GSISR: Generative Spatiotemporal Sequence Rec.

• GSISR is a recommendation approach that generates multi-step, parameterized intent sequences under temporal, geographic, and execution constraints.
• The methodology employs autoregressive sequence modeling with hierarchical generation and rule-based grounding to ensure logical and feasible plans.
• Key implications include improved sequence coherence, enhanced context responsiveness, and scalable industrial application for complex recommendation tasks.

Generative Spatiotemporal Intent Sequence Recommendation (GSISR) is a recommendation paradigm in which the model generates an ordered sequence of parameterized intents, rather than an isolated next item, under explicit temporal, geographic, and execution constraints. In the formulation introduced for Amap homepage recommendation, the input context is $\mathcal{X}=\{u,\mathcal{H},s,\mathcal{E}\}$, comprising user profile, interaction history, spatiotemporal context, and an intent library, while the output is a structured intent sequence $\mathbf{Y}=\{a_1,a_2,\dots,a_n\}$ with $a_k=(v_k,p_k)$, where $v_k\in\mathcal{E}$ is a selected tool and $p_k$ are its parameters. The learning objective is autoregressive sequence modeling,

$$\theta^* = \arg\max_\theta \sum_{(\mathcal{X}, \mathbf{Y}) \in \mathcal{D}} \sum_{t=1}^{T} \log P(y_t \mid y_{<t}, \mathcal{X}; \theta),$$

with the additional requirement that generated plans remain logically coherent and physically executable (Wang et al., 27 May 2026).

1. Scope, task boundaries, and relation to adjacent problems

GSISR is defined around intent flows rather than isolated actions. In the Amap setting, an intent is a parameterized invocation of an Amap homepage tool, and a card is the UI surface rendered from that invocation. The task therefore differs from standard next-item recommendation, which usually predicts one next action; from trajectory recommendation, which focuses on physical location sequences; from session-based recommendation, which emphasizes short-term co-occurrence; and from generic sequence generation, which need not satisfy application-specific schemas or real-world feasibility constraints (Wang et al., 27 May 2026).

This distinction is material because many recent generative recommenders remain single-step systems. AWARE formulates next-POI recommendation as conditional generation over mobility history and world knowledge, but still predicts only the next POI (Ding et al., 12 May 2026). ROS, GeoGR, and Spacetime-GR are likewise next-POI frameworks, even when they use hierarchical POI tokenization and explicit spatiotemporal conditioning (Lv et al., 8 Jan 2026, Wang et al., 11 Feb 2026, Lin et al., 22 Aug 2025). This suggests that GSISR should be regarded not as a synonym for generative next-location prediction, but as a stricter problem in which future recommendations form a structured, interdependent plan.

The output structure is also richer than in conventional ranking systems. In GSISR, later actions may consume entities or parameters established by earlier actions, and later spatial recommendations may have to be rebased around locations implied by earlier steps. The Amap travel example—ride to hotel, hotel-related service, nearby dining, return—illustrates why independent next-item scoring can yield incoherent plans when downstream actions ignore the state changes caused by previous recommendations (Wang et al., 27 May 2026).

2. Representation of intent, space, and time

A persistent issue in GSISR-adjacent research is that intent is rarely directly labeled. Instead, it is represented indirectly through text, prototypes, semantic codes, or behavior types.

In AWARE, intent is operationalized through a generated world-knowledge narrative $w$ that is conditioned on user trajectory, region, time, and retrieved public-web evidence. The final conditioning prompt is

$$x = \phi(\mathcal{H}_u,\mathcal{F}_u,w,\mathcal{C}_u,\mathcal{T}_u),$$

where $\mathcal{H}_u$ is time-aware history, $\mathcal{F}_u$ is a frequency prior, $\mathbf{Y}=\{a_1,a_2,\dots,a_n\}$0 is current context, and $\mathbf{Y}=\{a_1,a_2,\dots,a_n\}$1 is a transition prior. The paper explicitly characterizes $\mathbf{Y}=\{a_1,a_2,\dots,a_n\}$2 as a personalized, temporally grounded side signal that functions as an intent proxy and explanation (Ding et al., 12 May 2026). This is a textual approach to intent representation.

Other work uses latent prototype mixtures. MCLRec represents a sequence by a weighted combination of top-$\mathbf{Y}=\{a_1,a_2,\dots,a_n\}$3 latent intent centroids obtained from K-means over sequence embeddings, rather than by a single intent vector (Huang et al., 2024). BIPCL extends this idea bilaterally, learning shared prototype sets on both the user and item sides and injecting them into sequence and item representations through gated fusion (Zhang et al., 3 Apr 2026). IGR-SR uses learnable $\mathbf{Y}=\{a_1,a_2,\dots,a_n\}$4 tokens appended to a frozen encoder, treating their hidden states as multi-faceted latent intents that guide downstream reasoning via cross-attention (Shao et al., 16 Dec 2025). These systems model intent as continuous latent structure rather than discrete labels.

A third line of work encodes intent indirectly in hierarchical semantic identifiers. ROS represents each POI as

$\mathbf{Y}=\{a_1,a_2,\dots,a_n\}$5

with geospatial locality $\mathbf{Y}=\{a_1,a_2,\dots,a_n\}$6, functional semantics $\mathbf{Y}=\{a_1,a_2,\dots,a_n\}$7, and a uniqueness suffix $\mathbf{Y}=\{a_1,a_2,\dots,a_n\}$8 (Lv et al., 8 Jan 2026). GeoGR likewise learns hierarchical POI SIDs through contrastive POI representation learning, residual quantization, and EM-style SID refinement, with the stated aim of encoding behavioral semantics and cross-category spatio-temporal collaborative relationships (Wang et al., 11 Feb 2026). Spacetime-GR uses a two-token hierarchy $\mathbf{Y}=\{a_1,a_2,\dots,a_n\}$9, where the block is a 5km $a_k=(v_k,p_k)$0 5km geographic region and the inner token is the local POI index within that block (Lin et al., 22 Aug 2025). In GSISR terms, these constructions are not explicit intent variables, but they create a structured state space in which intent-conditioned generation becomes tractable.

Time and space are also encoded in heterogeneous ways. Spacetime-GR is explicit: each action is represented as $a_k=(v_k,p_k)$1, with

$a_k=(v_k,p_k)$2

so spatiotemporal context becomes part of the token stream itself (Lin et al., 22 Aug 2025). STGIN, by contrast, decomposes user preference into global, spatial, temporal, spatial-temporal, and real-time components, using multi-view subgraphs and a recent-behavior attention module rather than autoregressive generation (Liu et al., 2023). This suggests that GSISR can inherit either token-based or graph-conditioned context representations, depending on whether the design prioritizes generation or retrieval.

3. Generative modeling strategies

GSISR research draws from several generative recommendation templates, but extends them toward plan generation.

A foundational autoregressive formulation appears in GPTRec: $a_k=(v_k,p_k)$3 with a decoder-only GPT-2 backbone and a Next-K strategy that generates recommendation lists item by item, conditioning each choice on previously generated recommendations (Petrov et al., 2023). Although this work is not spatiotemporal, it clarifies the difference between independent Top-K ranking and sequential list generation, which is central to GSISR.

Hierarchical generation is now common in large-vocabulary POI recommendation. Spacetime-GR pretrains a decoder-only model to generate a block token and then an inner token for the next POI: $a_k=(v_k,p_k)$4 Only interest-based actions contribute to this loss (Lin et al., 22 Aug 2025). GLASS uses a related hierarchical SID decoder, but specializes long-history modeling: SID-Tier improves the first semantic token, semantic hard search retrieves long-term history after that first token is generated, and an adaptive gate fuses short- and long-term context (Cao et al., 5 Feb 2026). GeoGR similarly frames next-POI retrieval as autoregressive SID generation, preceded by continued pre-training on template-based corpora and followed by supervised fine-tuning on next-POI instruction data (Wang et al., 11 Feb 2026).

Within GSISR proper, GPlan addresses a different bottleneck: large models can reason, but direct industrial deployment is limited by high inference latency and by context-mismatched or physically infeasible plans. Its two core components are Progressive Implicit CoT Distillation, which compresses explicit reasoning processes into reserved latent tokens, and Spatiotemporal Counterfactual DPO, which aligns the model with counterfactual context-plan pairs to improve context sensitivity (Wang et al., 27 May 2026). This is a distinct design choice: rather than generating long reasoning traces online, the model internalizes planning logic.

The broader literature suggests a division between prompt-conditioned augmentation and architectural modification. AWARE makes no architectural changes to the generative backbone and instead enriches the input prompt with behavioral priors and world knowledge (Ding et al., 12 May 2026). BITRec, by contrast, injects structured inductive bias directly into self-attention through Hierarchical Behavior Aggregation and Transition Relation Encoding, steering generation with behavior-intensity and transition-aware attention biases (Yang et al., 27 Apr 2026). For GSISR, this suggests two non-exclusive paths: enrich the prompt with grounded planning context, or encode intent-transition structure directly in the model’s state dynamics.

4. Grounding, feasibility, and alignment

A defining property of GSISR is that generated sequences must satisfy execution constraints, not merely predictive likelihood.

In GPlan, plan validity is enforced through concrete rule layers: format checks such as JSON parsability and valid plan length; schema checks ensuring that every intent comes from the fixed tool library and that required parameters and enumerated fields are valid; and logic invariants such as resolvable parameter references and bounded counts for mutually exclusive intents, including at most one ride-hailing call per sequence (Wang et al., 27 May 2026). This is a stronger requirement than ordinary recommendation correctness.

Related next-POI work has developed partial analogues of this grounding problem. ROS makes geography a first-class decision variable through a three-stage Mobility CoT—Personality Modeling, Intent Space Construction, and Locality Informed Pruning—and then aligns the model with real-world geography through spatial-guided RL. Its unified reward is

$a_k=(v_k,p_k)$5

combining format correctness, hierarchical SID correctness, and a Haversine-distance-based spatial reward (Lv et al., 8 Jan 2026). GeoGR grounds POI semantics earlier in the pipeline, by constructing geographically constrained co-visited POI pairs and optimizing a contrastive POI-to-POI loss before SID quantization (Wang et al., 11 Feb 2026). AWARE grounds external knowledge through web search, web fetch, corroboration from at least two independent sources, and explicit region/time constraints, while also filtering that knowledge through frequency, transition, and periodic priors to avoid generic city descriptions (Ding et al., 12 May 2026).

These mechanisms address different failure modes. ROS and GPlan emphasize feasibility at generation time. AWARE and GeoGR emphasize grounding in external or collaborative evidence before inference. GRASP adds a complementary robustness argument: LLM-generated semantics should be treated as soft auxiliary context, not hard supervision, because hallucination increases when histories are short; its retrieval-augmented, holistic-attention design is explicitly motivated by this concern (Dai et al., 25 Nov 2025). A plausible implication is that a mature GSISR system will need both grounded state construction and conservative use of generated reasoning artifacts.

5. Empirical landscape and industrialization

Evidence directly on GSISR is still concentrated in GPlan, which reports improved sequence coherence and context responsiveness in both offline experiments and online A/B testing (Wang et al., 27 May 2026). Its latency measurements also establish why implicit reasoning is operationally important: explicit CoT increases inference time from 299 ms to 1.06 s for a 1.7B model and from 437 ms to 1.85 s for a 4B model, whereas PICD yields 366 ms and 503 ms respectively (Wang et al., 27 May 2026).

Adjacent next-POI and sequential works provide a broader empirical backdrop for the ingredients GSISR depends on.

Work	Setting	Indicative finding
GPlan (Wang et al., 27 May 2026)	GSISR in Amap	Improves sequence coherence and context responsiveness; PICD sharply reduces CoT latency overhead
AWARE (Ding et al., 12 May 2026)	Generative next-POI	Up to 12.4% relative improvement; +14.9% on hard cases in NYC
ROS (Lv et al., 8 Jan 2026)	Geographic generative next-POI	HR@1 gains of +11.2%, +11.0%, and +15.7% over strongest baselines
GeoGR (Wang et al., 11 Feb 2026)	Spatio-temporal aware POI generation	Public-set gains around 4%–5% recall; AMAP online gains include +2.31% WinRate
Spacetime-GR (Lin et al., 22 Aug 2025)	Large-scale online POI generation	+6% CTR and +4.2% CVR in deployment
STGIN (Liu et al., 2023)	Large-scale POI retrieval	Online +1.1% CTR and +6.3% RPM

Several patterns recur. First, explicit spatiotemporal encoding matters: removing spacetime information in Spacetime-GR drops hr@1 from 0.1525 to 0.1007 and hr@100 from 0.4721 to 0.3671 (Lin et al., 22 Aug 2025). Second, grounded world or geographic knowledge helps most when local continuation is insufficient: AWARE improves hard cases more than easy ones (Ding et al., 12 May 2026), and ROS shows particularly strong cross-city transfer under compositional spatial-semantic tokenization (Lv et al., 8 Jan 2026). Third, industrial deployment favors hybrid stacks: GeoGR and Spacetime-GR both act as generative retrieval front-ends that remain compatible with downstream ranking and post-processing (Wang et al., 11 Feb 2026, Lin et al., 22 Aug 2025).

6. Limitations, misconceptions, and research directions

A common misconception is that GSISR is simply next-POI generation with extra features. The current literature indicates otherwise. AWARE, ROS, GeoGR, and Spacetime-GR are all highly relevant, but they remain primarily single-step systems (Ding et al., 12 May 2026, Lv et al., 8 Jan 2026, Wang et al., 11 Feb 2026, Lin et al., 22 Aug 2025). GSISR, as formalized in GPlan, requires the generation of multi-step parameterized intent sequences under logical and physical constraints (Wang et al., 27 May 2026).

A second misconception is that “intent” is already a well-defined supervised variable. In most current work it is not. Intent is usually implicit: a generated contextual narrative in AWARE, prototype mixtures in MCLRec and BIPCL, latent intent tokens in IGR-SR, hierarchical semantic codes in GeoGR and ROS, or behavior-intensity strata in BITRec (Ding et al., 12 May 2026, Huang et al., 2024, Zhang et al., 3 Apr 2026, Shao et al., 16 Dec 2025, Wang et al., 11 Feb 2026, Lv et al., 8 Jan 2026, Yang et al., 27 Apr 2026). This suggests that explicit intent-state supervision remains an open problem rather than a solved component.

Several research directions follow directly from the present boundary of the field. One is explicit intent-state generation: current multi-intent methods are mostly discriminative, sequence-level, or auxiliary-loss-based (Huang et al., 2024, Zhang et al., 3 Apr 2026). Another is reasoning under deployment constraints: GPlan’s PICD shows one path, but the general problem of compressing planning logic without losing constraint sensitivity remains open (Wang et al., 27 May 2026). A third is hallucination-robust grounding: GRASP argues that LLM semantics should be fused as soft context rather than imposed as supervision, especially under sparse histories (Dai et al., 25 Nov 2025). A fourth is transition-structured generation: BITRec shows that explicit intensity and transition biases improve generative recommendation, suggesting a route toward intent-stage transition models with spatiotemporal relation features (Yang et al., 27 Apr 2026). Finally, online–offline decomposition remains important for scale: STGIN demonstrates the practicality of offline context priors plus online short-term updates, a pattern that GSISR systems may need if full plan generation is to meet industrial latency targets (Liu et al., 2023).

GSISR is therefore best understood as the convergence point of several lines of work: hierarchical generative recommendation, spatiotemporal conditioning, latent intent modeling, grounding against executable constraints, and deployment-aware reasoning compression. The field is moving from next-step destination prediction toward sequence-level service planning, but the transition is incomplete. Current systems show that spatiotemporal generation is practical at industrial scale; GPlan adds the stronger claim that the generated object can be a coherent and executable intent sequence rather than a single recommendation (Wang et al., 27 May 2026).