2.1 Group recommendation with geographic factors

Group recommender systems have long relied on preference aggregation strategies to support collective decision-making. Two main strategies exist: score aggregation, which merges individual recommendation results or ranked lists into a unified group list, and preference aggregation, which integrates members’ preference information to construct a single group profile and performs recommendations for this virtual user [12, 13, 14]. Building on these aggregation paradigms, recent research has increasingly adopted deep learning models to better capture complex preference signals and user-item relations in group settings. Specifically, prior work introduces deep learning techniques such as Graph Neural Networks (GNNs), attention mechanisms, and knowledge graphs to model complex user-item interactions [15, 16, 17]. However, these methods often overlook spatial mobility constraints, limiting their applicability to real-world location-based scenarios.

To address the lack of spatial factors in group recommendation, several studies have attempted to incorporate geographic modeling into group recommendation tasks. The context-based geoprocessing framework [18] searches meeting locations on road networks that minimize total group travel cost by combining path planning with heuristic search. Similarly, the optimal sequenced route for group meetup (OSR-G) model [19] applies origin-destination (OD) path optimization to reduce the overall travel distance of group members. While these approaches make progress in geographic optimization and are effective in spatial terms, they generally lack modeling support for user preference signals.

To account for individual differences, the method in [11] incorporates MBTI (a widely used but high-level personality typology) by mapping users into type-based categories and applying spatial clustering to improve fairness in meeting location selection. Beyond personality descriptors, other studies adopt topic-based user modeling approaches such as P-LAG [1]. While these designs introduce some degreeof personalization, they rely on static or coarse-grained representations and therefore struggle to capture fine-grained, context-aware preference variations, which limits their adaptability to diverse recommendation needs.

Overall, existing studies suggest that effective GMLR requires jointly considering mobility costs, user preferences, and location attributes to produce recommendations that are both preference-consistent and spatially feasible. To address these issues, we quantify travel costs and jointly model user preferences and merchant attributes, enabling recommendations that balance spatial feasibility, personal preferences, and group satisfaction.

Unlike traditional Point of Interest (POI) recommendation models focused on individual trajectory prediction [20, 21], as well as recent group POI recommendation models [22–24], our method targets group decision-making for meeting locations under spatial constraints while aligning collective preferences, making it well-suited for scenarios such as group meetups or collaborative meeting location planning.

2.2 LLMs for group recommendation

Recent research has increasingly integrated LLMs into recommender systems, leveraging their strong text understanding and generation capabilities to reformulate recommendation tasks as language modeling problems. Representative work, such as the P5 framework transforms user-item interactions and review texts into unified prompts, enabling recommendation tasks through language modeling [25]. Building on the semantics of reviews, RDRec and HDRec further enhance preference learning by incorporating review content and rationale distillation, enabling LLMs to capture deeper user intent beyond sparse interaction logs [26, 27].

To improve generalization ability, studies such as Zero-Shot and GenRec demonstrate the strong ada-ptability of LLMs in zero-shot and generative recommendation scenarios [28, 29]. FineRec further proposes extracting attribute-opinion pairs from reviews to construct fine-grained preference graphs for enhanced structural explainability [30], while POD [31] and ReviewRec [32] improve LLM efficiency and personalization through prompt compression and review-driven mechanisms.

Despite these advances in individual recommendation, the application of LLMs to group recommendation remains in its early stages. Some studies attempt to use LLMs to generate group recommendation rankings and rationales based on members’ rating lists [33], or extract topic labels from user reviews and learn group preference representations via graph neural networks before generating recommendations [34], or adopt graph-based aggregation, with some recent work beginning to explore fairness concerns [35].

However, these methods often rely on explicit topic modeling and static structures, making it difficult to capture the heterogeneity of users’ semantic preferences, and thus lack the capacity for dynamic negotiation during the recommendation process.

Our work fills this gap by constructing semantic profiles for users and merchants directly from free-form reviews, enabling fine-grained modeling of preferences and implicit intentions. We further adopt CoT prompting to reason over multiple profiles, balancing satisfaction and conflict via multi-turn inference. The resulting recommendations align more closely with both individual intent and group consensus.

3.1 MEET-LLM framework

To address the disconnect among multi-preference coordination, spatial feasibility, and real-world executability in GMLR, we propose MEET-LLM, an LLM-powered multi-agent collaborative recommendation framework [36]. MEET-LLM integrates semantic understanding, spatial optimization, and dynamic user modeling to bridge the digital preference space and the physical travel space in urban environments, where feasibility is shaped by geographic context and road-network routing. The framework moves beyond abstract rankings toward executable meeting location decisions that are accompanied by route-feasible plans.

As shown in Figure 2, the framework follows a perception-decision-action-feedback loop. Perception constructs mobility-aware contextual evidence, decision performs conflict-aware preference reasoning under feasibility constraints, action produces executable outputs that can be integrated with navigation services, and feedback updates historical context for continual improvement. For implementation, we instantiate this loop with seven components, L, C, M, S, D, O, and K.

Figure 2 The framework of MEET-LLM

L (Large Language Model): provides semantic understanding throughout the pipeline, supporting review analysis, preference extraction, and recommendation generation. It serves as the core reasoning agent within the perception-inference-feedback loop.

C (Character): builds foundational representations for perception by modeling users’ semantic preferences and mobility capabilities, as well as merchant attributes and target customer features. It constructs semantic and reachability representations between users and venues.

M (Memory): stores historical mobility data, preference evolution, and feedback logs of users. It provides contextual support for perception and decision-making, receives updates from the feedback module, and enables long-term adaptive learning of user profiles.

S (Perception): based on user-merchant profiles from C and historical data from M, this module integrates user geographic positions and mobility-related characteristics to construct a multimodal perception state that captures preference-relevant mobility constraints, which is then provided as input to D.

D (Decision): receives the perception state from S and the semantic reasoning capability from L, introduces the group-level mobility cost (MC) to filter out infeasible locations, and performs multi-round CoT reasoning to resolve preference conflicts and generate recommendations that are both preference-compatible and spatially feasible.

O (Action): based on the output from D, this module evaluates mobility feasibility and transfers the decision to real navigation or coordination services, generating a structured and executable recommendation, where the system outputs a ranked Top-k list (k = 3 by default) of candidate meeting locations, each accompanied by a reasoning explanation and navigation path.

K (Feedback): generates recommendation explana-tions, gathers feedback on acceptance and interpretability, and triggers re-inference through S and D when results are unsatisfactory or when the average semantic match (ASM) score falls below 0.4, including feeding back to S for adaptive spatial optimization when the number of recommended merchants is insufficient. All feedback is stored in M for continual optimization.

The entire system forms a closed-loop reasoning chain centered on LLMs—linking perception, decision, action, and feedback—to achieve integrated reasoning across semantic and geographic spaces. The L module provides semantic capabilities throughout the pipeline; C, S, and M jointly construct the perception foundation; D coordinates group preferences and spatial feasibility; O ensures real-world execution; and K enables adaptive optimization. Together, they form a dynamic multi-agent recommender system that transforms group recommendation from abstract semantic ranking into executable, collaborative urban-level tasks.

3.2 Core framework modules

The MEET-LLM framework aims to recommend meeting locations that are both preference-compatible and convenient for all group members to reach. First, the PPM extracts structured user preference profiles and merchant attribute profiles from historical review data. Second, the SOM computes mobility cost and identifies a set of spatially filtered candidate merchants. Third, the CRM integrates user preference profiles, merchant attribute profiles, and mobility cost through multi-turn LLM reasoning to generate the final meeting location recommendations.

3.3 Preference Profiling Module

Existing group recommendation methods often rely on ratings or predefined structures, which are insufficient to capture nuanced preferences expressed in user-generated text. To enable fine-grained and intent-aware preference modeling, we construct natural-language profiles from reviews. User reviews provide scalable, semantically rich, and behaviorally informative signals that reveal implicit decision cues beyond structured data [37]. Leveraging this, this module employs an LLM with guided CoT prompting to take historical reviews as input and generate expressive user preference profiles and merchant attribute profiles.

Building on prior work such as P5 [25], this module introduces a scene-differentiated strategy to extract fine-grained user preferences across diverse merchant contexts. Using carefully designed prompt templates and guided reasoning, the system distills structured preferences from reviews into a three-part representation: User Likes, User Dislikes, and Additional Insights. Likes and Dislikes summarize explicit, scene-specific positive and negative preferences expressed toward different types of merchants. Additional Insights consolidate implicit signals beyond surface-level statements, including (i) cross-scene and contrastive behavioral patterns (e.g., solitude-seeking in cafés but family-oriented behavior in shopping contexts) and (ii) inferred decision motives and psychological tendencies that help explain why certain attributes matter in a given scene. This dual-layer semantic modeling constructs a more complete user profile by combining surface-level preferences with inferred behavioral patterns and decision motives, thereby enhancing personalization and interpretability in downstream recommendation reasoning.

For each user-merchant interaction, the module extracts a structured user preference profile and a merchant attribute profile from the review text. The user preference profile summarizes what the user likes, dislikes, and additionally mentions in reviews. The merchant attribute profile consists of two parts: the functional attribute profile and the target audience profile.

To ensure profile completeness and semantic fidelity, the system aggregates each user's historical reviews and performs instance-level reasoning across dual layers. Merchant profiles are similarly derived by aggregating multi-user reviews to identify consistent features and positioning. Detailed prompt texts are provided in Tables S2 and S3 of the Supporting Information (SI).

3.4 Spatial Optimization Module

In this module, we assume that users can adopt feasible travel modes given their individual conditions and local infrastructure constraints. Accordingly, mobility cost is treated as an abstraction of travel burden, primarily determined by route distance and physical ability. Specific transportation modes can be selected by users at the final navigation stage.

To model group-level differences in physical mobility and preliminarily filter candidate meeting locations, this module aims to identify a candidate central area that offers relatively low mobility costs and is balanced across group members. The system begins by collecting the geographic coordinates (latitude and longitude) of all group members and calculating their arithmetic mean to determine an initial centroid. This centroid serves as the starting point for filtering candidate meeting locations.

The mobility burden of each user is determined by physical ability and route distance. This is simplified into an individual mobility cost (MC_i), as shown in Equation (1):$\begin{array}{c}{\text{MC}}_i=b_i\cdot d_i\end{array}$(1)where b_i is the physical ability weight (set to 2 for mobility-impaired users, 1.5 for elderly users aged ≥ 65, and 1 for average adults); d_i is the route distance from user i to the current centroid (computed using GraphHopper).

The specific weight values used in this study are illustrative and chosen to demonstrate the feasibility of the proposed framework. In practice, these parameters can be dynamically adjusted based on contextual data, regional characteristics, or learned from historical usage patterns.

While real-world mobility burden may involve multiple factors (e.g., weather and transportation availability), the current implementation uses physical ability and travel distance as illustrative components to model mobility cost. The proposed framework is inherently extensible and can readily incorporate additional factors as relevant data become available.

To reflect group heterogeneity, the system computes normalized weights based on the inverse of MC_i:$\begin{array}{c}{w}_i=\frac{1/{\text{MC}}_i}{{{\displaystyle \sum }}_{j=1}^{n}1/{\text{MC}}_j}\end{array}$(2)

Using these weights, the updated centroid coordinates (lat', lng') are calculated as:$\begin{array}{c}{\text{lat}}^{\prime }={\displaystyle \sum _{i=1}^n{w}_i}\cdot \mathrm{l}a{\text{t}}_i,{\text{lng}}^{\prime }={\displaystyle \sum _{i=1}^n{w}_i}\cdot \mathrm{l}\mathrm{n}{\mathrm{g}}_i\end{array}$(3)

This iterative process continues until convergence, as shown in Equations (2) and (3). Finally, the system uses the converged centroid as the center and a radius R to select businesses within this region as candidate meeting locations for downstream recommendation reasoning, with the radius adaptively adjusted through the feedback mechanism when necessary.

3.5 CoT Reasoning Module

To generate the final GMLR recommendations, the CRM feeds an LLM with user preference profiles, merchant attribute profiles, and user and merchant coordinates within a retrieval-augmented generation (RAG) structure. Guided by a CoT prompting process (Figure 3), the LLM outputs Top-3 meeting-location recommendations together with explanations and executable navigation routes for group members. LLM-based reasoning is sensitive to prompt design [38]. Excessive context can introduce noise and degrade inference [39]. We therefore use a multi-turn dialogue that decomposes reasoning across rounds while maintaining contextual continuity and retaining key information for multi-source decision integration. Detailed prompt texts are provided in Tables S4 and S5 of the SI.

Figure 3 Flowchart of the CRM

Leveraging CoT prompting [40], we divide the reasoning process into two structured rounds. Each round involves the following key prompt components:

Task Description: specifies the round's objective.

User Preference Profiles: loaded into the prompt to support inference.

Format Constraints: guide LLM output format to reduce ambiguity and enforce structured responses.

Merchant Attribute Profiles: filtered merchant sets generated from the SOM are introduced as the decision space.

MC and average mobility cost (AMC): the MC and AMC values are integrated to guide spatial feasibility decisions.

Meeting Intentions: user goals (e.g., dining and massage) may or may not be specified and are used to refine decision alignment when available.

By composing prompts with these structured components, the LLM performs layered reasoning, generating both final recommendations and interpretable explanations.

4.1 Experimental setup

This study evaluates the effectiveness of the proposed MEET-LLM system for GMLR, focusing on group acceptance. Experiments are conducted on the Yelp dataset [41], which provides user IDs, review histories, merchant IDs, and merchant latitude and longitude. We use Lutz as the primary evaluation city and additionally include Wesley Chapel to examine generalization in cross-city scenarios. To ensure data quality and obtain sufficiently stable preference profiles, we retain users with more than 10 historical review interactions. GPT-4o is used as the backbone model due to its strong reasoning capabilities. To ensure consistent output, we set the temperature to 0, and structured prompts with format constraints and multi-turn memory are adopted to enhance consistency. To simulate real-world group decision-making, we generate 10 groups each for sizes 3, 4, and 5. For spatial realism, user locations are sampled from residential areas and transport hubs, with five geographically distributed configurations created per group size. Detailed group composition information is provided in the SI, Section 3.

To benchmark MEET-LLM, we adopt four representative group recommendation strategies as baselines: CF-A, which aggregates predicted scores by averaging across group members [42]; CF-L, which follows the least-misery principle by selecting the minimum score among members [42]; TGR-RA, which implements recommendation aggregation [43]; LLM-SC, which prompts an LLM to apply social choice-based aggregation strategies to produce group recommendations [9]. We further evaluate robustness by replacing the backbone with GPT-3.5-turbo, GPT-4o-mini, and GPT-4.1 under the same prompts.

All baselines use user-based collaborative filtering [44] for score prediction:$\begin{array}{c}{r}_{u_i,x_j}=r_{u_i}^{-}+\frac{{{\displaystyle \sum }}_{u_k\in N}\left(r_{u_k,x_j}-r_{u_k}^{-}\right)\cdot \text{Similarity}\left(u_i,u_k\right)}{{{\displaystyle \sum }}_{u_k\in N}\left|\text{Similarity}\left(u_i,u_k\right)\right|}\end{array}$(4)where $r_{u_i,x_j}$ denotes the predicted rating of user u_i for merchant x_j; $r_{u_i}^{-}$ and $r_{u_k}^{-}$ represent the average ratings of users u_i and u_k, respectively; $r_{u_k,x_j}$ is the rating given by neighbor u_k to merchant x_j; Similarity (u_i, u_k) indicates the similarity between users u_i and u_k, calculated using the weighted Pearson correlation; and N is the set of neighbors similar to user u_i.

To evaluate the quality of generated explanations, we conducted a human study based on pairwise comparison. Since research on LLM-based group recommendation is still in its early stage, we compared our structured reasoning approach with two representative single-user explanation generation models: P5 [25], which uses a T5-based model, and XRec [45], which summarizes user reviews using GPT-4o. For each model, we randomly sampled 100 explanation instances for evaluation.

4.2 Evaluation strategy

To comprehensively evaluate MEET-LLM, we design a framework using six core metrics across four dimensions.

4.3 Sensitivity analysis of α

To examine the robustness of the CAS in Equation (5), we vary the weighting coefficient α and recompute CAS while keeping all other settings unchanged. As shown in Figure 4, we analyze for both 3-person and 5-person groups with α in {0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8}. Notably, our method remains consistently competitive and achieves the best CAS in most settings across the tested α range.

Figure 4 Sensitivity analysis of α on CAS

4.4 Experimental results

Across all group sizes, MEET-LLM consistently outperforms the baseline models in terms of semantic preference matching, spatial feasibility, and group-level acceptability, as shown in Table 1. Specifically, it achieves the highest scores on CAS, G@3, and MAR, indicating a strong ability to generate recommendations that are not only semantically aligned and geographically feasible but also broadly accepted by group members. G@3 and MAR reflect different dimensions of group acceptability-namely, unanimous agreement and majority consensus—and offer complementary perspectives on how well the model supports group coordination.

Although MEET-LLM does not always achieve the highest ASM, this is mainly due to its explicit consideration of MC. Unlike methods that simply maximize semantic preference alignment, MEET-LLM selects meeting locations that are convenient for all members, leading to significantly lower AMC values. This comprehensive approach has substantially enhanced group satisfaction. Compared with traditional baselines, CF-L improves preference matching by focusing on the least satisfied member, whereas CF-A and TGR-RA show unstable performance across different group settings. Notably, even with large language models, LLM-SC performs worse than MEET-LLM in the location recommendation setting. Ultimately, these baseline approaches lack a joint consideration of overall group satisfaction and spatial feasibility, resulting in suboptimal performance.

We further examine robustness across LLM back-bones and urban contexts. As shown in Figure 5, MEET-LLM maintains consistently strong CAS when switching the backbone model, and the overall advantage over baselines remains stable in both Lutz and Wesley Chapel. These results suggest that the observed gains are not specific to GPT-4o and that the proposed framework generalizes across different LLM backbones and city-level spatial distributions. We further report efficiency statistics of MEET-LLM, including end-to-end runtime and total token consumption across four LLM backbones, as shown in Table 2.

Figure 5 CAS comparison of CF-A, CF-L, TGR-RA, LLM-SC, and M-L (MEET-LLM) in Lutz (A) and Wesley Chapel (B) across different LLM backbones

Under the EQE protocol, MEET-LLM consistently outperforms the two baseline explainers across all six criteria. As illustrated in Figure 6, it achieves uniformly higher scores on clarity, reasoning, justification, trustworthiness, faithfulness, and overall preference. The most pronounced gains appear in reasoning and justification, indicating that MEET-LLM generates more structured and persuasive explanations.

Figure 6 Evaluation of pairwise quality comparisons of human-evaluated explanations generated by MEET-LLM and baseline models (p-value < 0.05)

Figure 7A illustrates a representative suburban case in which the initial spatial region yields insufficient candidate merchants. Through feedback-driven adaptive expansion of the search radius, the framework enables the CRM to identify a sufficient set of suitable merchants for recommendation. This example demonstrates the ability of the proposed framework to dynamically adjust the candidate search region and handle non-central or uneven user distributions. In addition, a recommendation scenario involving a larger group of ten users is provided as an illustrative example (Figure 7B), indicating the applicability of the framework under increased group size and spatial dispersion.

Figure 7 Adaptive spatial expansion and large-group recommendation in MEET-LLM using a case study in Lutz, Florida, USA. (A) Illustration of adaptive spatial expansion under a suburban user distribution. The inner dashed circle denotes the initial spatial region with insufficient candidate merchants; through feedback-driven refinement, the search radius is expanded to the outer circle to identify suitable merchants. (B) Recommendation results for a larger group of ten users

4.5 Ablation study

The MEET-LLM framework consists of three core modules: PPM, SOM, and CRM. To evaluate the contribution of each module to the overall recommendation pipeline, we conduct a series of ablation experiments.

Each ablation setting removes a critical component, as follows: No-PPM: disables PPM, directly feeds raw review text into CRM; No-SOM: omits the spatial filtering step in SOM and directly passes the unrefined candidate set to the reasoning stage; No-CRM: disables CRM and ranks candidates using a deterministic CAS-based weighted score of ASM and MC; No-MC: excludes all MC-related prompts from both SOM and CRM; and Only-MC, an MC-only baseline where only SOM is retained. We assess the effects of each ablation on preference, mobility feasibility, and group-level acceptance metrics. Results are summarized in Table 3.

Through the ablation study, we observe that No-PPM directly feeds raw review text into CRM. Compared with MEET-LLM, removing structured preference and attribute profiles leads to consistent drops in CAS, G@3, and MAR. This suggests that profile-based representations provide cleaner and more decision-relevant signals than raw review text, which is often noisy and redundant. Without structured profiles, the reasoning stage becomes less effective at capturing the meeting intention and identifying mutually acceptable options, resulting in lower group-level acceptance. The No-SOM setting exposes the model to an unfiltered merchant pool. While this increases flexibility for preference matching, it leads to higher AMC and reduced acceptance. Thanks to the downstream CRM, which still considers mobility, feasibility degradation is partially mitigated. The No-MC variant builds on No-SOM by further removing all mobility cost prompts from CRM. The system thus prioritizes semantic fit alone, achieving the highest ASM but causing AMC to spike and acceptance metrics to drop. This confirms the joint importance of SOM and CRM, particularly CRM, in balancing preferences and feasibility. The Only-MC configuration disables both PPM and CRM, selecting locations purely by travel cost. Although AMC is minimized, semantic preference match and group acceptance drop sharply, highlighting that mobility alone cannot replace the central role of preference reasoning. No-CRM disables CRM and instead applies a deterministic CAS-based weighted scoring rule. However, while this variant achieves a relatively high CAS, its acceptance-oriented metrics (G@3 and MAR) remain clearly below those of MEET-LLM. This suggests that the deterministic rule does not explicitly model the meeting intention, resolve preference conflicts, or provide nuanced trade-offs, which can reduce the perceived acceptability of the final recommendations.

Table 4 summarizes the ablation settings, and Table 5 reports the corresponding relative contribution ratios. The results show that SOM contributes the most to CAS sensitivity, CRM contributes most strongly to acceptance-oriented metrics such as G@3 and MAR, and PPM provides a stable contribution across all metrics. PPM extracts structured preferences from reviews to support alignment and explanation. SOM filters candidates based on travel feasibility. CRM plays a central role in final decision-making, balancing preference fit and mobility cost. Together, these modules enable recommendations that are both acceptable and feasible in real-world group settings.