WorksComputer Science
MILES-Rec: Uncovering Implicit Intents from Mobility Behaviors for Cross-Domain POI Recommendation
Abstract
Map applications bridge everyday mobility with local services, creating opportunities to recommend places by uncovering users' implicit intents. However, utility-driven map interactions provide only indirect and incomplete evidence of consumption demands, making it challenging to infer personalized service needs from observed behaviors. We propose MILES-Rec, a cross-domain POI recommendation framework that infers latent service intents from heterogeneous map behaviors and expresses them as prioritized natural language queries. It first learns from progressive reasoning demonstrations that synthesize user profiles and map histories for situation understanding, intent expansion, and query prioritization. We then introduce intent-aligned group relative policy optimization, using subsequent user interactions to reward query relevance and ranked service coverage while penalizing redundancy and excessive response length. A two-level credit assignment mechanism integrates shared list-level advantages with query-specific advantages, coordinating individual query quality with the complementary coverage of the complete query list. We further construct MILES-Bench from real user interaction logs on a large-scale mapping platform, supporting training and evaluation across diverse services under rich and sparse behavioral histories. Extensive experiments against general-purpose LLMs and specialized recommendation baselines demonstrate superior POI retrieval utility and strong service and semantic alignment, with consistent gains from behavioral policy optimization across four backbone scales.
Introduction
Local-service recommendation helps users discover places for dining, shopping, leisure, and other everyday activities. Modern map applications naturally encompass two distinct yet coupled functional domains: the spatial mobility domain, which records high-frequency utility actions such as destination searches, route requests, and navigation, and the local-service domain, which captures service exploration and consumption across diverse commercial categories. Because users rely on map platforms primarily as mobility tools, explicit service transactions are often sparse, whereas mobility trajectories carry rich contextual cues about companion lifestyle needs. Conventional next-POI prediction focuses on forecasting a user’s immediate spatial destination within the mobility domain [29], yet it largely overlooks the derivative service demands triggered by these trips. For instance, planning a route to a family park often entails subsequent target-domain needs for child-friendly dining or nearby parking. Bridging these two domains requires jointly interpreting heterogeneous map histories alongside user profiles to uncover latent consumption intents. This
motivates cross-domain POI recommendation from mobility behaviors to local services, which leverages abundant mobility evidence to proactively recommend relevant everyday services beyond explicitly stated user queries, as illustrated in Figure 1.

Figure 1. (a) Service recommendation; (b) next-location prediction; (c) MILES-Rec infers service intents from map behaviors and retrieves POIs using prioritized queries.
Large language models combine broad world knowledge with reasoning capabilities, offering a strong semantic foundation for linking observed mobility behaviors to implicit service needs. Expressing these latent connections in natural language provides an interpretable bridge between source-domain mobility evidence and target-domain service preferences. Recent studies have explored explicit intent modeling and query-based recommendation from various perspectives. Specifically, PIGEON [14] predicts open-set living needs to retrieve local services via embeddings, AIR [13] extracts atomic intents to transfer preferences across content and e-commerce domains, and EAGER [30] aligns supervised query generation with single-domain click preferences through reinforcement learning. Although these works demonstrate the promise of explicit intent representations and behavioral optimization, they primarily generate isolated or unranked intents from homogeneous item interactions. Consequently, a critical challenge remains in how to infer a prioritized set of complementary service intents from noisy, heterogeneous map logs and jointly optimize their list-wise target-domain utility through subsequent multi-service feedback.
Addressing this cross-domain task entails three fundamental challenges. First, contextualized multi-intent reasoning requires jointly interpreting heterogeneous mobility actions and user profiles in context. The model must distinguish user-specific service demands from generic commonsense associations, derive complementary expansion intents alongside primary needs, and establish their relative priorities. Second, behavioral supervision across domains is inherently indirect because interaction logs record only external actions rather than underlying motivations. Since multiple plausible query sets can correspond to the same mobility activity, merely imitating static reference queries via supervised learning fails to directly optimize alignment with subsequent target-domain interactions. Third, ranked query generation involves a trade-off between individual query relevance and overall list coverage. Optimizing solely for relevance often yields repetitive queries targeting a single dominant service, whereas pursuing broad coverage risks introducing weak queries. Because a single response-level reward conflates these factors across the entire output, fine-grained credit assignment is required to disentangle individual query quality from the collective utility of the complete list.
To overcome these challenges, we propose MILES-Rec, a two-stage framework that bridges the mobility domain with local-service discovery through prioritized natural-language queries. To support contextualized multi-intent reasoning, we construct progressive reasoning demonstrations that guide the model through situation understanding, intent expansion, and query prioritization via supervised fine-tuning. To learn effectively from indirect behavioral supervision, we introduce intent-aligned Group Relative Policy Optimization, refining the supervised policy with feedback from subsequent target-domain interactions. The behavioral reward jointly measures query-to-service relevance and one-to-one ranked service coverage while penalizing redundancy and excessive length. Furthermore, we design a two-level credit assignment mechanism to coordinate individual relevance with list coverage. A global list-level advantage guides the entire response including reasoning and structure, whereas a local query-specific advantage is selectively applied to the corresponding query-text tokens. At inference time, MILES-Rec generates prioritized service queries from user profiles and map histories to retrieve relevant local-service POIs.
We further construct MILES-Bench from real-world user profiles and heterogeneous interaction logs on a large-scale mapping platform. It provides 19,006 verified reasoning demonstrations for supervised fine-tuning, 3,000 interaction-grounded samples for reinforcement learning, and two user-disjoint test sets covering rich and sparse histories across ten cities. To assess both intermediate intent reasoning and downstream retrieval, we establish a unified evaluation protocol spanning POI Retrieval Utility over a two-million-POI pool, Service Intent Alignment, and Semantic Demand Alignment. Together, these dimensions evaluate retrieved POI utility alongside the category relevance, constraint compatibility, and service coverage of the generated natural-language intent queries.
Our main contributions are summarized as follows:
We formulate cross-domain POI recommendation from the spatial mobility domain to the local-service domain as an intent-based reasoning and prioritized query generation task over heterogeneous map histories and profiles.
We propose MILES-Rec, which integrates progressive reasoning supervision with intent-aligned reinforcement learning and employs a two-level credit assignment mechanism to jointly optimize individual query relevance and list-wise service coverage while suppressing redundancy.
We construct MILES-Bench from real-world map logs covering both rich and sparse histories, paired with a three-dimensional evaluation protocol linking POI retrieval utility with service and semantic alignment.
Extensive experiments across four backbone scales demonstrate consistent gains over supervised fine-tuning and strong improvements over state-of-the-art baselines, supported by comprehensive ablations, human evaluations, and retriever-transfer analyses.
Related Work
LLM-based Recommendation
Large language models provide a strong semantic foundation for modeling user behaviors and generating personalized recommendations. Early works unify recommendation tasks within text-to-text frameworks as in P5 [11] or adapt language models via instruction tuning as in TALLRec [2], while LLaRA [16] and ReLLa [17] further incorporate collaborative embeddings and retrieved historical interactions to mitigate long-sequence noise. Moving beyond supervised adaptation, recent studies optimize generation with downstream feedback via reinforcement learning, including Rec-R1 with fixed-recommender rewards [18], EAGER with click-preference alignment [30], and ReList with multi-objective query optimization [4]. However, these reinforcement learning methods typically rely on coarse response-level rewards over homogeneous item clicks. In contrast, MILES-Rec infers implicit service intents from heterogeneous map logs and user profiles, introducing a two-level credit assignment mechanism within group relative policy optimization (GRPO) to jointly optimize individual query relevance and list-wise service coverage without redundant intent generation.
Location-based Recommendation
Location-based recommender systems leverage user preferences and spatiotemporal contexts to suggest relevant places and services. One mainstream direction focuses on next-POI prediction within the mobility domain via sequential trajectory modeling [29] or language-model-based spatiotemporal reasoning [10, 21, 31]. Because trajectory forecasting alone cannot capture companion consumption demands, a complementary line of research explicitly models living needs for local-service discovery, evolving from predefined service categories in NEON [15] to open-set need descriptions in PIGEON [14], hierarchical agentic service reasoning [5], and spatiotemporally constrained intent sequences [25]. Although these studies advance explicit demand modeling, they primarily predict isolated needs or next spatial destinations from homogeneous check-in sequences. Our work instead jointly interprets heterogeneous mobility interactions alongside user profiles, uncovering both primary needs and complementary expansion intents as a prioritized query list for personalized multi-service POI discovery.
Cross-Domain Recommendation
Cross-domain recommendation transfers behavioral knowledge across domains to alleviate data sparsity and improve preference modeling. Recent approaches leverage language models to capture semantic relationships beyond item-identifier overlaps, such as combining unified semantic representations with hierarchical profiling in LLM4CDSR [19] or generating synthetic cross-domain interactions in LLMCDSR [28]. Closer to our formulation, AIR extracts atomic intents from content browsing to support e-commerce recommendation [13]. While these methods transfer preferences across item-consumption domains via latent embeddings or unranked atomic tags, they do not address the functional gap between tool-oriented mobility behaviors and commercial service consumption. MILES-Rec bridges the spatial mobility domain with the local-service domain by reasoning over map histories and user profiles to formulate prioritized queries for hybrid spatial retrieval.
Problem Formulation
We study cross-domain POI recommendation from the spatial mobility domain to the local-service domain. The source context is dominated by high-frequency mobility interactions alongside user profiles, while the target space encompasses local-service POIs across commercial categories. Because utility-oriented mobility actions and service POIs lack direct item-level overlap, prioritized natural-language queries serve as an explicit semantic bridge connecting behavioral observations to personalized service retrieval.
For each user , we group the observed profile attributes as
Here, contains age group, occupation, and city of residence; captures marital status, parenthood, life stage, children’s age groups, and pet ownership; and specifies car ownership, brand, and model series. Each group includes observed attributes.
The chronological interaction history observed before the prediction time is , where is the number of historical events and may be zero for cold-start users. Each event records a timestamp, an action type, and textual content. Historical actions are dominated by source-domain mobility behaviors such as destination searches, detail views, saves, route planning, and navigation alongside sparse historical service bookings, while the textual content includes search queries and place descriptions. Given the spatiotemporal context with available city context , the complete input state is defined as
Given , a policy model reasons over the user’s profile, mobility trajectory, and context to infer implicit service needs, expressing them as an ordered query list:
where is the number of generated queries. Each query expresses a distinct service intent and contextual constraints in natural language, and its position indicates its priority. The generator aims to identify primary intents grounded in recent mobility actions and uncover complementary expansion intents for unseen services.
Given a shared pool of local-service POIs , ranking candidates by relevance to yields the final recommendation:
where is the number of returned POIs and . Target-domain interactions observed at are excluded from and provide behavioral feedback for policy optimization and evaluation.
Method
Overview
Figure 2 illustrates the architecture of MILES-Rec, which learns cross-domain POI recommendation from the spatial mobility domain to the local-service domain via a two-stage paradigm. In Section 4.2, we construct progressive reasoning demonstrations covering situation understanding, intent expansion, and query prioritization to initialize the policy through supervised fine-tuning. In Section 4.3, we refine the policy with subsequent target-domain interactions via Group Relative Policy Optimization, introducing two-level credit assignment to jointly optimize individual query relevance and list-wise service coverage.

Figure 2. Overview of MILES-Rec: reasoning-supervised SFT and intent-aligned GRPO. Right: MILES-Bench service taxonomy and history distributions.
Progressive Reasoning Supervision
Directly generating target-domain service queries from mobility logs and user profiles is challenging because mobility actions primarily reflect movement rather than explicit consumption requests. We therefore construct progressive chain-of-thought demonstrations [27] that decompose intent inference into three stages before producing prioritized queries for personalized POI retrieval.
Three-Stage Reasoning Structure. Situation understanding jointly interprets profile attributes, chronological interactions, and prediction context to infer the user’s underlying activity or circumstances, citing supporting events. Intent expansion identifies primary service needs grounded in observed behaviors and derives complementary, alternative, or subsequent intents supported by the context, emphasizing distinct service purposes rather than synonymous queries. Query prioritization considers evidence strength, recency, and recurring behavioral patterns to organize up to ten concise queries. Primary queries precede related extensions, and the final list retains a primary or expansion label for each generated service query.
Demonstration Synthesis and Verification. We organize user-level platform records into a structured context containing available profile attributes, prediction context, and chronological interactions. Each event retains its identifier, timestamp, action type, and original content. We remove missing-value placeholders and duplicated contextual fields while preserving repeated interactions and their temporal order. This reduces input redundancy and supports event-based evidence citation. Given , Qwen-235B generates the three-stage reasoning and prioritized queries in a single response.
Building on model-generated supervision and AI feedback [9, 26], we refine these demonstrations using GPT-5.1 and DeepSeek-Pro feedback alongside human review and revision. The review examines evidence fidelity, reasoning consistency, and intent distinctness, guiding corrections to unsupported attributes, inconsistent inferences, and redundant queries. The resulting dataset contains 19,006 user-level demonstrations, each represented as , where contains relevant profile facts, cited events, and reasoning, and contains ordered service queries and their priority labels.
Supervised Fine-Tuning Objective. We fine-tune the policy on complete target responses using token-level negative log-likelihood:
where denotes the training users, is the target response length, and . The loss covers reasoning, queries, priority labels, and output structure, excluding all input tokens.
This supervision initializes a policy that connects mobility evidence with prioritized service intents. Subsequent RL explores alternative intent combinations and query priorities, using later user interactions to refine query relevance and service coverage.
Intent-Aligned GRPO
We use subsequent target-domain interactions to align generated queries with observed service needs. The behavioral reward combines individual query relevance with ranked service coverage, balancing intent quality and the breadth of the query list. Two-level credit assignment combines shared response-level guidance with additional local feedback for individual service queries.
Starting from the SFT initialization, we follow GRPO [23] and sample responses from the rollout policy . Each response contains a reasoning trace and an ordered query list . Subsequent interactions are used solely to compute rewards and are excluded from the policy input.
Behavioral Reward Construction. To connect open-ended intents with observable target-domain interactions, we adopt 263 fine-grained service categories organized into 12 broad groups covering dining, shopping, hotels, leisure and entertainment, transportation, automotive services, and other local services. We group POIs from subsequent interactions by service category into , with category descriptions and observed POI texts retained for each group. These groups provide common coverage targets for distinct POIs offering the same target-domain service.
We fine-tune BGE-reranker-v2-m3 [3] on Qwen-235B-annotated query–category pairs to assess service alignment across three relations, namely match, compatible, and mismatch. A pretrained Qwen3-Embedding-8B [32] encoder additionally measures query–POI semantic similarity. Both models remain frozen during reinforcement learning; annotation and matcher-training details appear in Appendix B. For query and service group , we compute
Here, and denote the predicted probabilities of the match and compatible relations, , and clips values to . The product favors queries supported by both service agreement and semantic similarity to observed POI texts.
We reward complementary coverage through maximum-weight one-to-one matching between queries and service groups. Let indicate whether repeats an earlier query, and let contain all feasible partial matchings. We define ranked coverage as
Each query matches at most one service group, and each group contributes at most once. The rank discount encourages placing well-supported intents earlier, while the one-to-one constraint prevents repeatedly rewarding the same target service.
For valid responses, we also penalize redundancy and excessive response length, yielding the following list-level reward:
where and . Repeated queries incur an explicit penalty, while applies a bounded linear penalty beyond a free response-length budget, counting both reasoning and queries. Invalid or truncated responses receive zero reward. Appendix B specifies the validity criteria and penalty parameters.
Two-Level Credit Assignment. A list-level reward evaluates the complete output but does not distinguish the relevance of individual queries within it [34]. We therefore derive a local relevance score for each query from the same service-matching matrix:
This score measures a query’s best service agreement without one-to-one matching or rank discounting, complementing list-level coverage with an assessment of individual query relevance.
We normalize each signal within the user’s response group:
Here, and are computed over all response rewards. For local statistics, let index valid responses. We compute the weighted mean and standard deviation using weights so that every valid response contributes equal total weight, preventing longer query lists from dominating the local baseline. The constants and stabilize the normalization of both signals.
Letting denote the token positions corresponding to the text of , we combine the list-level and query-level advantages as
where controls the local contribution. All response tokens receive the list-level advantage, whereas query-text tokens additionally receive their corresponding local advantage. Reasoning, priority labels, and structural tokens retain only the list-level signal, and invalid responses receive no local contribution. This assignment preserves overall list optimization while allowing queries within the same response to receive distinct learning signals.
Policy Optimization. We optimize the policy using the two-level advantages, enabling query-specific updates within the shared list-level objective while KL regularization constrains deviation from the frozen SFT reference policy. Writing , the probability ratio between the current and rollout policies is
The policy is optimized using the following training loss:
where is the response length, is the dual-clipped surrogate, and is the sampled KL penalty relative to the frozen SFT reference policy , weighted by . The expectation is taken over training users and sampled responses. The rollout policy is refreshed from the updated policy, while rewards and advantages remain fixed during each update, as detailed in Appendix B.
By coordinating query relevance, service coverage, and response length, GRPO turns target-domain behavioral feedback into targeted improvements in personalized service query generation.
Experiments
Experimental Setup
Datasets. We construct MILES-Bench from real user profiles and heterogeneous map interactions on a large-scale mapping platform. Following Section 4.2, we obtain 19,006 user-level reasoning demonstrations for SFT, while RL uses 3,000 separate users with subsequent target-domain interaction feedback and another 500 users for validation. We evaluate on user-disjoint MapRec-Rich and MapRec-Sparse, each containing 1,000 users across the same ten cities with rich and naturally sparse mobility histories, respectively. Subsequent target-domain interactions are held out as evaluation targets. Figure 2 illustrates the service taxonomy and history distributions, Table 1 summarizes the dataset statistics, and Appendix A provides further construction details.
| Statistic | Training | Evaluation | ||
| SFT | RL | Rich | Sparse | |
| Users | 19,006 | 3,000 | 1,000 | 1,000 |
| History events | 4,383,201 | 704,164 | 135,793 | 5,826 |
| Events/user | 230.62 | 234.72 | 135.79 | 5.83 |
| Action types/user | 5.33 | 5.37 | 3.07 | 1.63 |
| Profile fields/user | 6.65 | 7.19 | 6.63 | 5.65 |
| Queries/targets | 64,378 | 43,920 | 11,335 | 2,876 |
Table 1. Statistics of MILES-Bench. Per-user counts are means. The last row counts SFT queries or user–POI target pairs.
Baselines. Table 2 compares general-purpose LLMs with Base, SFT, and GRPO variants of Qwen3.5 at the 4B, 9B, 27B, and 122B-A10B scales. For each backbone, + SFT and + GRPO denote successive training stages, with + GRPO representing the complete MILES-Rec model. Table 3 adds cross-domain adaptations of GLoSS [1], Rec-R1 [18], and PIGEON [14] built on Qwen3.5-4B, where PIGEON-history uses historical service context and context proxies to generate a single query per user. We also evaluate Content-ComiRec-SA, a content-based adaptation of ComiRec-SA [6] that replaces item-ID embeddings with projected POI content embeddings to rank POIs directly by multi-interest matching scores, while SIA and SDA apply only to query-generating methods. Finally, Table 6 includes task-adapted Rank-GRPO [34] as a policy-optimization baseline.
| Model / Stage | MapRec-Rich Retrieval utility Hit@50 | Recall@50 | NDCG@50 | Service SIA@10 | Semantic SDA | MapRec-Sparse Retrieval utility Hit@50 | Recall@50 | NDCG@50 | Service SIA@10 | Semantic SDA |
| MapRec-Rich | MapRec-Sparse | |||||||||
| Retrieval utility | Service | Semantic | Retrieval utility | Service | Semantic | |||||
| Hit@50 | Recall@50 | NDCG@50 | SIA@10 | SDA | Hit@50 | Recall@50 | NDCG@50 | SIA@10 | SDA | |
| API-evaluated models | ||||||||||
| GPT-4o | 1.70 | 0.164 | 0.178 | 31.19 | 35.28 | 0.50 | 0.197 | 0.109 | 13.78 | 15.02 |
| DeepSeek-v4-pro | 1.30 | 0.157 | 0.158 | 28.94 | 30.89 | 0.50 | 0.277 | 0.201 | 12.46 | 12.82 |
| GLM-5.2 | 1.50 | 0.165 | 0.150 | 29.85 | 31.75 | 0.50 | 0.274 | 0.197 | 13.31 | 14.42 |
| Qwen3.8-Max | 1.70 | 0.206 | 0.205 | 32.95 | 36.08 | 0.70 | 0.308 | 0.178 | 14.43 | 15.45 |
| MiniMax-M2.7 | 1.30 | 0.145 | 0.101 | 27.51 | 29.42 | 0.40 | 0.177 | 0.114 | 12.00 | 13.33 |
| Locally evaluated models | ||||||||||
| GPT-OSS-120B | 3.00 | 0.350 | 0.329 | 25.53 | 22.59 | 0.80 | 0.337 | 0.254 | 10.84 | 12.15 |
| Llama3.1-8B-Instruct | 0.60 | 0.058 | 0.057 | 12.17 | 13.39 | 0.00 | 0.000 | 0.000 | 2.39 | 2.80 |
| Llama3.1-70B-Instruct | 0.90 | 0.156 | 0.095 | 28.48 | 37.39 | 0.70 | 0.261 | 0.182 | 12.04 | 14.06 |
| Qwen3-235B-A22B | 2.10 | 0.219 | 0.198 | 34.52 | 30.48 | 0.70 | 0.317 | 0.281 | 14.25 | 13.10 |
| Qwen3.5-397B-A17B | 0.90 | 0.110 | 0.079 | 30.30 | 33.52 | 0.40 | 0.063 | 0.079 | 12.62 | 15.22 |
| Backbone: Qwen3.5-4B | ||||||||||
| Base | 1.30 | 0.194 | 0.131 | 21.35 | 20.99 | 0.40 | 0.192 | 0.108 | 8.90 | 9.55 |
| + SFT | 1.50 | 0.159 | 0.171 | 32.50 | 34.97 | 0.50 | 0.269 | 0.276 | 14.41 | 15.77 |
| + GRPO | 3.40 | 0.338 | 0.325 | 39.93 | 51.37 | 0.80 | 0.342 | 0.282 | 18.20 | 25.80 |
| Backbone: Qwen3.5-9B | ||||||||||
| Base | 0.90 | 0.134 | 0.102 | 24.21 | 27.26 | 0.30 | 0.146 | 0.143 | 11.05 | 12.60 |
| + SFT | 1.30 | 0.134 | 0.143 | 33.38 | 35.61 | 0.40 | 0.083 | 0.085 | 14.02 | 15.03 |
| + GRPO | 2.60 | 0.273 | 0.252 | 40.01 | 52.60 | 0.40 | 0.177 | 0.195 | 18.42 | 27.67 |
| Backbone: Qwen3.5-27B | ||||||||||
| Base | 1.50 | 0.166 | 0.160 | 31.79 | 31.53 | 0.50 | 0.300 | 0.131 | 12.69 | 14.70 |
| + SFT | 1.70 | 0.168 | 0.178 | 33.28 | 33.87 | 0.50 | 0.277 | 0.203 | 14.01 | 15.57 |
| + GRPO | 3.30 | 0.381 | 0.297 | 42.45 | 72.19 | 0.60 | 0.293 | 0.213 | 21.35 | 43.43 |
| Backbone: Qwen3.5-122B-A10B | ||||||||||
| Base | 1.40 | 0.205 | 0.190 | 28.67 | 29.93 | 0.50 | 0.274 | 0.201 | 12.29 | 14.07 |
| + SFT | 2.00 | 0.222 | 0.211 | 32.81 | 33.79 | 0.60 | 0.203 | 0.141 | 14.07 | 14.51 |
| + GRPO | 3.10 | 0.331 | 0.258 | 41.32 | 61.51 | 0.70 | 0.295 | 0.192 | 19.63 | 34.35 |
Table 2. Query-generation results on MapRec-Rich and MapRec-Sparse (%). POI retrieval uses spatial hybrid retrieval over a shared pool of 2 million POIs. All 1,000 users and all valid queries per benchmark are evaluated. Bold marks column maxima.
| Method | Hit @50 | Recall @50 | NDCG @50 | SIA @10 | SDA |
| SFT | 1.50 | 0.159 | 0.171 | 32.50 | 34.97 |
| GLoSS | 1.90 | 0.218 | 0.211 | 28.03 | 17.23 |
| Rec-R1 | 1.70 | 0.167 | 0.121 | 25.61 | 50.40 |
| PIGEON-history | 0.30 | 0.049 | 0.044 | 21.56 | 44.94 |
| Content-ComiRec-SA | 1.80 | 0.202 | 0.212 | — | — |
| MILES-Rec | 3.40 | 0.338 | 0.325 | 39.93 | 51.37 |
| MapRec-Sparse | |||||
| SFT | 0.50 | 0.269 | 0.276 | 14.41 | 15.77 |
| GLoSS | 0.30 | 0.157 | 0.091 | 13.91 | 6.61 |
| Rec-R1 | 0.60 | 0.318 | 0.266 | 12.27 | 24.85 |
| PIGEON-history | 0.00 | 0.000 | 0.000 | 12.56 | 22.57 |
| Content-ComiRec-SA | 0.70 | 0.226 | 0.183 | — | — |
| MILES-Rec | 0.80 | 0.342 | 0.282 | 18.20 | 25.80 |
Table 3. Recommendation baselines (%). Query-based methods use Qwen3.5-4B; — denotes inapplicable query metrics.
| Variant | NDCG @50 | SIA @10 | SDA | NDCG @50 | SIA @10 | SDA | |
| MapRec-Rich | MapRec-Sparse | ||||||
| Rank-GRPO | — | 0.098 | 33.55 | 61.45 | 0.147 | 17.82 | 35.29 |
| List only | 0 | 0.317 | 37.31 | 37.96 | 0.217 | 16.68 | 20.16 |
| Shared credit | 0.5 | 0.295 | 37.59 | 41.04 | 0.234 | 15.68 | 21.37 |
| Full method | 0.5 | 0.325 | 39.93 | 51.37 | 0.282 | 18.20 | 25.80 |
| Relaxed | 0 | 0.262 | 37.57 | 42.08 | 0.288 | 16.45 | 20.94 |
| Relaxed | 0.5 | 0.338 | 39.93 | 52.18 | 0.208 | 17.51 | 24.49 |
Table 6. GRPO variants and component ablations with Qwen3.5-4B (%). Bold marks column maxima.
Evaluation Metrics. We evaluate performance across three complementary dimensions. For POI retrieval utility, we report Hit@50, Recall@50, and NDCG@50 [12] over a shared two-million-POI pool while retaining all future targets, including those outside the pool. Query-based methods share a spatial hybrid retrieval backend in which semantic and keyword ranks are fused and geographically weighted, with each POI receiving its maximum score across valid queries, whereas Content-ComiRec-SA retains its native multi-interest ranking. For service intent alignment, SIA@10 measures category-level NDCG@10 over 263 service categories ranked by a frozen service matcher, using categories observed in future interactions as relevance labels. For semantic demand alignment, a frozen GPT-OSS-120B evaluator assesses service matching, constraint compatibility, specific-service matching, and target coverage, aggregated as
where is the number of test users. The first three components assess individual queries, while coverage evaluates the complete query list against future targets. All metrics are macro-averaged over all 1,000 users in each test set and reported as percentages, with full definitions and evaluation details in Appendix C.
Implementation Details. We train MILES-Rec across 64 NVIDIA H20 GPUs using LlamaFactory [33] for SFT and verl [24] for GRPO, with detailed training configurations and hyperparameters provided in Appendix B.
Main Results
Table 2 compares MILES-Rec with general-purpose LLMs under a shared spatial hybrid retrieval backend over two million POIs.
Overall performance. MILES-Rec combines strong target-domain POI retrieval with superior service and semantic alignment. Its 27B variant achieves the highest SIA@10 and SDA on both benchmarks alongside the highest Recall@50 on MapRec-Rich. Meanwhile, the 4B variant achieves the highest Hit@50 on MapRec-Rich and Recall@50 on MapRec-Sparse, while maintaining NDCG@50 competitive with the strongest general-purpose baselines. These results demonstrate that intent-guided query generation effectively bridges heterogeneous mobility behaviors with both concrete POI retrieval and the local-service demands reflected in subsequent target-domain interactions.
Effectiveness of two-stage training. Supervised fine-tuning initializes structured cross-domain intent reasoning, whereas GRPO further refines its recommendation utility using subsequent target-domain behavioral feedback. Across all four backbone scales and both benchmarks, GRPO consistently improves Recall@50, NDCG@50, SIA@10, and SDA over the corresponding SFT models, while Hit@50 improves in seven settings and remains tied in one. For the 4B backbone on MapRec-Rich, Hit@50 more than doubles from 1.50% to 3.40%, and NDCG@50 increases from 0.171% to 0.325%. These consistent gains support reinforcement learning with target-domain feedback for aligning generated intents with the service needs reflected in subsequent user interactions.
Performance across model scales. Peak performance across metrics is distributed across model sizes, as the 4B variant leads several POI retrieval metrics while the 27B variant delivers the strongest service and semantic alignment. Scaling up backbone parameters alone therefore does not uniformly improve cross-domain recommendation quality. Instead, the substantial gains from GRPO across all four scales confirm the need to adapt general language capabilities to mobility-grounded intent inference and query generation.
Comparison with recommendation methods. Table 3 compares MILES-Rec with domain-adapted query-based and direct multi-interest recommendation baselines. MILES-Rec achieves the best performance across all three retrieval metrics on both benchmarks and leads both alignment metrics among query-generating methods. Specifically, its NDCG@50 reaches 0.325% on MapRec-Rich and 0.282% on MapRec-Sparse, substantially outperforming Content-ComiRec-SA at 0.212% and 0.183%, respectively. Although PIGEON-history attains SDA scores of 44.94% on MapRec-Rich and 22.57% on MapRec-Sparse, it yields markedly lower POI retrieval accuracy, indicating that service-level semantic alignment alone is insufficient to recover specific target POIs. Together with the consistent improvements over GLoSS and Rec-R1, these results validate our cross-domain intent reasoning and two-level credit assignment.
Ablation Studies
We use Qwen3.5-4B to examine input evidence, supervision formats, and GRPO components under the main evaluation protocol.
Input evidence. Personalized intent inference connects user characteristics with heterogeneous mobility evidence. Table 4 compares independently trained models with the same input intervention during SFT and evaluation. All variants retain the original teacher targets. Removing interaction history reduces NDCG@50 to 0.000% on both benchmarks while substantially lowering service and semantic alignment, underscoring the importance of mobility context for retrieving relevant target-domain POIs. Removing user profiles yields mixed results on MapRec-Rich but lowers all three metrics on MapRec-Sparse, showing that profile attributes provide complementary personalization when behavioral evidence is sparse. Retaining the full sixty-day history also outperforms the seven-day variant across all three metrics on both benchmarks, supporting the use of broader temporal mobility context.
| Input | MapRec-Rich NDCG@50 | MapRec-Rich SIA@10 | MapRec-Rich SDA | MapRec-Sparse NDCG@50 | MapRec-Sparse SIA@10 | MapRec-Sparse SDA |
| Full input | 0.171 | 32.50 | 34.97 | 0.276 | 14.41 | 15.77 |
| No profile | 0.168 | 31.80 | 35.51 | 0.185 | 12.93 | 13.35 |
| No history | 0.000 | 15.24 | 19.15 | 0.000 | 7.77 | 8.32 |
| 7-day history | 0.041 | 23.90 | 26.61 | 0.147 | 5.90 | 5.63 |
Table 4. SFT input ablations with Qwen3.5-4B (%). Bold marks column maxima.
SFT supervision. Structured supervision connects raw mobility evidence, intermediate intent reasoning, and prioritized query generation. Table 5 compares three supervision formats trained on the same 19,006 users for the same number of optimization steps. Full SFT retains evidence restatement, three-stage intent analysis, and prioritized queries, whereas Direct-SFT retains only queries and Analysis-SFT removes evidence restatement. Full SFT achieves the highest NDCG@50 and SIA@10 on both benchmarks and the highest SDA on MapRec-Rich, while Direct-SFT attains a slightly higher SDA on MapRec-Sparse. These results support retaining both evidence grounding and intermediate reasoning to connect mobility understanding with retrieval-oriented query generation, establishing a task-adapted policy for reinforcement learning.
| Supervision | MapRec-Rich NDCG@50 | MapRec-Rich SIA@10 | MapRec-Rich SDA | MapRec-Sparse NDCG@50 | MapRec-Sparse SIA@10 | MapRec-Sparse SDA |
| Full SFT | 0.171 | 32.50 | 34.97 | 0.276 | 14.41 | 15.77 |
| Direct-SFT | 0.137 | 31.69 | 31.41 | 0.191 | 14.35 | 15.95 |
| Analysis-SFT | 0.128 | 30.33 | 32.70 | 0.204 | 14.14 | 14.55 |
Table 5. SFT supervision ablations with Qwen3.5-4B (%). Bold marks column maxima.
GRPO components. Table 6 examines credit allocation and coverage matching in GRPO, with controlling local credit in our internal ablations. Under one-to-one matching, the full method improves all three metrics on both benchmarks over the list-only variant with and the shared-credit variant. Shared credit preserves the token-weighted sum of local advantages by broadcasting their token-weighted mean across all query-text tokens. This comparison supports assigning local feedback to the corresponding query under shared list-level coverage guidance. A context-blinded evaluation also yields higher context-supported coverage estimates for Full than Shared with complete outputs under both LLM judges (Table 10). Compared with task-adapted Rank-GRPO [34], the full method achieves higher NDCG@50 and SIA@10 on both benchmarks despite lower SDA, showing that stronger semantic assessment scores do not necessarily translate into better retrieval of specific target POIs. With query-specific local credit, replacing one-to-one matching with relaxed matching slightly improves NDCG@50 and SDA on MapRec-Rich but lowers all three metrics on MapRec-Sparse. Among internal ablations on MapRec-Sparse, the full method leads SIA@10 and SDA, whereas relaxed matching without local credit leads NDCG@50.
| Queries | Dataset | DeepSeek-v4-pro Full | DeepSeek-v4-pro Shared | DeepSeek-v4-pro | GPT-4o |
| First five | Rich | 50.15 | 49.55 | +0.60 | +0.69 |
| Sparse | 32.71 | 31.97 | +0.74 | +3.64 | |
| All | Rich | 64.37 | 63.66 | +0.71 | +0.63 |
| Sparse | 42.71 | 39.12 | +3.59 | +5.55 |
Table 10. Context-supported coverage (%). : Full minus Shared in percentage points. DeepSeek differences use displayed scores.
Further Analysis
We analyze query budget, query expression, and human preference using Qwen3.5-4B outputs under the main evaluation protocol.
Query budget. Figure 3 compares prefixes of one, three, or five queries against complete query lists. SFT gains largely saturate within three queries, whereas GRPO continues to benefit from later queries. On MapRec-Rich, GRPO leads at every budget, with NDCG@50 rising from 0.235% at three queries to 0.325% with complete lists. On MapRec-Sparse, SFT leads at one, three, and five queries, whereas GRPO surpasses SFT with complete lists at 0.282% versus 0.276%, indicating greater retrieval gains from complete lists.

Figure 3. POI retrieval across query budgets. All denotes the complete query list.
Query expression. Figure 4 replaces each generated query with its top coarse service category predicted by the frozen matcher while preserving query slots and the retrieval pipeline. NDCG@50 drops sharply for both SFT and GRPO on both benchmarks, falling for GRPO from 0.325% to 0.004% on MapRec-Rich and from 0.282% to 0.000% on MapRec-Sparse. Whole category mapping also affects this comparison, the results support retaining descriptive query content beyond coarse service labels.

Figure 4. POI retrieval with natural-language queries versus mapped coarse service category labels.
Human assessment. In side-by-side human evaluation on MapRec-Sparse in Table 9, annotators consistently favor GRPO over both Base and SFT with preference rates of 70% to 80%. Appendix § 11.§ 3 presents a qualitative case in Figure 7.
| Assessor | Comparator | Win | Tie | Loss | Pref. (%) |
| 1 | 4B Base | 61 | 18 | 21 | 70.00 |
| 1 | 4B SFT | 72 | 16 | 12 | 80.00 |
| 2 | 4B Base | 65 | 12 | 23 | 71.00 |
| 2 | 4B SFT | 64 | 18 | 18 | 73.00 |
Table 9. Human preferences for 4B GRPO on MapRec-Sparse. Each row contains 100 comparisons.
Figure 7. A qualitative comparison of MILES-Rec and Base. We show 12 of 116 historical interactions and three selected queries per model, retaining their original query indices.
Conclusion
We presented MILES-Rec for intent-based cross-domain POI recommendation, bridging heterogeneous mobility behaviors and user profiles with personalized local service retrieval by combining progressive reasoning supervision with intent-aligned GRPO. Leveraging subsequent behavioral feedback from the target domain, our two-level credit assignment mechanism jointly optimizes individual query relevance and collective service coverage while suppressing redundancy. Extensive experiments on MILES-Bench across four backbone scales demonstrate consistent gains in both concrete POI retrieval utility and semantic demand alignment across rich and sparse mobility histories, underscoring the value of implicit intent inference in connecting everyday mobility traces with downstream local service discovery.
References
A MILES-Bench Details
Sections 3 and 4.2 define disjoint user splits, pre-prediction inputs, and future targets deduplicated by POI per user. RL validation contains 500 users and 7,507 user–POI pairs. MapRec-Rich includes 100 users each from Beijing, Shanghai, Guangzhou, Shenzhen, Hangzhou, Nanjing, Wuhan, Chengdu, Chongqing, and Xi’an. MapRec-Sparse includes 200 users per history-length band of 0, 1–2, 3–5, 6–10, or 11–20 events, retaining available profiles and city context. Figure 2 shows 263 fine categories within twelve broad groups. Table 7 provides additional statistics.
| Statistic | MapRec-Rich | MapRec-Sparse |
| Median history events/user | 104.5 | 4.0 |
| 90th-percentile history events | 265 | 15 |
| History-length range | 7–1,002 | 0–20 |
| Distinct target POIs | 10,872 | 2,832 |
| Observed fine categories | 208 | 164 |
| Mean broad categories/user | 3.95 | 1.71 |
| Multi-service users (%) | 94.10 | 40.20 |
| History-unseen target pairs (%) | 97.27 | 97.18 |
Table 7. Additional benchmark characteristics. Multi-service users have targets in at least two broad categories.
Privacy protection. Mobility logs, vehicle attributes, and family information are de-identified before dataset construction, training, and evaluation. We will release a de-identified version of MILES-Bench and our code upon acceptance.
Implementation Details
Demonstrations and Service Matcher
Qwen-235B generates three-stage demonstrations with 1–10 queries labeled primary or expansion, ordered primary-first. Checks cover required fields, nonempty analysis and queries, labels, duplicates, and event citations. GPT-5.1, DeepSeek-Pro, and human review assess evidence fidelity, temporal consistency, reasoning coherence, and intent distinctness. SFT supervises assistant tokens across 19,006 retained demonstrations containing user profile facts, cited events, staged reasoning, and prioritized queries.
Qwen-235B labels generated queries and historical searches paired with service definitions as match, compatible, or mismatch, excluding uncertain cases. User- and normalized-query-disjoint splits contain 621,949 training, 39,737 validation, and 38,545 test pairs. We fully fine-tune BGE-reranker-v2-m3 [3] with a three-class head using cross-entropy and AdamW for three epochs, with batch size 512, length 256, weight decay 0.01, 3% warmup, and linear decay. Learning rates are for pretrained parameters and for the head. The best validation macro-F1 checkpoint and Qwen3-Embedding-8B [32] remain frozen during GRPO.
Reward and Optimization Details
SFT freezes the vision encoder and projector, training all language-model parameters with AdamW [20], BF16, cosine decay, two epochs, batch size 64, and length 98,304. Learning rates are for 4B and 9B, and for 27B and 122B-A10B. GRPO uses 125 updates at , sampling 24 users and eight responses per user. We set temperature to 1.0, top- to 0.95, , , and input and output limits to 49,152 and 8,192.
Targets are grouped by service-category path for one-to-one matching in § 4.3. Broad-only groups include all child categories and are omitted if that user has a fine-category target under the same parent. A later query is redundant if normalized texts match, character-bigram Jaccard similarity is , or cosine similarity is with the same best service group and no score improvement. Redundant queries contribute zero coverage, incur a 0.05 list penalty, and receive local score . Total response length incurs
Invalid or truncated responses receive zero reward and no local correction. List advantages cover all response tokens, while local advantages cover corresponding query text. Local normalization weights valid responses equally, with both stabilizers . (13) uses dual clipping with and :
The sampled KL term is
The reference policy remains fixed at the SFT model, while rollouts follow the updated policy. Group rewards and token-level advantages remain fixed within each policy update.
GRPO training dynamics. Figure 5 shows overall upward trends in training rewards across all four backbones over 125 updates, with the 27B variant achieving the highest rewards at later steps.
Figure 5. GRPO training rewards across four backbone sizes.
Near-line deployment. We are developing asynchronous intent inference triggered by searches and navigation, caching queries and embeddings for online POI retrieval and ranking.
Evaluation Details
Metrics are macro-averaged over all 1,000 users and reported as percentages. Query-based methods share frozen retrieval and evaluation models. Content-ComiRec-SA uses native POI scores for retrieval utility only. Empty queries and later whitespace-normalized duplicates become empty positions. Retrieval uses all valid queries, while SDA retains original positions in its denominator. Outputs without valid queries score zero.
POI Retrieval Utility
A fixed seed samples two million POIs without replacement, proportional to retrievable catalog sizes across ten cities, without city prefiltering or target injection. Each query retrieves the top 100 with Qwen3-Embedding-8B and top 1,000 each with field-BM25 [22] and term-boosted BM25. Dense scores are recomputed over the deduplicated union across queries. Scaled reciprocal rank fusion [8] combines dense and field-BM25 ranks:
Here ranks the union and ranks the field-BM25 set ; term-boosted BM25 only expands candidates. Geographic weighting is , with in kilometers. Uniquely resolved locations use . Without explicit locations, and measures distance to the nearest of up to three historical anchors weighted by frequency and recency with a 14-day half-life. Unresolved locations or missing coordinates give . User scoring is
Hit@50, Recall@50, and NDCG@50 are computed over distinct future target POIs. All ground-truth targets remain in metric denominators, including those outside the candidate pool.
Service Intent Alignment
Target categories are deduplicated within the 263-category catalog and ranked by the frozen matcher’s maximum query score:
SIA@10 is category-level Exact-NDCG@10. Unmapped categories are excluded; users without mapped targets score zero.
Semantic Demand Alignment
GPT-OSS-120B receives queries, target POI names, and business categories, excluding model identities, reasoning, and retrieval scores. We use temperature 0, top- 1, seed 42, and medium reasoning effort. Core service relations are specific, broad, unrelated, or unknown, weighted , respectively. Brand, location, price, and facility constraints are labeled supported, contradicted, or unknown. Compatibility iff none is contradicted. Unknown constraints remain unverified. Pair scores are
These scores distinguish core relevance, constraint compatibility, and specific-service agreement without requiring the same store. For future target POIs , original query count , and , matching and coverage are
where . Matching averages each query’s best match, whereas coverage measures the list’s matched target fraction. Empty and duplicate positions contribute zero. Each query can cover multiple targets, each counted once. Equation (14) averages all four components before rounding.
Additional Results and Analysis
Semantic Components
Table 8 shows that all four SDA components improve from SFT to GRPO across four backbones and both benchmarks.
| Backbone | Core | Compat. | Spec. | Cov. |
| MapRec-Rich | ||||
| 4B | +11.89 | +12.24 | +6.78 | +34.67 |
| 9B | +11.87 | +12.93 | +7.60 | +35.57 |
| 27B | +29.28 | +36.59 | +34.42 | +53.01 |
| 122B-A10B | +20.64 | +25.15 | +21.44 | +43.64 |
| MapRec-Sparse | ||||
| 4B | +8.01 | +8.93 | +3.74 | +19.44 |
| 9B | +9.55 | +10.51 | +4.93 | +25.57 |
| 27B | +24.43 | +26.15 | +17.38 | +43.48 |
| 122B-A10B | +15.85 | +17.60 | +11.73 | +34.15 |
Table 8. SDA component gains from SFT to GRPO in percentage points, computed before rounding.
Human assessment of query quality. Two human assessors independently compare anonymous 4B GRPO outputs with Base and SFT on 100 MapRec-Sparse users. Table 9 reports preference rates of 70% to 80%, calculated as (Win 0.5 Tie)/100, supporting improved overall query quality.
Context-supported service coverage. GPT-4o and DeepSeek-v4-pro independently judge anonymous 4B Full and Shared outputs for 100 users per benchmark. They assess intent and key-constraint support from profiles and histories before seeing future targets, then keep these judgments fixed. Coverage counts future categories matched by supported queries, excluding unknown or conflict-ing key constraints. Each category counts once in the 12-category taxonomy, with all future categories retained in the denominator. Sparse averages use sampling-design weights.
Table 10 shows that Full consistently achieves higher context-supported coverage than Shared credit under both LLM judges on MapRec-Rich and MapRec-Sparse, for both five-query prefixes and complete lists. These results support the effectiveness of query-specific credit assignment in improving grounded service coverage.
Retrieval Analysis
These analyses use saved 4B outputs under the main protocol.
Primary versus expansion queries. For users with both query types, and count primary and expansion queries. We compare the first of each type. Primary NDCG@50 is higher for both SFT and GRPO on both benchmarks. GRPO’s primary and expansion scores are 0.1823% and 0.1812% on MapRec-Rich, and 0.1587% and 0.1250% on MapRec-Sparse, respectively.
Query ordering under a fixed budget. Using up to five GRPO queries, original, random, and reversed orders yield NDCG@50 of 0.2874%, 0.2383%, and 0.1779% on Rich and 0.2721%, 0.2404%, and 0.2424% on Sparse. Random ordering averages 20 permutations. Original order gains 0.0491 and 0.0317 percentage points over random, supporting prioritization under limited budgets.
Transfer across retrieval encoders. With the pool, lexical retrieval, and geographic weighting fixed, reward-independent BGE-M3 [7] yields positive GRPO-over-SFT NDCG@50 gains of 0.0031–0.1884 percentage points across all scales and benchmarks in Figure 6.
Figure 6. NDCG@50 gains of GRPO over SFT with BGE-M3.
Case Study
Figure 7 compares queries for a parent with campsite, scenic-area, and water-park histories. The Chimelong Water Park interaction and parent profile ground MILES-Rec’s parent–child activities in Q1. The Ten-Mile Gallery interaction and earlier hotel and dining interests support Q2’s complementary services. Repeated campsite and scenic-area interactions inform Q6’s broader family outings, combining several observed interests.
References
- [1]Krishna Acharya, Aleksandr V Petrov, and Juba Ziani. 2025. GLoSS: Generative Language Models with Semantic Search for Sequential Recommendation. arXiv:2506.01910 [cs.IR] https://arxiv.org/abs/2506.01910
- [2]Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. 2023. TALLRec: An Effective and Efficient Tuning Framework to Align Large Language Model with Recommendation. In Proceedings of the 17th ACM Conference on Recommender Systems. 1007–1014. doi:10.1145/3604915.3608857DOI
- [3]Beijing Academy of Artificial Intelligence. [n. d.]. BGE-Reranker-v2-M3: Model Card. Hugging Face. Retrieved October 14, 2026 from https://huggingface.co/BAAI/bge-reranker-v2-m3
- [4]Shuxian Bi, Chenxu Wang, Wenjie Wang, Yueqi Mou, Fuli Feng, Tang Biao, and Peng Yan. 2026. ReList: A Multi-objective Reasoning Framework for Diversified Listwise Query Recommendation. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track). Association for Computational Linguistics, San Diego, California, USA, 1392–1405. doi:10.18653/v1/2026.acl-industry.97DOI
- [5]Shiting Cao, Xiaochong Lan, Yuwei Du, Jie Feng, Yinxing Liu, Xinlei Shi, and Yong Li. 2026. Enhancing Local Life Service Recommendation with Agentic Reasoning in Large Language Model. arXiv preprint. arXiv:2604.14051 https://arxiv.org/abs/2604.14051
- [6]Yukou Cen, Jianwei Zhang, Xu Zou, Chang Zhou, Hongxia Yang, and Jie Tang. 2020. Controllable Multi-Interest Framework for Recommendation. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, 2942–2951. doi:10.1145/3394486.3403344DOI
- [7]Jianlyu Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024. M3-Embedding: Multi-Linguality, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation. In Findings of the Association for Computational Linguistics: ACL 2024. Association for Computational Linguistics, Bangkok, Thailand, 2318–2335. doi:10.18653/v1/2024.findings-acl.137DOI
- [8]Gordon V. Cormack, Charles L. A. Clarke, and Stefan Buettcher. 2009. Reciprocal Rank Fusion Outperforms Condorcet and Individual Rank Learning Methods. In Proceedings of the 32nd International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM, 758–759. doi:10.1145/1571941.1572114DOI
- [9]Ganqu Cui, Lifan Yuan, Ning Ding, Guowei Yao, Bingxiang He, Wei Zhu, Yuan Ni, Guotong Xie, Ruobing Xie, Yankai Lin, Zhiyuan Liu, and Maosong Sun. 2024. UltraFeedback: Boosting Language Models with Scaled AI Feedback. In Proceedings of the 41st International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 235). PMLR, 9722–9744. https://proceedings.mlr.press/v235/cui24f.htmlarxiv.org/abs/2310.01377
- [10]Qiyu Ding, Heng-Da Xu, Wei Zhang, Dongyi Lv, Changda Xia, Feng Xiong, and Mu Xu. 2026. Why Users Go There: World Knowledge-Augmented Generative Next POI Recommendation. arXiv:2605.11807 [cs.AI] https://arxiv.org/abs/2605.11807
- [11]Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2022. Recommendation as Language Processing (RLP): A Unified Pretrain, Personalized Prompt & Predict Paradigm (P5). In Proceedings of the 16th ACM Conference on Recommender Systems. ACM, 299–315. doi:10.1145/3523227.3546767
- [12]Kalervo Järvelin and Jaana Kekäläinen. 2002. Cumulated Gain-based Evaluation of IR Techniques. ACM Transactions on Information Systems 20, 4 (2002), 422–446. doi:10.1145/582415.582418DOI
- [13]Zhuohang Jiang, Yuxin Chen, Shijie Wang, Haohao Ou, Jindong Zhou, Wengi Fan, Qing Li, Dongxu Liang, and Jun Wang. 2026. Atomic Intent Reasoning: Bringing LLM Semantics to Industrial Cross-Domain Recommendations. In Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2. ACM, 7501–7510. doi:10.1145/3770855.3818320
- [14]Xiaochong Lan, Jie Feng, Yizhou Sun, Chen Gao, Jiahuan Lei, Xinlei Shi, Hengliang Luo, and Yong Li. 2025. Open-Set Living Need Prediction with Large Language Models. In Findings of the Association for Computational Linguistics: ACL 2025. Association for Computational Linguistics, Vienna, Austria, 5454–5472. doi:10.18653/v1/2025.findings-acl.285
- [15]Xiaochong Lan, Chen Gao, Shiqi Wen, Xiuxi Chen, Yingge Chen, Han Zhang, Huazhou Wei, Hengliang Luo, and Yong Li. 2023. NEON: Living Needs Prediction System in Meituan. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’23). ACM, 5292–5302. doi:10.1145/3580305.3599874DOI
- [16]Jiayi Liao, Sihang Li, Zhengyi Yang, Jiancan Wu, Yancheng Yuan, Xiang Wang, and Xiangnan He. 2024. LLaRA: Large Language-Recommendation Assistant. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM, 1785–1795. doi:10.1145/3626772.3657690DOI
- [17]Jianghao Lin, Rong Shan, Chenxu Zhu, Kounianhua Du, Bo Chen, Shigang Quan, Ruiming Tang, Yong Yu, and Weinan Zhang. 2024. ReLLa: Retrieval-enhanced Large Language Models for Lifelong Sequential Behavior Comprehension in Recommendation. In Proceedings of the ACM Web Conference 2024. ACM, 3497–3508. doi:10.1145/3589334.3645467DOI
- [18]Jiacheng Lin, Tian Wang, and Kun Qian. 2025. Rec-R1: Bridging Generative Large Language Models and User-Centric Recommendation Systems via Reinforcement Learning. Transactions on Machine Learning Research (2025). https://openreview.net/forum?id=YBRU9MV2vEarxiv.org/abs/2503.24289
- [19]Qidong Liu, Xiangyu Zhao, Yejing Wang, Zijian Zhang, Howard Zhong, Chong Chen, Xiang Li, Wei Huang, and Feng Tian. 2025. Bridge the Domains: Large Language Models Enhanced Cross-domain Sequential Recommendation. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM, 1582–1592. doi:10.1145/3726302.3729911DOI
- [20]Ilya Loshchilov and Frank Hutter. 2019. Decoupled Weight Decay Regularization. In International Conference on Learning Representations. https://arxiv.org/abs/1711.05101
- [21]Dongyi Lv, Qiuyu Ding, Heng-Da Xu, Zhaoxu Sun, Zhi Wang, Feng Xiong, and Mu Xu. 2026. Reasoning Over Space: Enabling Geographic Reasoning for LLM-Based Generative Next POI Recommendation. arXiv:2601.04562 https://arxiv.org/abs/2601.04562
- [22]Stephen Robertson and Hugo Zaragoza. 2009. The Probabilistic Relevance Framework: BM25 and Beyond. Foundations and Trends in Information Retrieval 3, 4 (2009), 333–389. doi:10.1561/1500000019DOI
- [23]Zihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Jaan Zuo. 2024. DeepSeek-Math: Pushing the Limits of Mathematical Reasoning in Open Language Models. arXiv:2402.03300. doi:10.48550/ARXIV.2402.03300
- [24]Guangming Sheng, Chi Zhang, Zilingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu. 2025. HybridFlow: A Flexible and Efficient RLHF Framework. In Proceedings of the Twentieth European Conference on Computer Systems. ACM, 1279–1297. doi:10.1145/3689031.3696075DOI
- [25]Sicong Wang, Ruiting Dong, Yue Liu, Bowen Zheng, Jun Meng, Jie Li, Shuaijun Guo, Yu Gu, Fanyi Di, and Xin Li. 2026. Generative Spatiotemporal Intent Sequence Recommendation via Implicit Reasoning in Amap. In Proceedings of the 20th ACM Conference on Recommender Systems. ACM, 1119–1127. doi:10.1145/3773078.3831861DOI
- [26]Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2023. Self-Instruct: Aligning Language Models with Self-Generated Instructions. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 13484–13508. doi:10.18653/v1/2023.acl-long.754
- [27]Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc V. Le, and Denny Zhou. 2022. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. In Advances in Neural Information Processing Systems, Vol. 35. 24824–24837. doi:10.5220/068431-1800arxiv.org/abs/2201.11903
- [28]Haoran Xin, Ying Sun, Chao Wang, and Hui Xiong. 2025. LMCDSR: Enhancing Cross-Domain Sequential Recommendation with Large Language Models. ACM Transactions on Information Systems 43, 5, Article 120 (2025), 33 pages. doi:10.1145/3715099DOI
- [29]Song Yan, Jiamou Liu, and Kaiqi Zhao. 2022. GETNext: Trajectory Flow Map Enhanced Transformer for Next POI Recommendation. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’22). ACM, 1144–1153. doi:10.1145/3477495.3531983DOI
- [30]Shuwei Yuan, Mingqian Ding, Luxin Liu, Rong Xiao, and Xiaoyi Zeng. 2026. EAGER: Enrich-and-Align Generative Query Recommendation from Clicked Items in E-commerce Search. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track. Association for Computational Linguistics, Budapest, Hungary, 1900–1912. https://aclanthology.org/2026.emnlp-industry.111/
- [31]Penglong Zhai, Jie Li, Fanyi Di, Yue Liu, Yifang Yuan, Jie Huang, Peng Wu, Sicong Wang, Mingyang Yin, Tingting Hu, Yao Xu, and Xin Li. 2025. Cognitive-Aligned Spatio-Temporal Large Language Models for Next Point-of-Interest Prediction. arXiv:2510.14702 [cs.AI] https://arxiv.org/abs/2510.14702
- [32]Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Wei, An Yang, Dayiheng Liu, Junyang Lin, Fei Huang, and Jingren Zhou. 2025. Qwen3 Embedding: Advancing Text Embedding and Reranking Through Generative Models. arXiv:2506.05176 [cs.CL] https://arxiv.org/abs/2506.05176
- [33]Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, and Zheyan Luo. 2024. LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations). Association for Computational Linguistics, Bangkok, Thailand, 400–410. doi:10.18653/v1/2024.acl-demos.38DOI
- [34]Yaochen Zhu, Harald Steck, Dawen Liang, Yinhua He, Vito Ostuni, Jundong Li, and Nathan Kallus. 2026. Rank-GRPO: Training LLM-based Conversational Recommender Systems with Reinforcement Learning. In International Conference on Learning Representations. https://proceedings.iclr.cc/paper_files/paper/2026/hash/bf8f2d4e90018c0fae060df09a91ce3-Abstract-Conference.htmlarxiv.org/abs/2510.20150


