Paper deep dive
Behavior2Trip: Towards Personalized Travel Planning via User Behavior Trajectory
Zihao Cheng, Yingyu Shan, Hongru Wang, Zeming Liu, Xinyi Wang, Xiangrong Zhu, Yuhang Guo, Wei Lin, Yunhong Wang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 8/28/2026, 4:13:15 AM
Summary
This paper introduces Behavior2Trip, a new benchmark and task for Behavior-Aware Travel Planning, which infers user preferences from past behavioral trajectories (clicks, favorites, bookings) rather than explicit instructions. The benchmark contains 11,400 instances derived from a major Chinese online travel platform. The authors propose B2T-Agent, a reinforcement learning-based agent using Qwen3-8B that leverages external tools and internal memory to generate personalized travel plans. Experiments show B2T-Agent significantly outperforms baselines, including GPT-4.1, on the Behavior2Trip benchmark and demonstrates strong generalization to the TravelPlanner benchmark.
Entities (7)
Relation Signals (5)
B2T-Agent → builtupon → Qwen3-8b
confidence 95% · B2T-Agent built upon Qwen3-8B outperforms all baselines
Behavior2Trip → usedfor → Behavior-Aware Travel Planning
confidence 95% · To facilitate research on this task, we introduce Behavior2Trip... comprising 11,400 instances.
B2T-Agent → outperforms → GPT-4.1
confidence 92% · B2T-Agent built upon Qwen3-8B outperforms all baselines... Qwen3-8B trained with B2T-Agent also outperforms GPT-4.1 on the TravelPlanner benchmark
Behavior2Trip → derivedfrom → Chinese online travel platform
confidence 90% · Behavior2Trip, a benchmark constructed from one of the largest Chinese online travel platforms
B2T-Agent → optimizedby → GRPO
confidence 90% · we adopt GRPO (Shao et al., 2024b) to optimize the policy model
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Travel planning agents assist users in generating personalized travel plans by modeling their individual preferences. Existing agents either rely on explicit user instructions or engage in multi-turn clarification to elicit user preferences. However, both approaches overlook the rich behavioral signals latent in users' past behaviors, which implicitly encode their preferences. This over-reliance on active user input increases interaction burden and limits plan personalization. To bridge this gap, we introduce a new task, Behavior-Aware Travel Planning, which infers user preferences directly from past behaviors and generates personalized travel plans. To facilitate research on this task, we introduce Behavior2Trip, a benchmark constructed from one of the largest Chinese online travel platforms, comprising 11,400 instances. Each instance represents an average of 39.8 past user behaviors spanning 14 attributes across 5 preference dimensions. We further propose B2T-Agent, a reinforcement learning-based agent that leverages user behavior trajectories, interacts with external tools for preference-aligned retrieval, and maintains an internal memory module. Experiments on Behavior2Trip show that GPT-4.1 achieves a full-constraint pass rate of only 0.5\% on the hardest tasks, while B2T-Agent built upon Qwen3-8B outperforms all baselines, highlighting the substantial challenge of this task. Moreover, Qwen3-8B trained with B2T-Agent also outperforms GPT-4.1 on the TravelPlanner benchmark, demonstrating strong generalization. Code and data are available at this https URL
Tags
Links
- Source: https://arxiv.org/abs/2608.26807v1
- Canonical: https://arxiv.org/abs/2608.26807v1
Trouble viewing inline? Open PDF directly →
Full Text
80,060 characters extracted from source content.
Expand or collapse full text
Behavior2Trip: Towards Personalized Travel Planning via User Behavior Trajectory Zihao Cheng 1* , Yingyu Shan 2* , Hongru Wang 3* , Zeming Liu 1† , Xinyi Wang 4 , Xiangrong Zhu 4‡ , Yuhang Guo 2 , Wei Lin 4 , Yunhong Wang 1 1 School of Computer Science and Engineering, Beihang University, Beijing 2 Beijing Institute of Technology 3 University of Edinburgh 4 Meituan Inc. * Equal Contribution † Corresponding Author ‡ Project Leader Email: zihaocheng@buaa.edu.cn, zmliu@buaa.edu.cn Abstract Travel planning agents assist users in gener- ating personalized travel plans by modeling their individual preferences. Existing agents either rely on explicit user instructions or en- gage in multi-turn clarification to elicit user preferences. However, both approaches over- look the rich behavioral signals latent in users’ past behaviors, which implicitly encode their preferences. This over-reliance on active user input increases interaction burden and limits plan personalization. To bridge this gap, we introduce a new task, Behavior-Aware Travel Planning, which infers user preferences di- rectly from past behaviors and generates per- sonalized travel plans. To facilitate research on this task, we introduce Behavior2Trip, a benchmark constructed from one of the largest Chinese online travel platforms, comprising 11,400 instances. Each instance represents an average of 39.8 past user behaviors spanning 14 attributes across 5 preference dimensions. We further propose B2T-Agent, a reinforcement learning-based agent that leverages user behav- ior trajectories, interacts with external tools for preference-aligned retrieval, and maintains an internal memory module. Experiments on Be- havior2Trip show that GPT-4.1 achieves a full- constraint pass rate of only 0.5% on the hard- est tasks, while B2T-Agent built upon Qwen3- 8B outperforms all baselines, highlighting the substantial challenge of this task. Moreover, Qwen3-8B trained with B2T-Agent also out- performs GPT-4.1 on the TravelPlanner bench- mark, demonstrating strong generalization 1 . 1 Introduction Travel planning agents have become an essential role for assisting users in complex travel decision- making (Xie et al., 2024; Shao et al., 2025; Zhang et al., 2024; Ni et al., 2025; Deng et al., 2025). With the advancement of Large Language Models 1 Code and data are available athttps://github.com/ BUAA-IRIP-LLM/Behavior2Trip. + Resident:Chongqing Preferences:Enjoy eating spicy foodand prefer staying in 5-star hotels, avoiding repetitive tourist attractions. ...... [Click]The Ritz-Carlton, Five Star,Four Seasons Hotel, Five Star,Tiananmen Square,Classic Chongqing SpicyRestaurant,... [Favorite]The Ritz-Carlton,Tiananmen Square, Classic Chongqing Restaurant,... [Book]The Ritz-Carlton,Tiananmen Square, Classic Chongqing Restaurant,... BehaviorTrajectory [Day 1, Lunch] ChongqingSpicy Hotpot, 11:35–13:00, ¥275 [About 30 minutes by car, 32 minutes by Metro Line 8] [Day 1, Attraction] The Forbidden City, 14:00, ¥70 [About 26 minutes by car, 41 minutes by Metro Line 1] [Day1,Accommodation]JW Marriott Luxury Hotels,¥1275 ...... Explaination:Based on your behavior trajectory, it’s clear that you enjoy spicy food and5-starhotels. I also noticed that you visited Tiananmen Square two months ago, so I will not include it in future recommendations.. Profile I’mplanningatriptoBeijingforthreedays,couldyougivemeatravelplan? UserInstruction Input: UserInstructionBehaviorTrajectory ReflectUserPreference Figure 1: An example of Behavior2Trip, where the agent analyzes a user’s behavioral trajectory to implic- itly infer preferences, enabling the direct generation of personalized travel plans from implicit instructions. (LLMs) (Ma et al., 2025; He et al., 2025; Cheng et al., 2025a, 2026a; Gao et al., 2025), these agents are gaining the ability to interact with external en- vironments to generate comprehensive travel plans (Jiang et al., 2024; JU et al., 2024; Luo et al., 2025; Zhang et al., 2025b; Singh et al., 2025; Zhang et al., 2025a), paving the way for more practical and in- telligent travel solutions. Existing travel planning agents follow two main paradigms: (1) Single-turn Explicit Instruction (Xie et al., 2024; Shao et al., 2025; Ni et al., 2025; de la Rosa et al., 2024a), which requires users to provide all preferences upfront in a single turn; and (2) Multi-turn Clarification (Zhang et al., 2024; Chen et al., 2024), which interactively elicits prefer- ences through dialogue. However, both paradigms heavily rely on users’ active input, overlooking be- havioral signals from users’ past interactions (e.g., clicks, favorites, and bookings) that naturally en- code their preferences (Ye et al., 2011; Chen et al., 2023). This over-reliance on active input increases interaction burden and limits plan personalization. arXiv:2608.26807v1 [cs.CL] 27 Aug 2026 DatasetUBTSIIOPUPFLIBUP# POIs# POI Attr.# ConstraintsInstances TravelPlanner (Xie et al., 2024)✗ ✗ ✓ ✗ ✗exp.19.9k6.61520 ChinaTravel (Shao et al., 2025)✗ ✗ ✓ ✗ ✗exp.12.2k6.823154 Ask–before–plan (Zhang et al., 2024) ✗ ✓ ✗ ✗ ✗exp.19.9k6.6122,000 ITINERA (Tang et al., 2024)✗ ✗ ✓ ✗ ✗exp.7.6k9.0– TRIP–PAL (de la Rosa et al., 2024a) ✗ ✗ ✓ ✗ ✗exp.1.8k–2100 Travel–Agent (Chen et al., 2024)✗ ✗ ✗ ✗ ✗exp.–2020 TP–RAG (Ni et al., 2025)✗ ✗ ✓ ✗ ✗exp.85.6k–2,348 Behavior2Trip (ours)✓imp.80.9k17.73211,400 Table 1: Comparison between the Behavior2Trip and other benchmarks. UBT: User Behavior Trajectory, SII: Support Implicit Instruction, OP: One–shot Planning, UPF: User Profile, LIB: Low Interaction Burden, UP: User Preference. The ✓ indicates full support, the ✗ indicates no support, “–” denotes not mentioned in paper. To bridge this gap, we introduce a new task, Behavior-Aware Travel Planning, which infers user preferences directly from past behaviors and generates personalized travel plans without requir- ing active user input. To support this task, we construct Behavior2Trip, a benchmark grounded in real user data from one of the largest Chinese on- line travel platforms, comprising 11,400 instances, each with an average of 39.8 past user behaviors spanning 14 attributes across 5 preference dimen- sions. The dataset is built upon a sandbox environ- ment with 80.6k real-world POIs, where behavior trajectories are generated from diverse preference profiles and validated through a rigorous two-stage quality control process. To address the challenge of implicitly infer- ring user preferences from behavior trajectories while satisfying long-horizon planning constraints, we propose B2T-Agent, a reinforcement learning- based agent that generates personalized travel plans from user behavior trajectories. Unlike prompt- driven methods (Yuan et al., 2025; Xie et al., 2024), B2T-Agent features a structured action space sup- porting both external tool invocation and internal memory management, optimized by a compos- ite reward function that rewards plan quality and penalizes invalid actions. We conduct comprehensive experiments on Be- havior2Trip to evaluate both open-sourced and closed-sourced models.Results highlight the substantial challenge of this task: even GPT-4.1 achieves a full-constraint pass rate of only 0.5% on the hardest tasks, while B2T-Agent built upon Qwen3-8B significantly outperforms all baselines. Moreover, Qwen3-8B trained with B2T-Agent also outperforms GPT-4.1 on the TravelPlanner (Xie et al., 2024), demonstrating strong generalization. Overall, the contributions of this paper are as fol- lows: •We identify a novel task, Behavior-Aware Travel Planning, which generates personalized travel plans by inferring user preferences directly from past behaviors, without requiring explicit or iterative user input. •To support this task, we introduce Behav- ior2Trip, a large-scale benchmark grounded in real user data, comprising 11,400 instances with rich behavioral signals spanning 14 attributes across 5 preference dimensions. • To tackle the challenges of this task, we pro- pose B2T-Agent, a reinforcement learning-based agent with structured external tool invocation and internal memory management. Experiments show that B2T-Agent built upon Qwen3-8B sig- nificantly outperforms GPT-4.1 and generalizes well to the TravelPlanner benchmark. 2 Related Work 2.1 Travel Planning Benchmarks Travel planning is a representative task for eval- uating agents’ long-horizon planning capabilities, as it requires understanding user instructions, per- forming multi-turn tool use to gather information, and generating constraint-satisfying travel plans. Existing benchmarks fall into two paradigms: (1) Single-turn Explicit Instruction (Xie et al., 2024; Shao et al., 2024a; de la Rosa et al., 2024b; Singh et al., 2024), where the agent generates a complete travel plan from a single, explicit user request; and (2) Multi-turn Clarification (Zhang et al., 2024; Chen et al., 2024), where the agent interactively elicits user preferences through dialogue to refine the plan. However, as shown in Table 1, both paradigms require users to actively express their preferences and cannot address scenarios where users express their preferences implicitly. In con- trast, Behavior2Trip evaluates agents’ ability to (b) Preference-Driven User Behavior Generation ConstraintManualAutomatic (c) Chain of Action Annotation (d) Quality Assesment Toolset × N POI Database Preference Sample Format Authenticity Reasonableness Information Richness Personalization Toolset Attraction FlightHotel Restraunt Train WeatherPOI Database FlightSearch AttraSearch MapGuide Python TrainSearch WeaSearch RestSearch HotelSearch Constraint Extarct POI Category Behavior2Trip Filter Next Step Prompt Chain of Action Iterative Action GenerationAction with Query and Preference Action Enter. Think1 Action1 Obv1 Think2 Action2 Obv2 ... ... ... System LogPreferenceRaw ProfileProfile Behavior Trajectory： [CLICK] The Ritz-Carlton, Five Star [CLICK] Four Seasons Hotel, Five Star [CLICK] Orange Hotel, Three Star [FAVORITE] The Ritz-Carlton [BOOK] The Ritz-Carlton ...... Preferences: minimum of five-star hotels, ... Instruction：I’m planning a trip to Beijing and my budget is 5,000 RMB. (a) Sandbox Environment Construction Figure 2: The overall construction pipeline of Behavior2Trip, comprising four phases: (a) Sandbox Environment Construction builds a realistic platform with real-world POIs, tools, and constraints; (b) User Behavior Generation produces user trajectories across three difficulty levels; (c) Chain of Action Annotation collects agent trajectories with chain-of-action reasoning paths; and (d) Quality Assessment ensures data quality through two-stage automated and human verification. infer user preferences from behavior trajectories and generate preference-aligned travel plans, en- abling more realistic evaluation of personalized planning in real-world scenarios. 2.2 Travel Planning Agents To build travel planning agents, existing methods generally fall into two paradigms. (1) End-to-End LLM Agent (Xie et al., 2024; Chen et al., 2024; Zhang et al., 2024; Tang et al., 2024; Liu et al., 2026b; Lan et al., 2026), which relies on manually designed prompts to guide information retrieval and plan generation. However, such handcrafted workflows lack flexibility and struggle to general- ize across diverse user inputs. (2) Hybrid LLM- Symbolic Solver (Hao et al., 2024; Ju et al., 2024; de la Rosa et al., 2024b; Shao et al., 2024a), which improves logical consistency by incorporating sym- bolic solvers, but typically applies them only at the final planning stage, leaving the step-by-step information gathering process unguided. To ad- dress these limitations, we propose B2T-Agent, a reinforcement learning-based agent that elimi- nates handcrafted components and learns to interact with external tools and manage internal memory throughout the entire planning process. 3 Behavior2Trip 3.1 Problem Definition We define the user’s action trajectory asA = [a 1 ,a 2 ,...,a t ] , where eacha i denotes an action (e.g., click, favorite, book) at time stepi, im- plicitly encoding the user’s travel preferencesP. Given a natural-language travel requestI, the model queries the POI databaseDvia toolset Tto generate a personalized travel planR = Model(A,I,T ,D) , subject to two constraints:① Commonsense Constraint (C) (R ∈ C C ): the plan must be logically feasible and factually grounded;②User Preference Constraint (UPC) (R ∈ C UPC ): the plan must align with the user’s inferred preferences. 3.2 Data Collection As shown in Figure 2, this section describes the step-by-step process of constructing a diverse and realistic dataset, which includes four key stages: (a) Sandbox Environment Construction, (b) Preference-Driven User Behavior Generation, (c) Chain of Action Annotation, and (d) Quality As- sessment. 3.2.1 Sandbox Environment Construction To create a realistic and fully functional sandbox en- vironment, we establish a POI Database, a Toolset, and Constraints sourced from one of the largest Chinese online travel platforms. All data is fully anonymized with no user privacy information in- volved. POI Database We collect 80.6k real POIs from 24 popular Chinese travel cities, including Attrac- tion, Hotel, Restaurant, and Entertainment. Each POI has an average of 17.7 fields, including:① Statistic TrainTest EMHEMH Dataset Samples300030003000800800800 Trajectory Actions14.538.266.714.237.667.2 Preferences1.02.44.01.02.34.0 POIs10.327.247.810.126.948.0 Instruction Trip Duration4.94.95.04.94.95.0 Companions2.22.12.32.22.22.2 Table 2: Statistics of Behavior2Trip. E, M, and H represent Easy, Medium, and Hard levels, respectively. basic info (e.g., name, category),②spatiotem- poral data (e.g., location, hours, pricing),③user feedback (e.g., reviews, ratings). We also include 728.4k records on Flight, Train, and Weather, form- ing a comprehensive POI databaseD. See Ap- pendix A.1 for details. Toolset We design a fine-grained toolset where each POI category has a dedicated search tool sup- porting optional filtering fields such as location, price range, rating, and opening hours. We also implement MapGuide, which helps interpret spa- tial data, and Python for general-purpose comput- ing. In total, we develop 9 tools; details are in Appendix A.2. Constraints To enable fine-grained evaluation of travel plans while preserving real-world fidelity, we design a constraint set grounded in authentic online travel service scenarios. For C, we define four dimensions: Format ensures structural consis- tency, Authenticity verifies POI-related correctness, Reasonableness checks logical feasibility, and In- formation Richness reflects plan completeness. For UPC, we introduce the Personalization dimension to capture alignment with user preferences. Al- together, we define 32 constraints; details are in Appendix A.3. 3.2.2 User Behavior Generation To construct user trajectories that are both preference-grounded and behaviorally coherent, we first derive a structured preference schema from real logs, instantiate user profiles with controlled difficulty, and then use LLM role-playing to gener- ate long-term click–favorite–book sequences. Preference SchemaWe derive a structured pref- erence schema from anonymized user logs by ex- tracting 14 high-frequency attributes across 5 di- mensions. For each attribute, we calibrate its value range and POI-category mapping against real-log distributions, with details in Appendix A.4. User Profile We instantiate user profilesPat three difficulty levels to control preference com- plexity and behavioral noise. Easy includes one preference category without noise; Medium com- bines multiple categories with one noise behavior; and Hard covers all categories with a higher noise ratio. For each level, we sample attribute combina- tions from the schema and use GPT-4.1 to remove inconsistent profiles, with the prompt shown in Fig- ure 8. Behavior TrajectoryMotivated by prior studies showing that user check-in trajectories are shaped by sequential, spatio-temporal, and semantic/cate- gory patterns (Cheng et al., 2013; Tian et al., 2023; Wang et al., 2022), we model real user sessions as coherent travel-intent trajectories. To capture this structure, for each profileP, we construct a candi- date pool containing preference-matched POIs and a difficulty-controlled set of noise POIs. Gemini- 3-Flash is then prompted to role-play the user and generate coherent click→add_to_favorites→ book sequences (prompt shown in Figure 9). Noise POIs are restricted to click actions to simulate ex- ploratory browsing, while favorite and book ac- tions remain preference-aligned. Repeating this process across sessions yields long-term trajecto- riesA, which are paired with real travel instruc- tionsIfrom system logs to form(A,P,I)triplets. 3.2.3 Chain of Action Annotation After obtaining the(A,P,I)triplets, we anno- tate each instance with a Chain of Action (CoA) label that records the full reasoning trajectory— including tool calls and intermediate decisions— to provide step-level supervision for model train- ing. To generate these trajectories, we use Open- Manus (Liang et al., 2025) as the rollout har- ness, feedingAandIas inputs. Since uncon- strained rollouts tend to drift from the correct planning direction, after each(think, action, observation) cycle we manually inject a tailored Next Step Prompt (Figure 7) to steer the agent toward the next correct step, ensuring the resulting trajectory remains coherent and usable as a super- vision signal. 3.2.4 Quality Assessment To ensure dataset quality, we use a two-stage filter- ing process (Liu et al., 2020, 2021; Cheng et al., Preferences: Prefers spicy food; has visited Tiananmen Square. Interaction History: searched for top hotels and spicy food in Beijing. Travel Plan: Day 1 Accommodation, Park Plaza Beijing Wangfujing. (a) Structured Action Sequence Generation(b) Multi-component reward modeling (c) Loss Mask Penalty Format Check User Preference Constraint Commonsense Constraint Travel Plan Reward Total Reward Think Tool Mem T Res M Res Step 1 Think Step K Travel Plan ...... 1 1 1 1 0 Think Tool Call × N Mem Operation × M Tool Response × N Mem Response × M Step 1 Think Step K Travel Plan ...... Query Rollout 1 Rollout 2 ...... Reward Model Reference Model R 1 R 2 ...... A 1 A 2 ...... L 1 L 2 ...... Group Computation POI Database Internal Memory Policy Model Figure 3: Overview of the proposed B2T-Agent framework. B2T-Agent incorporates external tool invocation and internal memory management, optimized via Group Relative Policy Optimization with a multi-component reward function that rewards plan quality and penalizes invalid actions. 2026b). In the first stage, we utilized Deepseek-R1 to automatically score each instance on a 0-3 scale with respect to three aspects: trajectory complete- ness, reflection of user preferences, and constraint satisfaction (Figure 10). Only instances that re- ceived a score of 3 were retained. Subsequently, three independent annotators reviewed a random sample of 1000 instances, reporting no observable errors and achieving 94.6% agreement. 3.3 Data Statistics Table 2 summarizes dataset statistics across three difficulty levels. The full dataset comprises 11,400 instances (9,000 train / 2,400 test). With increas- ing difficulty, the number of actions, preference categories, and involved POIs all grow, making it progressively harder for models to infer user pref- erences from behavioral trajectories and generate constraint-satisfying plans. At each difficulty level, instructions span varied trip durations and group sizes, ensuring the benchmark captures a broad dis- tribution of real-world travel planning scenarios. 4 B2T-Agent In this section, we present B2T-Agent, an RL-based agent that supports external tool invocation and in- ternal memory management for personalized travel planning. As shown in Figure 3, B2T-Agent com- prises three key components: (a) Structured Action Sequence Generation (§4.1), (b) Multi-Component Reward Modeling (§4.2), and (c) Training with Loss Mask (§4.4). 4.1 Structured Action Sequence Generation To support iterative POI retrieval from the database and coherent context management across long- horizon planning, we define the agent’s rollout as a structured action sequence y: y : = (a 1 ,a 2 ,...,a k ,R) where each actiona i is one of the following types: External Tool UseT, Internal Memory Manage- mentM, or Chain-of-Thought ReasoningC. The sequence concludes with a final ResultR. External Tool Use The agent wraps each tool invocation in<tool_call>and receives the result in<tool_response>. Built on Qwen-Agent and following the Model Context Protocol 2 , it can ac- cess any tool in our sandbox. To improve efficiency, multiple tools can be called within a single reason- ing step. Internal Memory Management To mitigate context window saturation from accumulating tra- jectories and tool responses, the agent maintains an internal memory module organized as key-value pairs, accessible via<memory>. The model can read entries by key and write new entries to store user preferences, tool outputs, and interaction his- tory across planning steps. Reasoning and ResultTo infer user preferences from behavior trajectories and guide subsequent 2 https://w.anthropic.com/news/ model-context-protocol tool invocations, the model performs explicit rea- soning within<think>to analyze retrieved infor- mation and intermediate results. Once all neces- sary interactions are complete, it generates the final travel plan within <answer>. 4.2 Multi-Component Reward Modeling Given the structured action sequence defined above, we design a multi-component reward function to supervise both the process of tool invocation and the quality of the final travel plan, comprising an Incorrect Action Penalty and a Travel Plan Reward. Incorrect Action Penalty To improve the reli- ability and robustness of the execution process, we introduce a penalty for invalid tool or mem- ory invocations.An error is identified when <tool_response>or<memory_response>con- tains an explicit error message, upon which a penalty ofP action = −1is applied to the entire sequence, and P action = 0 otherwise. Travel Plan Reward To evaluate the quality of generated travel plans against both commonsense and user preference constraints, we design a two- stage gating reward. A format scoreR format ∈ 0, 1 first verifies structural correctness; only if passed, the Commonsense Constraint scoreR C and User Preference Constraint scoreR UPC are as- sessed, ensuring the model learns to produce well- formed outputs before optimizing for content qual- ity: R plan = R format · (R C + R UPC ) The total reward combines both components: R total = P action + R plan This design penalizes any action-level failure while rewarding high-quality, preference-aligned plans, jointly encouraging reliable execution and person- alized planning. 4.3 Policy Optimization via GRPO Inspired by recent advances in RL-based LLM training (Pan et al., 2025; Team, 2025; Wang et al., 2025a,b; Liu et al., 2026a), we adopt GRPO (Shao et al., 2024b) to optimize the policy model by max- imizing the following objective: J GRPO (θ) =E [y i G i=1 ∼π θ (Y|q)] 1 G G X i=1 1 |y i | |y i | X t=1 n min h r i,t (θ) ˆ A i,t , clip (r i,t (θ), 1− ε, 1 + ε) ˆ A i,t i − βD KL [π θ ||π ref ] with r i,t (θ) = π θ (y i,t |q,y i,<t ) π θ old (y i,t |q,y i,<t ) , whereπ θ andπ ref stand for the policy model and reference model respectively.D KL is the KL- divergence, scaled by a coefficientβ.qdenotes a query, andyrepresents a sequence of structured actions. 4.4 Training with Loss Mask During training, the rollout contains externally sourced tokens within<tool_response>and <memory_response> that are not generated by the model. Including them in the loss computation in- troduces noise, so we follow Jin et al. (2025) and apply a loss mask to exclude these segments from gradient updates, ensuring the model is optimized only on its own generated tokens. 5 Experiments 5.1 Experimental Setup BaselinesandModels Followingprior works (Xie et al., 2024; Cheng et al., 2025b), we adopt two baselines: ReAct (Yao et al., 2023) and SFT, alongside our proposed B2T-Agent, all with interactive reasoning where the model interacts with the environment to gather information. To ensure fair comparison, all methods are given the full user behavior trajectory. We evaluate open-source models including Qwen3-8/14/32B (Yang et al., 2025) and Deepseek-V3 (Liu et al., 2024), as well as frontier proprietary models including GPT-4o and GPT-4.1. Metrics Following Xie et al. (2024), we report the Delivery Rate (DR), which measures the pro- portion of plans that successfully produce a struc- turally valid travel itinerary. For constraint adher- ence, we compute the Micro Pass Rate (Micro PR) and Macro Pass Rate (Macro PR) separately for Commonsense Constraints (C) and User Prefer- ence Constraints (UPC), defined as: Micro PR = P p∈P P c∈C p 1 passed(c,p) P p∈P |C p | ModelMethod EasyMediumHard Avg. DRCCUPCFinalLLMDRCCUPCFinalLLMDRCCUPCFinalLLM Qwen3-8B ReAct 40.234.2/0.010.1/10.10.027.534.131.8/0.012.3/0.00.023.511.224.1/0.015.9/0.00.012.80.0 Qwen3-14B44.332.9/0.314.2/14.20.332.632.832.9/0.419.7/1.81.828.721.531.9/1.221.1/0.00.023.20.7 Qwen3-32B53.132.4/0.314.3/14.30.334.341.532.5/0.819.8/0.80.829.931.432.4/1.322.7/0.00.033.30.4 GPT-4o90.436.5/1.316.8/16.81.336.697.736.4/4.219.4/3.13.131.996.233.1/3.422.2/0.30.326.91.5 GPT-4.195.833.6/2.217.7/17.72.236.495.033.4/3.028.4/7.07.031.297.833.5/3.122.2/0.50.528.23.2 DeepSeek-V394.334.7/3.320.1/20.13.141.694.933.9/3.728.4/4.44.438.897.034.1/4.131.8/0.80.534.12.7 Qwen3-8B SFT 77.254.0/5.017.7/17.75.073.776.554.4/4.626.5/3.73.762.357.454.5/4.132.9/0.30.335.83.0 Qwen3-14B87.358.7/5.217.8/17.85.273.188.659.0/5.529.0/4.14.166.961.555.2/5.033.2/0.30.336.13.2 Qwen3-8B100.056.8/5.225.0/25.04.777.296.357.3/5.237.2/10.35.270.293.353.3/5.136.9/4.34.368.54.7 Qwen3-14B B2T-Agent 98.959.8/7.327.9/27.97.381.199.059.3/6.240.3/13.56.275.297.459.4/6.141.2/5.05.074.36.2 Table 3: Performance on Different Methods and Difficulties. C and UPC show Micro/Macro Pass Rate. Results are reported in percentage (%). The best and second-best results are marked in bold and underlined. Macro PR = P p∈P 1 passed(C p ,p) |P| For overall performance, we report the Final Pass Rate (Final), the percentage of plans satisfying all constraints, and the LLM Pass Rate (LLM) for flexible LLM-based evaluation. Implementation Details We adapt the ReAct baseline from TravelPlanner (Xie et al., 2024) to our expanded toolset and sandbox constraints, and conduct SFT using LLaMAFactory (Zheng et al., 2024). B2T-Agent is built on RLFactory (Simple- Efficient, 2025) with the prompt template shown in Figure 11. 5.2 Main Results Table 3 reports the main results and we highlight three key findings. Existing models struggle with Action-Aware Travel Planning. Even the strongest proprietary model, GPT-4.1, achieves only 0.5 Final PR on Hard tasks, underscoring the severity of the chal- lenge. This stems from a fundamental shift in what drives success: while Easy tasks can largely be solved by satisfying C, Medium and Hard tasks increasingly require alignment with UPC, a more personalized and dynamic requirement that current models consistently fail to capture. While LLM- based evaluation assigns relatively higher scores than rule-based Final due to its tolerance for out- put variations, both metrics confirm the same trend: UPC satisfaction remains the key bottleneck as task difficulty scales up. B2T-Agent achieves the best performance on Action-Aware Travel Planning.B2T-Agent con- sistently achieves the best performance across all difficulty levels, with B2T-Agent-14B leading all baselines on both Final PR and LLM PR across Easy, Medium, and Hard splits. Furthermore, this advantage becomes especially pronounced on harder tasks: on the Hard split, B2T-Agent-14B achieves 5.0 Final PR compared to 0.5 for the best ReAct baseline (GPT-4.1), and 74.3 LLM PR compared to 28.2. Notably, even B2T-Agent with Qwen3-8B surpasses GPT-4.1 in average Final PR (4.7 vs. 3.2), highlighting the efficiency of our approach. B2T-Agent Achieves Superior Performance on User Preference Constraints. Since C con- straints are shared across all queries, SFT can learn fixed patterns to improve C satisfaction; however, UPC constraints vary per user and require dynamic adaptation, which SFT fails to provide. In contrast, B2T-Agent leverages reward signals to drive the model toward personalized planning through inter- action, rather than relying on memorized response patterns. This translates to substantial UPC gains: on the Hard split, B2T-Agent-14B achieves 5.0 UPC Macro PR, compared to 0.3 for SFT-14B and 0.8 for the best ReAct baseline (DeepSeek-V3). 6 Analysis In this section, we present a comprehensive anal- ysis aimed at addressing the following research questions. RQ1: Does incorporating behavior trajectory really help with personalized travel plan- ning? (§6.1) RQ2: How does increasing trajec- tory complexity affect method performance? (§6.2) RQ3: How does each component affect the perfor- mance of B2T-Agent? (§6.3) RQ4: Can B2T-Agent generalize to other travel planning benchmarks? (§6.4) Figure 4: Impact of Trajectory and Clarification Turns on Final PR. Incorporating user trajectory sig- nificantly improves initial plan quality and reduces the need for clarification turns. 6.1 RQ1: Impact of Behavior Trajectory To investigate whether user behavior trajectory benefits personalized travel planning, we evaluate Qwen3-8B-B2T-Agent using only implicit user in- structions, without trajectory. Instead, GPT-4.1 is used to simulate user feedback, allowing the model to iteratively refine plans through clarification turns. As shown in Figure 4, with zero clarification turns, Final PR is nearly zero, indicating failure to align plans with user preferences. To match the per- formance of trajectory-augmented planning, Easy, Medium, and Hard tasks require approximately 2, 3, and 5 clarification turns, respectively. These results confirm that user trajectory substantially im- proves initial plan quality and reduces reliance on multi-turn clarification. 6.2 RQ2: Impact of Trajectory Complexity 204060 Actions 0 20 40 60 80 100 DR (%) 204060 Actions 0 1 2 3 4 5 6 Final PR (%) 204060 Actions 10 20 30 40 50 60 70 80 LLM PR (%) Qwen3-8B (ReAct) Qwen3-8B (SFT) Qwen3-8B (HTA) GPT-4.1 (ReAct) Figure 5: Relationship between the number of actions in trajectories and the performance of DR, Final PR, and LLM PR. We compare Qwen3-8B and GPT-4.1 across DR, Final PR, and LLM PR as trajectory complexity increases (measured by the number of actions). As shown in Figure 5, Qwen3-8B under SFT and ReAct shows sharp drops across all metrics, as longer trajectories introduce more user preferences and edge cases beyond their capacity. In contrast, GPT-4.1 maintains strong DR but suffers notable declines in Final PR and LLM PR. Qwen3-8B- B2T-Agent performs best overall, sustaining high DR and stable PR scores; reward-driven interac- tive training enables it to handle edge cases and capture user preference–POI relationships that SFT and prompting cannot cover. 6.3 RQ3: Ablation Study MethodDRFinal PRLLM PR B2T-Agent96.524.7571.98 - w/o. M97.46 (↑0.97%)4.21 (↓11.37%)64.24 (↓10.76%) - w/o. P86.70 (↓10.17%)4.32 (↓9.05%)65.22 (↓9.39%) - w/o. L78.18 (↓18.99%)3.79 (↓20.21%)57.97 (↓19.47%) Table 4: Ablation study on Qwen3-8B. We ablate each key component of B2T-Agent on Qwen3-8B to assess their individual contributions. As shown in Table 4, removing the memory mod- ule (w/o. Memory) slightly improves DR (+0.97%), as the model skips memory operations and favors faster responses. However, without memory, the model loses access to user preferences, causing Final PR and LLM PR to drop by 11.4% and 10.8%, respectively. Without the penalty term (w/o. Penalty), the model fails to correct inappropriate tool or memory actions, leading to repetitive loops and a 10.2% drop in DR. Finally, removing loss masking (w/o. Loss Mask) introduces noisy gradi- ents from non-answer tokens, causing the largest overall decline, with DR, Final PR, and LLM PR dropping by 19.0%, 20.2%, and 19.5%, respec- tively. 6.4 RQ4: Cross-Benchmark Evaluation MethodDRCMiCMaHMiHMaFPR Mistral-7B † 7.04.80.00.00.00.0 GPT-4-Turbo † 93.163.32.910.55.50.6 GPT-4.199.475.15.710.77.81.5 Qwen3-8B-B2T-Agent10076.825.026.510.29.0 Table 5: Performance on the TravelPlanner ( † results from the original paper). To validate the generality of B2T-Agent for travel planning, we further conduct a cross-benchmark evaluation on TravelPlanner (Xie et al., 2024). As shown in Table 5, Qwen3-8B-B2T-Agent consis- tently outperforms all baselines, with especially large gains on constraint satisfaction (CMa: 25.0 vs. 5.7) and overall plan quality (FPR: 9.0 vs. 1.5). These results confirm that B2T-Agent general- izes effectively beyond our benchmark to standard travel planning settings. 7 Conclusion In this paper, we identify a new task, Behavior- Aware Travel Planning, which infers user prefer- ences from behavioral trajectories to generate per- sonalized travel plans without requiring explicit in- put. To support this task, we present Behavior2Trip, a benchmark combining user behavior data, and B2T-Agent, a reinforcement learning framework that enables models to interact autonomously us- ing external tools and internal memory. Exten- sive experiments demonstrate that B2T-Agent ef- fectively captures user intent, significantly reducing the need for explicit user interaction and enabling more seamless travel planning experiences. Limitations While B2T-Agent achieves high-quality personal- ized planning by combining reward-driven multi- turn tool interaction with reinforcement learning, it incurs notable computational overhead from two main sources. First, during training, GRPO re- quires sampling multiple rollouts per query, which substantially increases computational cost. Sec- ond, at inference time, the multi-turn tool invoca- tions lead to high first-token latency. Reducing this overhead without compromising planning quality remains an important direction for future work. Ethics Statement POI and user-preference data can contain substan- tial sensitive information (e.g., consumers’ per- sonally identifiable information), making it essen- tial to protect user privacy from both the data and model perspectives. Specifically, as described in Section 3.2.1, we first remove all explicit person- ally identifiable information from the POI database by using a locally deployed GPT-oss-120B (Agar- wal et al., 2025) to redact user reviews, business phone numbers, and POI descriptions. We then conduct random manual audits to verify that no privacy-sensitive content remains. In addition, dur- ing preference extraction, we model only high-level attributes rather than individual users, thereby fur- ther reducing privacy risk and mitigating potential bias when learning the mapping between prefer- ences and POI categories. Acknowledgments Thanks for the insightful comments and feedback from the reviewers. This work was supported by the National Natural Science Foundation of China (No. 62406015). References Sandhini Agarwal, Lama Ahmad, Jason Ai, Sam Alt- man, Andy Applebaum, Edwin Arbus, Rahul K Arora, Yu Bai, Bowen Baker, Haiming Bao, and 1 others. 2025. gpt-oss-120b & gpt-oss-20b model card. arXiv preprint arXiv:2508.10925. Aili Chen, Xuyang Ge, Ziquan Fu, Yanghua Xiao, and Jiangjie Chen. 2024. Travelagent: An ai assistant for personalized travel planning. arXiv preprint arXiv:2409.08069. Xiaoqing Chen, Zhitao Li, Weike Pan, and Zhong Ming. 2023. A survey on multi-behavior sequential recom- mendation. arXiv preprint arXiv:2308.15701. Chen Cheng, Haiqin Yang, Michael R Lyu, and Irwin King. 2013. Where you like to go next: Succes- sive point-of-interest recommendation. In IJCAI, vol- ume 13, pages 2605–2611. Zhangtao Cheng, Yuhao Ma, Jian Lang, Kunpeng Zhang, Ting Zhong, Yong Wang, and Fan Zhou. 2025a. Generative thinking, corrective action: User- friendly composed image retrieval via automatic multi-agent collaboration. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2, pages 334–344. Zihao Cheng, Zeming Liu, Yingyu Shan, Xinyi Wang, Xiangrong Zhu, Yunpu Ma, Hongru Wang, Yuhang Guo, Wei Lin, and Yunhong Wang. 2026a. Mem2evolve: Towards self-evolving agents via co- evolutionary capability expansion and experience dis- tillation. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Vol- ume 1: Long Papers), pages 20784–20831. Zihao Cheng, Hongru Wang, Zeming Liu, Yuhang Guo, Yuanfang Guo, Yunhong Wang, and Haifeng Wang. 2025b. Toolspectrum: Towards personalized tool utilization for large language models. In Findings of the Association for Computational Linguistics: ACL 2025, pages 20679–20699. Zihao Cheng, Hongru Wang, Zeming Liu, Xinyi Wang, Xiangrong Zhu, Yuhang Guo, Wei Lin, Jeff Z. Pan, and Yunhong Wang. 2026b. Terminal-world: Scal- ing terminal-agent environments via agent skills. Preprint, arXiv:2605.20876. Tomas de la Rosa, Sriram Gopalakrishnan, Alberto Pozanco, Zhen Zeng, and Daniel Borrajo. 2024a. Trip-pal: Travel planning with guarantees by combin- ing large language models and automated planners. Preprint, arXiv:2406.10196. Tomas de la Rosa, Sriram Gopalakrishnan, Alberto Pozanco, Zhen Zeng, and Daniel Borrajo. 2024b. Trip-pal: Travel planning with guarantees by combin- ing large language models and automated planners. arXiv preprint arXiv:2406.10196. Bin Deng, Yizhe Feng, Zeming Liu, Qing Wei, Xian- grong Zhu, Shuai Chen, Yuanfang Guo, and Yunhong Wang. 2025. Retail: Towards real-world travel plan- ning for large language models. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 14881–14913. Fan Gao, Hongqiang Li, Zhilong Chen, Yunai Yi, Shi- hao Nie, Zihao Cheng, Zeming Liu, Yuanfang Guo, Shumin Liu, Qizhen Qin, and 1 others. 2025. A chemical autonomous robotic platform for end-to- end synthesis of nanoparticles. Nature Communica- tions, 16(1):7558. Yilun Hao, Yongchao Chen, Yang Zhang, and Chuchu Fan. 2024. Large language models can solve real- world planning rigorously with formal verification tools. arXiv preprint arXiv:2404.11891. Yingzhi He, Xiaohao Liu, An Zhang, Yunshan Ma, and Tat-Seng Chua. 2025. Llm2rec: Large language mod- els are powerful embedding models for sequential recommendation. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2, pages 896–907. Song Jiang, Da JU, Andrew Cohen, Sasha Mitts, Aaron Foss, Justine T Kao, Xian Li, and Yuandong Tian. 2024. Towards full delegation: Designing ideal agentic behaviors for travel planning. Preprint, arXiv:2411.13904. Bowen Jin, Hansi Zeng, Zhenrui Yue, Jinsung Yoon, Sercan Arik, Dong Wang, Hamed Zamani, and Jiawei Han. 2025. Search-r1: Training llms to reason and leverage search engines with reinforcement learning. Preprint, arXiv:2503.09516. Da Ju, Song Jiang, Andrew Cohen, Aaron Foss, Sasha Mitts, Arman Zharmagambetov, Brandon Amos, Xian Li, Justine Kao, Maryam Fazel-Zarandi, and 1 others. 2024. To the globe (ttg): Towards language- driven guaranteed travel planning. In Proceedings of the 2024 Conference on Empirical Methods in Nat- ural Language Processing: System Demonstrations, pages 240–249. Da JU, Song Jiang, Andrew Cohen, Aaron Foss, Sasha Mitts, Arman Zharmagambetov, Brandon Amos, Xian Li, Justine T Kao, Maryam Fazel-Zarandi, and Yuandong Tian. 2024. To the globe (ttg): To- wards language-driven guaranteed travel planning. Preprint, arXiv:2410.16456. Tianwei Lan, Jiaqi Wu, Zeming Liu, Zhaoxin Fan, Haifeng Wang, and Yuhang Guo. 2026. Peap: Proac- tive embodied action sequence planning with joint understanding of vision and audio perception. In Proceedings of the 64th Annual Meeting of the As- sociation for Computational Linguistics (Volume 1: Long Papers), pages 23118–23138. Xinbin Liang, Jinyu Xiang, Zhaoyang Yu, Jiayi Zhang, Sirui Hong, Sheng Fan, and Xiao Tang. 2025. Open- manus: An open-source framework for building gen- eral ai agents. Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, and 1 others. 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437. Jingjing Liu, Ziye Huang, Zihao Cheng, Zeming Liu, Jiahong Wu, Yuhang Guo, Kehai Chen, Yunhong Wang, and Haifeng Wang. 2026a. Docos: Towards proactive document-guided actions in gui agents. Preprint, arXiv:2605.18048. Jingjing Liu, Silin Li, Zeming Liu, Zihao Cheng, Yuhang Guo, Yuanfang Guo, Yunhong Wang, and Haifeng Wang. 2026b.Towards multi-language repository-level code generation: From-scratch to guided tasks. Neurocomputing, page 133204. Zeming Liu, Haifeng Wang, Zheng-Yu Niu, Hua Wu, and Wanxiang Che. 2021. Durecdial 2.0: A bilingual parallel corpus for conversational recommendation. In Proceedings of the 2021 Conference on Empiri- cal Methods in Natural Language Processing, pages 4335–4347. Zeming Liu, Haifeng Wang, Zheng-Yu Niu, Hua Wu, Wanxiang Che, and Ting Liu. 2020. Towards con- versational recommendation over multi-type dialogs. In Proceedings of the 58th annual meeting of the as- sociation for computational linguistics, pages 1036– 1049. Junyu Luo, Weizhi Zhang, Ye Yuan, Yusheng Zhao, Jun- wei Yang, Yiyang Gu, Bohan Wu, Binqi Chen, Ziyue Qiao, Qingqing Long, Rongcheng Tu, Xiao Luo, Wei Ju, Zhiping Xiao, Yifan Wang, Meng Xiao, Chenwu Liu, Jingyang Yuan, Shichang Zhang, and 7 others. 2025. Large language model agent: A survey on methodology, applications and challenges. Preprint, arXiv:2503.21460. Zhiyuan Ma, Jiayu Liu, Xianzhen Luo, Zhenya Huang, Qingfu Zhu, and Wanxiang Che. 2025. Advanc- ing tool-augmented large language models via meta- verification and reflection learning. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2, pages 2078–2089. Hang Ni, Fan Liu, Xinyu Ma, Lixin Su, Shuaiqiang Wang, Dawei Yin, Hui Xiong, and Hao Liu. 2025. Tp-rag: Benchmarking retrieval-augmented large lan- guage model agents for spatiotemporal-aware travel planning. Preprint, arXiv:2504.08694. Jiayi Pan, Junjie Zhang, Xingyao Wang, Lifan Yuan, Hao Peng, and Alane Suhr. 2025.Tinyzero. https://github.com/Jiayi-Pan/TinyZero. Accessed: 2025-01-24. Jie-Jing Shao, Xiao-Wen Yang, Bo-Wen Zhang, Baizhi Chen, Wen-Da Wei, Lan-Zhe Guo, and Yu-feng Li. 2024a. Chinatravel: A real-world benchmark for language agents in chinese travel planning. arXiv preprint arXiv:2412.13682. Jie-Jing Shao, Bo-Wen Zhang, Xiao-Wen Yang, Baizhi Chen, Si-Yu Han, Wen-Da Wei, Guohao Cai, Zhen- hua Dong, Lan-Zhe Guo, and Yu feng Li. 2025. Chinatravel: An open-ended benchmark for lan- guage agents in chinese travel planning. Preprint, arXiv:2412.13682. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, and 1 others. 2024b. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300. Simple-Efficient. 2025.RL-Factory: Train your Agent model via our easy and efficient frame- work.https://github.com/Simple-Efficient/ RL-Factory. Commit 216e841. Harmanpreet Singh, Nikhil Verma, Yixiao Wang, Man- asa Bharadwaj, Homa Fashandi, Kevin Ferreira, and Chul Lee. 2024. Personal large language model agents: A case study on tailored travel planning. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 486–514. Joykirat Singh, Raghav Magazine, Yash Pandya, and Akshay Nambi. 2025.Agentic reasoning and tool integration for llms via reinforcement learning. Preprint, arXiv:2505.01441. Yihong Tang, Zhaokai Wang, Ao Qu, Yihao Yan, Zhaofeng Wu, Dingyi Zhuang, Jushi Kai, Kebing Hou, Xiaotong Guo, Jinhua Zhao, and 1 others. 2024. Itinera: Integrating spatial optimization with large language models for open-domain urban itinerary planning. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 1413–1432. Kimi Team. 2025. Kimi k1.5: Scaling reinforcement learning with llms. Preprint, arXiv:2501.12599. Jing Tian, Zilin Zhao, and Zhiming Ding. 2023. Next point-of-interest recommendation based on joint mining of spatial–temporal and semantic sequen- tial patterns. ISPRS International Journal of Geo- Information, 12(7):297. Hongru Wang, Cheng Qian, Wanjun Zhong, Xiusi Chen, Jiahao Qiu, Shijue Huang, Bowen Jin, Mengdi Wang, Kam-Fai Wong, and Heng Ji. 2025a. Acting less is reasoning more! teaching model to act efficiently. Preprint, arXiv:2504.14870. Xican Wang, Yanheng Liu, Xu Zhou, Xueying Wang, and Zhaoqi Leng. 2022. A point-of-interest recom- mendation method exploiting sequential, category and geographical influence. ISPRS International Journal of Geo-Information, 11(2):80. Zhenhailong Wang, Xuehang Guo, Sofia Stoica, Haiyang Xu, Hongru Wang, Hyeonjeong Ha, Xiusi Chen, Yangyi Chen, Ming Yan, Fei Huang, and Heng Ji. 2025b. Perception-aware policy optimization for multimodal reasoning. Preprint, arXiv:2507.06448. Jian Xie, Kai Zhang, Jiangjie Chen, Tinghui Zhu, Renze Lou, Yuandong Tian, Yanghua Xiao, and Yu Su. 2024. Travelplanner: A benchmark for real- world planning with language agents. arXiv preprint arXiv:2402.01622. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025.Qwen3 technical report. arXiv preprint arXiv:2505.09388. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. React: Synergizing reasoning and acting in language models. In 11th International Conference on Learn- ing Representations, ICLR 2023. Mao Ye, Peifeng Yin, Wang-Chien Lee, and Dik-Lun Lee. 2011. Exploiting geographical influence for collaborative point-of-interest recommendation. In Proceedings of the 34th international ACM SIGIR conference on Research and development in Informa- tion Retrieval, pages 325–334. Siyu Yuan, Kaitao Song, Jiangjie Chen, Xu Tan, Dong- sheng Li, and Deqing Yang. 2025. EvoAgent: To- wards automatic multi-agent generation via evolu- tionary algorithms. In Proceedings of the 2025 Con- ference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 6192–6217, Albuquerque, New Mexico. Asso- ciation for Computational Linguistics. Shaokun Zhang, Yi Dong, Jieyu Zhang, Jan Kautz, Bryan Catanzaro, Andrew Tao, Qingyun Wu, Zhid- ing Yu, and Guilin Liu. 2025a. Nemotron-research- tool-n1: Exploring tool-using language models with reinforced reasoning. Preprint, arXiv:2505.00024. Xuan Zhang, Yang Deng, Zifeng Ren, See Kiong Ng, and Tat-Seng Chua. 2024. Ask-before-plan: Proac- tive language agents for real-world planning. In Find- ings of the Association for Computational Linguistics: EMNLP 2024, pages 10836–10863. Yuxiang Zhang, Yuqi Yang, Jiangming Shu, Xinyan Wen, and Jitao Sang. 2025b. Agent models: Inter- nalizing chain-of-action generation into reasoning models. Preprint, arXiv:2503.06580. Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. 2024. Llamafactory: Unified efficient fine-tuning of 100+ language models. In Proceedings of the 62nd Annual Meeting of the Association for Compu- tational Linguistics (Volume 3: System Demonstra- tions), Bangkok, Thailand. Association for Computa- tional Linguistics. A Behavior2Trip A.1 POI Database Details Table 6 presents all the real Points of Interest (POIs) information used in this paper. It includes the POI types, the number of POIs of each kind, and the names of fields associated with each type. This comprehensive dataset provides a realistic and di- verse sandbox environment for our research. A.2 Toolset Details Table 7 presents all the tools, including the tool names, descriptions, and their respective param- eters. A total of nine tools are provided, each supporting fine-grained combinations based on user preference information, enabling precise re- trieval of the most relevant data from the large POI database. A.3 Constraint Details Table 8 lists the constraints used to evaluate the quality of travel plans. A total of 32 constraints are categorized into five groups: Format (1), Authentic- ity (4), Reasonableness (15), Personalization (10), and Information Richness (2). Each constraint is derived from real-world business scenarios, provid- ing a realistic and effective basis for assessing the quality of a travel plan. A.4 Preference Details Table 9 presents the fields and value ranges related to travel preferences defined in user profiles, in- cluding Historical Travel Experiences, Hotel Pref- erences, Attraction Preferences, Dining Require- ments, and Transportation Preferences. Each field is closely aligned with POIs in the database, ensur- ing that user instructions with diverse preferences can be effectively addressed. B Additional Experimentss B.1 RQ5: Case Study As illustrated in Figure 6, we compare the per- formance of B2T-Agent and TravelPlanner ReAct when processing the same input. The ReAct model performs tool invocation in a strictly sequential manner, relying entirely on prompts to execute ac- tions step by step. As the number of turns increases, the dialogue context grows rapidly, making it dif- ficult for the model to retain previous interaction history and user preferences. This often leads to repeated tool usage and a lack of filtering based on user preferences. Such inefficiency not only increases the number of interaction turns but may also prevent the generation of a final travel plan due to the maximum turn limit. In contrast, B2T-Agent supports parallel tool in- vocation and incorporates a memory module to read and write critical user information. During the planning process, the model proactively ac- cesses stored user preferences and interaction his- tory, thereby improving tool invocation efficiency and generating travel plans that better align with user needs. B.2 RQ6: Error Analysis We perform a detailed error analysis on the two best-performing models from our main experiment and our proposed B2T-Agent. From each model’s output, we randomly sampled 30 generated trajec- tories and manually examined them (total of 180 samples), identifying and categorizing five classes of recurring errors. B.2.1 Constraint Violations Agents often generate travel plans that ignore real- world constraints, such as geographic feasibility, time budgets, and user preferences. While the out- puts usually satisfy format and completeness re- quirements, they fail to consider whether the recom- mended POIs are realistically visitable or aligned with user needs. As a result, most plans are not directly usable. B.2.2 Erroneous Tool or Memory Loops When interacting step-by-step with external tools or internal memory, the models sometimes gener- ate invalid tool call formats or incorrect parameters. Once an error occurs, the model does not revise its behavior based on feedback, leading to repeated failures and ultimately crashing the generation pro- cess. B.2.3 Hallucinated POIs Agents occasionally include POIs that do not exist in the database. This type of hallucination indi- cates a poor understanding of the model’s knowl- edge boundaries. Instead of retrieving POIs from verified data, the model relies on memorized or incorrect internal representations. B.2.4 Contextual Reasoning Limits Each generation involves extensive interaction with memory and environment, leading to long contex- tual sequences, often exceeding 20k tokens. The I’mplanningatriptoBeijingforthreedays,couldyougivemeatravelplan? UserInstruction BehaviorTrajectory [“action_type”:“click”,“timestamp”:“2024-06-05T19:51:49”,“poi_details”:“poi_type”:“Food& Beverage”,“name”:“JunkangYayuan Rooftop Garden Restaurant”,“subcategory”:“HunanCuisine”,“action_type”:“click”,“timestamp”:“2024-06- 05T19:51:50”,“poi_details”:“poi_type”:“Food& Beverage”,“name”:“OldHunan Cuisine (DenggaoRoad Branch)”,“subcategory”:“Hunan Cuisine”,“action_type”:“book”,“timestamp”:“2024-06-05T19:52:02”,“poi_details”:“poi_type”:“Food& Beverage”,“name”:“Authentic HuofuXiangtan Cuisine (Hexi WangfujingBranch)”,“subcategory”:“HunanCuisine”,“action_type”:“click”,“timestamp”:“2024-04- 23T21:53:52”,“poi_details”:“poi_type”:“Hotel”,“name”:“NanjingFengdaInternational Hotel”,“primary_category”:“Hotel”,“secondary_category”:“Five- Star/Luxury”,“tertiary_category”:“Luxury”,“action_type”:“book”,“timestamp”:“2024-04- 23T21:54:04”,“poi_details”:“poi_type”:“Hotel”,“name”:“BeijingBaihewangChaolinSongyuan Hotel”,“primary_category”:“Hotel”,“secondary_category”:“Five-Star/Luxury”,“tertiary_category”:“Luxury”,......] Hist-Travel-AgentTravelplannerReAct <|im_start|> assistant <think> Okay, the user has now requested a three-day trip to Beijing. First, I need to analyze the user’s preferences based on their behavior trajectory. It’s clear that the user prefers spicy food, and the hotel history shows only 5-star or above hotels. I will first record the user’s preferences and then search for dining, lodging, and attractions in Beijing. </think> <memory>write(key=‘preference’, value=‘restaurant’: ‘spicy’, ‘hotel_requirement’: ’5- star’)</memory> <tool_call>‘name’: ‘accommodationsearch’, ‘arguments’: [‘city’: ‘beijing’, ‘n’:5, ‘preference’:‘Star Rating Requirement’:’Five-Star/Luxury’]</tool_call> <tool_call>>‘name’: ‘restaurantsearch’, ‘arguments’: [‘city’: ‘beijing’, ‘n’:10, ‘preference’:‘Preferred Cuisine’:’HunanCuisine’]</tool_call> <tool_call>>‘name’: ‘attractionsearch’, ‘arguments’: [‘city’: ‘beijing’, ‘n’:10]</tool_call> <|im_emd|> <|im_start|>user < memory _response>Data saved for key ‘preference’.</ memory _response> <tool_response> ... </tool_response> <tool_response> ... </tool_response> <tool_response> ... </tool_response><|im_end|> ....... <|im_start|> assistant <think>Okay, the three tools were successfully called in the last step. We’l select The Forbidden City, JW Marriott Luxury Hotels, and Chongqing Spicy Hotpot as the POIs for Day 1. At the same time, I will use MapGuideto query the route.</think> <memory>write(key=‘day1’, value=‘accommodation’: ‘JW Marriott Luxury Hotels’, ‘attraction’:’TheForbidden City’, ‘restaurant’: ’Chongqing Spicy Hotpot’)</memory> <tool_call>>‘name’: ‘mapguide’, ‘arguments’: [‘origin’: ‘The Forbidden City’, ‘destination’: ‘JW Marriott Luxury Hotels’]</tool_call><|im_emd|> <|im_start|> user < memory _response>Data saved for key ‘day1’.</ memory _response> <tool_response> ... </tool_response><|im_end|> ....... <|im_start|> assistant <think>I need to check the POIs I’ve already collected and confirm the user’s preferences again...</think> <memory>read(key=’all’)</memory> <|im_start|> user < memory _response>‘restaurant’: ‘spicy’, ‘hotel_requirement’: ’5-star’, ’day1’: ‘accommodation’: ‘JW Marriott Luxury Hotels’, ‘attraction’:’TheForbidden City’, ‘restaurant’: ’Chongqing Spicy Hotpot’ ,......</ memory _response> <|im_end|> ....... <|im_start|> assistant <think>I now have enough information. From the memory response, I can see the user likes spicy food and prefers 5-star hotels. Let me generate the full plan for the user in one go.</think> <result>“Day 1”:“Lunch”:“name”:“ChongqingSpicy Hotpot”,“time”:“11:35– 13:00”,“price”:275,“transport”:“car”:“~30 minutes”,“metro”:“Line8 ~32 minutes”,“Attraction”:“name”:“TheForbidden City”,“time”:“14:00”,“price”:70,“transport”:“car”:“~26 minutes”,“metro”:“Line1 ~41 minutes”,“Accommodation”:“name”:“JWMarriott Luxury Hotels”,“price”:1275, ......</result> <|im_end|> Thought1: The user wants to plan a three-day trip to Beijing. According to the user's request, I will first analyze the user's preferences based on the behavior trajectory. The user likes spicy food and wants to stay in 5-star hotels. Let me first collect POI information on 5-star hotels in Beijing. Action1:‘name’: ‘accommodationsearch’, ‘arguments’: [‘city’: ‘beijing’, ‘n’:5, ‘preference’:‘Star Rating Requirement’:’Five-Star/Luxury’] Observation1: ... Thought2: Then, I will continue collecting restaurant information and pay attention to the user’s food preferences —the user prefers Hunan cuisine. Action2: ‘name’: ‘restaurantsearch’, ‘arguments’: [‘city’: ‘beijing’, ‘n’:10, ‘preference’:‘Preferred Cuisine’:’HunanCuisine’] Observation2: ... Thought3: I have already collected restaurant and accommodation information. Now I need to plan the first day’s attractions. Action3: ‘name’: ‘attractionsearch’, ‘arguments’: [‘city’: ‘beijing’, ‘n’:1] Observation3: ... Thought4: I already have enough information to create the plan for the first day. I will choose Chongqing Spicy Hotpot for lunch, and arrange The Forbidden City and JW Marriott Luxury Hotels as the attraction and accommodation. According to the user’s request, let’s continue planning the city transportation route. Action4: ‘name’: ‘mapguide’, ‘arguments’: [‘origin’: ‘the forbidden city’, ‘destination’: ‘JW Marriott Luxury Hotels’] Observation4: ... ...... Thought20: Now I will collect hotel information for the first day. Action20: ‘name’: ‘accommodationsearch’, ‘arguments’: [‘city’: ‘beijing’, ‘n’:5] Observation20: ... Thought21: I will continue collecting restaurant information. Action21: ‘name’: ‘restaurantsearch’, ‘arguments’: [‘city’: ‘beijing’, ‘n’:10, ‘preference’:‘Preferred Cuisine’:’HunanCuisine’] Observation21: ... Thought21: Continue searching for attractions in Beijing. Action21: ‘name’: ‘attractionsearch’, ‘arguments’: [‘city’: ‘beijing’, ‘n’:1] Observation21: ... ... Exceeded the maximum number of 25 turns, unable to provide the final plan. Forget interaction history and user preferences. Serial execution is inefficient and repeatedly collects already available information, resulting in exceeding the maximum number of turns. Figure 6: Case study comparing the reasoning trajectories of TravelPlanner ReAct and B2T-Agent. While the ReAct baseline suffers from context loss and redundant tool usage due to strictly sequential execution, B2T-Agent utilizes parallel tool invocation and a memory module to retrieve user preferences, ensuring efficient and accurate plan generation. POI TypeCountFeatures Restaurant29,604Name, Details, Address, Phone, Province, City, Location, Longitude, Lati- tude, Tertiary Category, Secondary Category, Primary Category, Business Hours, Rating Rank, Parking Information, Average Price, Average Rating, User Reviews Flights710,633Flight Number, Airline, Airline Code, Departure City, Arrival City, De- parture City Code, Arrival City Code, Departure Airport, Arrival Airport, Departure Airport Code, Arrival Airport Code, Departure Date, Arrival Date, Departure Time, Arrival Time, Flight Duration, Departure Terminal, Arrival Terminal, Aircraft Model, Cabin Type, Cabin Code, Adult Fare, Child Fare, Adult Total Price, Child Total Price, Adult Fuel Surcharge, Adult Airport Construction Fee, Discount, On-time Rate, Meal Availability, Codeshare Flight Hotels22,473Name, Details, Cover Image, Address, Phone, Province, City, District, Lon- gitude, Latitude, Tertiary Category, Secondary Category, Primary Category, Business Hours, Rating Rank, Parking Information, WiFi Availability, Av- erage Price, Average Rating, User Reviews, Room Information Entertainment20,250Name, Details, Cover Image, Address, Phone, Province, City, District, Longitude, Latitude, Tertiary Category, Secondary Category, Primary Category, Business Hours, Rating Rank, Average Price, Average Rating Trains10,371Departure City, Arrival City, Departure Station, Arrival Station, Train Number, Train Type, Travel Time, Travel Minutes, Departure Time, Arrival Time, Fare Attractions8,229Name, Details, Cover Image, Address, Phone, Province, City, Location, Longitude, Latitude, Category Name, Attraction Type, Custom Type, Open- ing Hours, Best Visiting Season, Recommended Visiting Duration, Average Price, User Reviews, Ticket Information Weather7,440City, Date, Max Temperature, Min Temperature, Weather, Wind Direction, Air Quality Table 6: Statistics of the collected real-world POI database. The dataset comprises 7 distinct categories totaling 809k records, with an average of 17.7 attribute features per category. models struggle to extract relevant signals from such complex contexts, which negatively impacts their ability to generate coherent and relevant plans. B.2.5 Formatting Errors Despite clear instruction prompts, the models some- times fail to follow the required format for travel plans. These formatting issues hinder both auto- matic evaluation and real-world usability, revealing limitations in instruction following and output con- sistency. C Prompt Template In this section, we present the prompts used in this study. Figures 7, 8, 9, 10, and 11 illustrate the Next Step Prompt, Profile Filtering, Behav- ior Trajectory Generation, Quality Control, and B2T-Agent, respectively. ToolDescriptionParameter AccommodationSearchAccommodation options in a speci- fied city are found, with support for preference-based filtering. Main Parameters: • city: str • n: int (optional) • preferences: dict (optional) Preference Keys: • star_rating: list of str • amenities: list of str AttractionSearchAttractions in a specified city are found, with support for preference-based filter- ing. Main Parameters: • city: str • n: int (optional) • preferences: dict (optional) Preference Keys: • attraction_type: list of str • walk_tolerance: list of str EntertainmentSearchEntertainment options in a specified city are found, with support for preference- based filtering. Main Parameters: • city: str • n: int (optional) • preferences: dict (optional) Preference Keys: • type: list of str FlightSearchFlight information is retrieved, with sup- port for preference-based filtering. Main Parameters: • origin: str • destination: str • departure_date: str • n: int (optional) • preferences: dict (optional) Preference Keys: • cabin_class: list of str MapGuideNavigation routes from a departure loca- tion to a destination are searched, includ- ing walking, driving, and public transit options, either within the same city or between different cities. • origin_cityName: str • origin: str • destination_cityName: str • destination: str RestaurantSearchDining options in a specified city are explored, with support for preference- based filtering. Main Parameters: • city: str • n: int (optional) • preferences: dict (optional) Preference Keys: • cuisine: list of str • price_level: list of str • environment: list of str TrainSearchTrain information is retrieved. • origin: str • destination: str • n: int (optional) WeatherSearchWeather information for a specified city and date is retrieved. • city: str • date: str (optional) • n: int (optional) PythonPython code is executed, with the ability to display outputs and handle errors. • code: str Table 7: The toolset definition comprising 9 specialized APIs. Each retrieval tool is equipped with fine-grained parameters and preference keys (e.g., cuisine, star_rating) to support precise, multi-criteria filtering of the POI database. CategoryConstraintDescription Format (1)JSON Format ValidationValidates overall JSON structure and required fields Authenticity (4) POI AuthenticityVerifies that Points of Interest exist in real city databases Major Transport AuthenticityConfirms availability of intercity transport like flights/trains Local Transport AuthenticityEnsures local transport durations are reasonable Business Hours Authenticity Checks that POI visits occur within operating hours Reasonableness (15) Complete Meal ArrangementEnsures three meals are arranged each day Non-duplicate RestaurantsNo repeated restaurant recommendations across the plan Reasonable Meal TimesMeals occur at appropriate times of day Daily Accommodation ArrangementA hotel must be arranged for each night Reasonable Accommodation TimeHotel check-in should be scheduled after 18:00 Avoid Midnight TravelNo transport activities between 23:00 and 06:00 No Attractions After Late Major Transport No sightseeing after arriving at a city after 18:00 Local Transport Time LimitTotal daily local transport should not exceed 2 hours Weather and City Info IncludedWeather and city background must be included for each destination At Least One AttractionEach day must feature at least one valid attraction Attraction Quantity LimitNo more than 4 attractions per day No Time ConflictsActivities should not overlap, with at least 30- minute gaps Reasonable Time RangeActivities must occur between 09:00 and 21:00 Inter-attraction Transport TimeTransport between attractions should not exceed 1 hour No Long Continuous GapsNo idle gaps longer than 2 hours during 09:00–21:00 Personalization (10) Hotel Star Level MatchingHotel meets user’s minimum star rating Hotel Amenities MatchingHotel has required facilities (e.g. WiFi, parking) Attraction Type MatchingAttractions align with user’s preferences Walking Tolerance MatchingPlan respects walking endurance limitations Cuisine Type MatchingMeals reflect preferred cuisine types Price Level MatchingAll items stay within specified price levels Dining Environment Matching Dining settings match user preferences (e.g. quiet, lively) Transport Mode MatchingIntercity transport matches preferred modes Seat Class MatchingMatches user’s preferred flight/train seat class Budget ConstraintTotal cost stays within user-defined budget Information Richness (2) Information SufficiencyEach activity includes all required fields Information RichnessExtra details (e.g. descriptions, tips) are included when possible Table 8: Taxonomy of constraints derived from real-world business scenarios. The system enforces a total of 32 distinct constraints spanning 5 categories (e.g., Authenticity, Reasonableness) to ensure the feasibility and quality of travel plans. FieldData TypeValue Historical Travel Experiences accommodationsarray (of object)- restaurantsarray (of object)- attractionsarray (of object)- transportationarray (of object)- Hotel Preferences star_requirementarray (of string) 2-Star & Below/Economy, 4-Star/Upscale, 3-Star/Comfort, 5-Star/Luxury, City Homestay essential_facilitiesarray (of string) Free WiFi, Private Bathroom, Parking, Gym, Swimming Pool, Restaurant Attraction Preferences preferred_attraction_typesarray (of string) Park/Square, Museum/Exhibition, Cultural/Historic Site, Cityscape, Natural Landscape, Comprehensive Scenic Area, Religious Site, Theme Park, Outdoor Adventure, Water Activities, Hot Spring/Wellness, Rural/Folk Village, Animal Watching, Ice/Snow Sports, Water Conservancy Facility walking_tolerancestringLow (1-3km), Medium (3-5km), High (5km+) Dining Requirements cuisine_preferencesarray (of string) Beverages, Snacks/Fast Food, Other Cuisines, Hot Pot, Western Food, Cantonese, Other Chinese Food, Shanghai/Zhejiang Cuisine, Sichuan Cuisine, Bakery/Desserts, Barbecue, Buffet, Northern Chinese Food, Hunan Cuisine, Seafood, Japanese/Korean Cuisine, Southeast Asian Cuisine price_levelstringBudget (30-80 CNY/person), Moderate (80-150 CNY/person), Generous (150-300 CNY/person), Luxury (300-500 CNY/person) dining_environmentstringQuiet, Scenic View, Family-friendly, Lively, Unique Decor Transportation Preferences intercity_transportstringHigh-speed Rail, Regular Train, Airplane, Long-distance Bus intracity_transportstringSubway, Bus, Taxi, Ride-sharing, Walking seat_classstringEconomy Class, Business Class, First Class Seat, Second Class Seat Table 9: The schema of the user preference profile derived from real-world travel data. It consists of 14 fine-grained attributes across 5 dimensions (e.g., Dining, Transportation), defining both historical records and specific constraints to enable personalized planning. Prompt Template for Next Step Prompt Review the user's goals, summarize the information already collected from past conversations, and then briefly think about what additional information is needed to create a travel guide as detailed as the template, before selecting multiple tools to execute. Notes: Do not aim too high. You have multi-turn problem-solving ability, so to ensure the completeness and accuracy of your output, you can solve only part of the problem each time. Keep your reasoning concise: output your reasoning in with no more than 100 words, focusing only on the most important information and avoiding repetitive details. You must use all the information obtained through tools to summarize your answer. In the section, do not guess dates, destination city attractions, or attraction details. In the first reasoning step, you can only call 3 tools to determine three pieces of information , and only after that can you execute other operations. If the city is different from the user's current location, you need to check for round-trip flights/trains. If the user hasn't specified a departure city, assume departure is from the local city, in which case no train/flight information is needed. You must end the task using <result> </result>. The <result> </result> tag must stand alone and cannot be combined with other tool calls. Figure 7: Prompt Template for Next Step Prompt. Prompt Template for Filtering Profile Evaluate whether the following persona is reasonable: Age: age Education Background: education Occupation: occupation Location: location Family Status: family_status Sightseeing Preferences: sightseeing_preference Food Preferences: food_preference Hotel Preferences: hotel_preference Transportation Preferences: transportation_preference Evaluation criteria include but are not limited to: 1. Whether the age matches the education level (e.g., a 20-year-old PhD is unlikely; a 40-year- old with only elementary education is rare) 2. Whether the education background aligns with the occupation (e.g., someone with only a primary school education being a surgeon is unreasonable; a master's degree holder in a research or senior management role makes more sense) 3. Whether the age aligns with the family status (e.g., an 18-year-old being married with children is uncommon; a 60-year-old single person with no children may require justification) 4. Whether the occupation logically fits the location (e.g., a tech product manager is usually based in major cities; a full-time fashion influencer in a rural town would need context) 5. Whether the income level inferred from the occupation matches the stated preferences (e.g., someone earning a low income preferring luxury hotels and Michelin restaurants is inconsistent) 6. Whether the transportation preferences match the age or financial condition (e.g., a 70-year -old backpacking solo on a tight budget is uncommon; older people tend to prefer comfort and convenience) 7. Whether sightseeing/food/hotel preferences align with the overall persona (e.g., a highly educated researcher preferring trendy Instagram spots may seem off; a low-income person favoring ultra-luxury hotels may be unrealistic) 8. Whether the overall persona is logically coherent (e.g., a culturally inclined person should show that interest in their dining, travel, or accommodation preferences) Please respond strictly in the following JSON format and do not add any extra content: "reasonable": true/false, "reason": "specific reason" Output: output Figure 8: Prompt Template for Filtering Profile. Prompt Template for Behavior Trajectory Generation # ROLE You are role-playing a real user on a Chinese online travel platform (e.g., Ctrip / Fliggy). Your task is to generate a single-session behavior trajectory that authentically reflects the given user profile, as if you were genuinely browsing and exploring the platform during one visit. # USER PROFILE user_profile # SESSION CONTEXT Current session number: session_id Previous session history (for cross-session continuity): previous_sessions # CANDIDATE POI POOL ## Preference-Matched POIs (aligned with the user's profile): matched_pois ## Noise POIs (NOT aligned with the user's profile): noise_pois # ACTION SPACE - "click" : The user is browsing or curious. Low commitment. - "add_to_favorites" : The user finds the POI appealing and saves it for later. - "book" : The user commits to the POI. # BEHAVIORAL RULES 1. Preference Grounding: "add_to_favorites" and "book" MUST use Preference-Matched POIs only. Noise POIs MUST only receive "click" actions. 2. Behavioral Funnel: click -> add_to_favorites -> book. A POI must be clicked before it is favorited; favorited before it is booked. Most POIs stop at click. 3. Sequential Coherence: - Within session: consecutive actions should be topically related. - Cross-session: if previous sessions are provided, continue the established preference trajectory; do not re-book already-booked POIs; early sessions reflect broad exploration, later sessions narrow toward commitment. 4. Spatio-Temporal Consistency: POIs in one session should be geographically plausible. Decide transportation and accommodation before attractions and dining. 5. Semantic / Category Diversity: cover multiple POI categories per trajectory. 6. Noise Behavior: noise clicks should be sporadically interleaved, not clustered. 7. Realism Constraints: do NOT invent POIs outside the candidate pool. Do NOT mention preferences explicitly. Do NOT include reasoning or commentary. # OUTPUT FORMAT Return a JSON array for THIS SESSION ONLY. Each record: [ "action_type": "click" | "add_to_favorites" | "book", "poi_id": "<string from candidate pool>", "poi_name": "<string from candidate pool>" , ... ] Timestamps and full POI details are added automatically in post-processing. Figure 9: Prompt Template for Behavior Trajectory Generation. Prompt Template for Quality Control You are a professional travel-user behavior data-quality analyst. Your task is to evaluate the given user data. You must imitate the example below, score each of the three dimensions--'Completeness of Trajectory Information','Clarity of Preference Reflection in the Trajectory', and'Degree to Which the Plan Meets Constraints'--with either 0 or 1 point, and provide a structured JSON output. The output should include the scores, your reasoning, and specific instructions that can be used to automate data modifications. Scoring Criteria 1. Completeness of Trajectory Information - 1 point (Meets): The trajectory contains a clear behavior chain from browsing/exploration to final decision (e.g., multiple behaviors such as clicks, favorites, bookings/orders) and ends with a definitive decision point. - 0 points (Does Not Meet): The trajectory is overly single-patterned (e.g., only a few clicks), lacks a final decision step, or the behavior chain is broken/illogical. 2. Clarity of Preference Reflection in the Trajectory - 1 point (Meets): The user's key behaviors (such as favoriting, booking) are highly consistent with the preferences stated in persona scenic spot preference. - 0 points (Does Not Meet): The user's key behaviors conflict with or are completely unrelated to the stated preferences. 3. Degree to Which the Plan Meets Constraints - 1 point (Meets): The generated travel plan satisfies all constraints. - 0 points (Does Not Meet): The generated travel plan clearly conflicts with or is completely unrelated to the constraints. Output Format You must strictly output according to the JSON structure shown in the example and do not add any explanatory text outside the JSON object. Predefined Operation-Type Codes - No action required: Data quality is high; no modification is needed. - Correct user preference: When behavior conflicts with declared preferences, adjust the preferences in the persona based on behavioral history. - Mark incomplete trajectory: When the trajectory lacks a final decision step, mark the data so it can be down-weighted or filtered in downstream tasks. - Mark for deletion: When data quality is too low to fix, recommend deleting the entire record. Input: input Output: output Figure 10: Prompt Template for Quality Control. Prompt Template for B2T-Agent You are a seasoned travel planner with extensive experience in generating personalized plans. Your core task is to deeply understand the user's requirements, gather information by calling tools, and ultimately design a thorough and meticulous travel plan. You possess an internal dictionary-structured memory module for storing and retrieving key information. Use <memory></memory> to wrap any operation that reads or writes session memory using "action": "read" or "write" with a "key" (and optional "value" for write). After gathering all necessary information and crafting the complete plan, please wrap the final itinerary in JSON format between the "<answer>" and "</answer>" tags. Note: The delivered travel plan must be complete, covering every day of the trip. Please note that the final output should be a JSON list (array) that includes all travel days. Each day is an independent JSON object that follows the same internal structure. # Tools "name": "AccommodationSearch", "parameters": "type": "object", "properties": "city": "title ": "City", "type": "string", "n": "default": 5, "title": "N", "type": "integer", " preferences": "additionalProperties": true, "default": null, "title": "Preferences", "type": " object", "required": ["city"] "name": "RestaurantSearch", "parameters": "type": "object", "properties": "city": "title": "City", "type": "string", "n": "default": 5, "title": "N", "type": "integer", "preferences": "additionalProperties": true, "default": null, "title": "Preferences", "type": "object", " required": ["city"] "name": "FlightSearch", "parameters": "type": "object", "properties": "origin": "title": " Origin", "type": "string", "destination": "title": "Destination", "type": "string", " departure_date": "title": "Departure Date", "type": "string", "n": "default": 5, "title": "N ", "type": "integer", "preferences": "additionalProperties": true, "default": null, "title": "Preferences", "type": "object", "required": ["origin", "destination", "departure_date"] "name": "TrainSearch", "parameters": "type": "object", "properties": "origin": "title": " Origin", "type": "string", "destination": "title": "Destination", "type": "string", "n": " default": 5, "title": "N", "type": "integer", "required": ["origin", "destination"] "name": "WeatherSearch", "parameters": "type": "object", "properties": "city": "title": " City", "type": "string", "date": "default": null, "title": "Date", "type": "string", "n": " default": 5, "title": "N", "type": "integer", "required": ["city"] "name": "EntertainmentSearch", "parameters": "type": "object", "properties": "city": "title ": "City", "type": "string", "n": "default": 5, "title": "N", "type": "integer", " preferences": "additionalProperties": true, "default": null, "title": "Preferences", "type": " object", "required": ["city"] "name": "AttractionSearch", "parameters": "type": "object", "properties": "city": "title": "City", "type": "string", "n": "default": 5, "title": "N", "type": "integer", "preferences": "additionalProperties": true, "default": null, "title": "Preferences", "type": "object", " required": ["city"] "name": "MapGuide", "parameters": "type": "object", "properties": "origin_cityName": "title ": "Origin Cityname", "type": "string", "origin": "title": "Origin", "type": "string", " destination_cityName": "title": "destination_cityname", "type": "string", "destination": " title": "Destination", "type": "string", "required": ["origin_cityName", "origin", " destination_cityName", "destination"] Behavior trajectory: behavior trajectory Instruction: instruction Figure 11: Prompt Template for B2T-Agent.