GRUT: Time-Aware Generative Recommender

• The paper introduces GRUT, a novel framework that integrates continuous-time embeddings and dual-horizon interest routing to capture evolving user preferences and periodic trends.
• It employs advanced RoPE-based techniques and time-aware objective functions to enhance retrieval, ranking, and prediction performance.
• Empirical benchmarks demonstrate significant improvements in hit rates and CTR, validating GRUT’s effective modeling of temporal dynamics in recommender systems.

Generative Recommender Using Time Awareness (GRUT) is an advanced paradigm for sequential recommendation that explicitly incorporates temporal dynamics within the generative modeling of user-item interactions. Unlike conventional architectures that focus primarily on item order or static sequence encoding, GRUT frameworks leverage sophisticated temporal context representations, continuous-time positional embeddings, and time-aware training objectives, enabling the model to capture evolving preferences and periodic interaction rhythms. These principles underlie architectures that unify retrieval and ranking, optimize for both temporal recency and long-term interests, and deliver state-of-the-art performance across diverse benchmarks and production-scale deployments (Gao et al., 26 Sep 2025, Yi et al., 16 Nov 2025, Lin et al., 22 Aug 2025, Wei et al., 23 Oct 2025).

1. Foundations of Time Awareness in Generative Recommendation

GRUT systems arise from a critical limitation in traditional generative recommenders, which typically encode user histories as discrete index-ordered sequences, neglecting wall-clock time intervals between events and thereby failing to model preference evolution and periodic temporal signals. Time-awareness in generative recommendation refers to the explicit encoding and utilization of temporal features—such as timestamps, time gaps, periodicity, and time-decay—in the architecture and inference process. This enables models to (a) distinguish between otherwise similar behavioral sequences, (b) react to trends and recency, and (c) disambiguate bursts and revisits.

The paradigm shift is typified by continuous-time rotary positional embedding (RoPE) mechanisms (Wei et al., 23 Oct 2025), dual-horizon interest routers (Yi et al., 16 Nov 2025), time-aware sequence tokenization (Lin et al., 22 Aug 2025), and time-aware objective functions (Gao et al., 26 Sep 2025, Gao et al., 26 Sep 2025, Yi et al., 16 Nov 2025). These components produce a recommender that generates next-item predictions or slates globally conditioned on both temporal and sequential order features.

2. Architectural Components and Temporal Encoding

2.1 Tokenization and Context Fusion

GRUT models structure the input sequence as heterogeneous events $$H = \{(ITEM_t, ACTION_t, TIME_t, QUERY_t)\}_{t=1}^T$$, mapping each field to learned or frozen semantic embeddings. Temporal context is represented either as continuous timestamps (Gao et al., 26 Sep 2025, Wei et al., 23 Oct 2025) or as multi-scale discrete features (month, day, hour) (Lin et al., 22 Aug 2025). The model fuses these features through an MLP or via direct token embedding, enabling fine-grained access to both order and time in self-attention and feed-forward blocks.

2.2 Time-Aware Positional Embedding

RoPE-based methods supplant position-index with actual event time, rotating query-key pairs by angles dependent on either or both discrete index $i$ and normalized timestamp $\tau_i$ (Wei et al., 23 Oct 2025, Gao et al., 26 Sep 2025, Gao et al., 26 Sep 2025). The unified Time-and-Order RoPE (TO-RoPE) formulation,

$$\theta_k(i) = (1 - \lambda_k) \alpha^{p}_k \, i \, \omega^{p}_k + \lambda_k \alpha^{t}_k \, \tau_i \, \omega^{t}_k$$

introduces learnable gates $\lambda_k$ that allocate capacity between index-locality and time-awareness, resulting in split-by-dimension or split-by-head routing strategies. This mechanism allows attention layers to preferentially leverage recency, periodicity, or order, improving retrieval and ranking metrics (Wei et al., 23 Oct 2025).

2.3 Dual-Horizon Interest Routing

Dual-Branch Long/Short-Term Router (DBR) partitions the user history into separate windows for stable long-term behavior ($\mathcal{H}_{\rm long}$) and transient short-term signals ($\mathcal{H}_{\rm short}$). Pooling and cosine similarity computations select the more relevant window in training, while both branches are evaluated and merged during inference (Yi et al., 16 Nov 2025). This architectural module enables explicit modeling of dynamic preference drift and intent bursts, equipped for rapid context adaptation.

2.4 Spatiotemporal and Multimodal Expansion

In specialized domains such as point-of-interest (POI) recommendation, spatiotemporal encoding is augmented with geographic features using hash-based geo-embeddings and hierarchical block indexing, permitting scalable autoregressive decoding across multimillion-vocabulary spaces (Lin et al., 22 Aug 2025). Incorporation of multimodal embeddings—textual and image features—fuses external semantic signals into POI representations, mitigating sparsity and long-tail item challenges.

3. Time-Aware Objective Functions and Decoding Strategies

3.1 Exposure-Aware Next-Token Prediction (ENTP-Loss)

The ENTP loss function penalizes the model for generating “stale” items that have been exposed to but ignored by the user, enforcing fast time decay of non-relevant interests. For training tuples $(x_i, s_i^{(1:L)}, c_i)$, where $c_i$ is a click/exposure label, the loss is:

$$\mathcal{L}_{\rm ENTP} = \frac{1}{N}\sum_{i=1}^N \left[ c_i \sum_{\ell=1}^L (-\log p_i^{(\ell)}) + (1-c_i)(-\alpha \log(1 - p_i^{(1)})) \right]$$

This explicit handling of negative feedback is critical for robust time-aware recommendation (Yi et al., 16 Nov 2025).

3.2 InfoNCE and Hybrid Ranking Loss

SynerGen's GRUT implementation couples an InfoNCE retrieval objective with a hybrid pointwise-pairwise ranking loss, balancing retrieval accuracy and click-probability calibration (Gao et al., 26 Sep 2025). Joint optimization regularizes generation, producing candidate slates and their ranking scores in a single backbone.

3.3 Search-Based Hierarchical Decoding

Hierarchical ID or block-based decoding restricts fine-grained candidate generation to relevant buckets selected via coarse-level decisions, reducing compute, noise, and enhancing intra-class consistency (Yi et al., 16 Nov 2025, Lin et al., 22 Aug 2025). Training and inference workflows are adapted via search-based selection and modular token attention.

4. Deployment, Adaptation, and Scalability Considerations

4.1 Post-Training Adaptation Pipelines

Extensive pretraining on behavioral sequences is followed by supervised fine-tuning (embedding-based or generative ranking SFT) and alignment via Direct Preference Optimization (DPO) (Lin et al., 22 Aug 2025). Such adaptation enables production-ready deployment with versatile outputs—user/POI embeddings for ranking, generative scores, or direct ID generation—supporting ranking, end-to-end recommendation, and large-scale recall.

4.2 Computational Complexity and Latency Management

GRUT models maintain sublinear complexity in softmax operations and cache-efficient autoregressive decoding, enabled by hierarchical indexing and search-based restrictions. Time-aware RoPE incorporates only $O(d)$ additional parameters, compatible with FlashAttention and other optimized kernels (Wei et al., 23 Oct 2025). Dual-branch and bucketed decoding parallelize inference without latency penalty (Yi et al., 16 Nov 2025).

4.3 Practical Impact in Industrial Contexts

GRUT-style architectures have demonstrated significant improvements in hit rate, diversity, and freshness of recommendations in commercial deployments. For example, DualGR achieved +0.527% video views and +0.432% watch time uplift in production A/B test, while Spacetime-GR delivered +6.0% CTR and +4.2% CVR (Yi et al., 16 Nov 2025, Lin et al., 22 Aug 2025). Empirical ablations consistently show 1–5 percentage point gains in recall@K and NDCG metrics attributable to time-awareness modules.

5. Benchmark Results and Performance Analysis

Quantitative evidence from public and industrial datasets substantiates the effectiveness of GRUT frameworks. Experimental results from SynerGen, Spacetime-GR, and DualGR highlight the following impacts:

Model/Method	Recall@K (Improvement)	NDCG@K (Improvement)	Context
SynerGen (ID)	+37% @Recall@10	---	Book Review (Gao et al., 26 Sep 2025)
SynerGen (Semantic)	+1.26pp R@1 (rec)	+0.98pp NDCG@10	eBook Search Sessions (Gao et al., 26 Sep 2025)
Spacetime-GR (SFT)	+1.86pp AUC	+2.29pp AUC	POI ranking (Lin et al., 22 Aug 2025)
DualGR (full)	HR@100 = 6.827%	HR@1000 = 19.529%	Kuaishou production (Yi et al., 16 Nov 2025)
TO-RoPE split-dim	HR@10 = 0.3406	NDCG@10 = 0.2059	MovieLens-20M (Wei et al., 23 Oct 2025)

Performance gains are consistently attributed to time-tokenization, rotary embeddings, dual-horizon modeling, and exposure-aware objectives.

6. Implications, Extensions, and Research Directions

The GRUT paradigm generalizes across domains demanding dynamic preference modeling, periodic recommendation, and real-time adaptation. Recent extensions include:

• Multi-scale and multi-frequency temporal embeddings, capturing calendar periodicities and user-specific rhythms (Wei et al., 23 Oct 2025).
• Spatiotemporal fusion for geo-sensitive recommendation, applicable to POI, ride-hailing, and real-world event streams (Lin et al., 22 Aug 2025).
• Negative sampling methods conditioned on temporal proximity, enhancing model robustness to recency and drift (Gao et al., 26 Sep 2025).
• Modular adaptation strategies for transfer learning, multi-objective optimization, and fine-tuning in low-data regimes (Lin et al., 22 Aug 2025).

A plausible implication is that as temporal signals become more granular and multimodal, GRUT frameworks will further blur boundaries between retrieval, ranking, and event forecasting, serving as deployable foundation models for unified information access in recommendation-centric platforms.