (Kashfi et al., 2016). Open-source issue tracker research reports that UX professionals are scarce in usability issue discussions, but when they participate they tend to produce more factual, better-supported issue reports and six recurring follow-up behaviors such as sharing design ideas and giving extra information (Sanei et al., 24 Feb 2025). Privacy-oriented design research similarly treats UI/UX designers as privacy advocates whose work is shaped by business goals, team power relations, technical implementation, and the tension between minimum compliance and user-centered privacy (Yao et al., 27 Apr 2026). In dataset search, UX is organized into six areas—Initial Interaction, Search Process, Dataset Exploration, Filtering and Sorting, Dataset Actions, and Assistance and Feedback—and linked to ten interface design prototypes centered on metadata accessibility and decision support (Zhao et al., 2024).

These papers do not define the UxSID recommendation architecture. They do, however, suggest a wider pattern in current research: whether in recommender systems, science gateways, AI oversight systems, software processes, open-source collaboration, privacy design, or dataset search, the decisive question is often how user-facing experience is tied to underlying semantic structure, workflow organization, and long-term system viability.