Source-linked AI summary
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
TL;DR
Existing travel-planning agents depend on explicit or iterative user input, overlooking behavioral signals that encode preferences. The paper introduces behavior-aware planning with the Behavior2Trip benchmark and B2T-Agent, whose results show severe task difficulty, strong baseline performance, and generalization to TravelPlanner.
Problem
Existing agents overlook behavioral signals from past interactions, relying on active user input that increases interaction burden and limits plan personalization.
Method
The paper introduces Behavior-Aware Travel Planning, the Behavior2Trip benchmark, and B2T-Agent, an RL agent using external tools and internal memory.
Results
B2T-Agent built on Qwen3-8B significantly outperforms all baselines and also outperforms GPT-4.1 on TravelPlanner, while GPT-4.1 reaches only 0.5% full-constraint pass rate on hardest tasks.
Takeaways & Limitations
Behavior2Trip enables evaluation of preference-aligned planning from implicit behavior trajectories, and B2T-Agent reduces reliance on explicit user interaction within the reported scope.
Takeaways & Limitations
B2T-Agent incurs substantial training cost from multiple GRPO rollouts and high inference-time first-token latency from multi-turn tool invocation.
Abstract
from arXiv · showhide
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 https://github.com/BUAA-IRIP-LLM/Behavior2Trip
1 Introduction
Existing travel agents require explicit or iterative user input, overlooking behavioral signals that encode preferences. Behavior2Trip introduces behavior-aware planning and B2T-Agent, with experiments showing severe task difficulty and strong benchmark performance.
- Motivation: Existing agents either collect preferences in one explicit turn or elicit them through multi-turn clarification, overlooking signals from clicks, favorites, and bookings.This reliance on active input increases interaction burden and limits plan personalization.
- New Task: Behavior-Aware Travel Planning infers user preferences directly from past behaviors and generates personalized travel plans without requiring active user input.The task targets implicit instructions rather than explicit or iterative preference expression.
- Benchmark: Behavior2Trip contains 11,400 instances grounded in real user data, averaging 39.8 behaviors across 14 attributes and 5 preference dimensions.Its sandbox includes 80.6k real-world POIs, and trajectories undergo two-stage quality control.
- Agent: B2T-Agent is a reinforcement learning-based agent with structured external tool invocation and internal memory management for behavior-aware travel planning.Its structured action space supports preference-aligned retrieval and memory management during long-horizon planning.
- Results: 0.5%: GPT-4.1’s full-constraint pass rate on the hardest tasks, while B2T-Agent built on Qwen3-8B significantly outperforms all baselines and generalizes to TravelPlanner.These results underscore the substantial challenge of behavior-aware travel planning.
2 Related Work
Existing travel benchmarks focus on explicit requests or clarification dialogue, requiring users to actively express preferences. Behavior2Trip instead evaluates preference inference from behavior trajectories for more realistic personalized-planning scenarios.
- Travel Planning Benchmarks: Existing travel benchmarks use either one-shot explicit requests or multi-turn clarification to obtain user preferences.Both paradigms require active preference expression and do not address implicit preferences.
- Benchmark Construction: The benchmark is constructed through sandbox environment construction, preference-driven behavior generation, chain-of-action annotation, and quality assessment.Its pipeline includes realistic POIs, tools, constraints, generated trajectories, reasoning paths, and two-stage verification.
- Travel Planning Benchmarks: Behavior2Trip evaluates agents’ ability to infer user preferences from behavior trajectories and generate preference-aligned travel plans.This supports evaluation of personalized planning when preferences are expressed implicitly.
- Travel Planning Agents: End-to-end LLM agents rely on handcrafted prompts for retrieval and generation, while hybrid LLM-symbolic solvers incorporate symbolic methods to improve logical consistency.The supplied related-work passage notes that handcrafted workflows lack flexibility and struggle across diverse inputs.
3 Behavior2Trip
Behavior2Trip formalizes behavior-aware travel planning and builds a realistic benchmark from anonymized travel-platform data. It combines structured user preferences, behavior trajectories, tools, constraints, and quality-controlled action traces across difficulty levels.
- Problem Definition: Behavior-Aware Travel Planning infers travel preferences from action trajectories and generates a plan satisfying commonsense and user-preference constraints.The formulation maps actions such as clicks, favorites, and bookings to implicit preferences, then uses a POI database and toolset to produce the plan.
- Sandbox Environment Construction: The sandbox includes 80.6k real POIs, 728.4k flight, train, and weather records, and 9 tools supporting category-specific retrieval and spatial interpretation.POIs cover attractions, hotels, restaurants, and entertainment, with fields for basic, spatiotemporal, and user-feedback information.
- Sandbox Environment Construction: The benchmark defines 32 constraints spanning format, authenticity, reasonableness, information richness, and personalization.The first four dimensions evaluate commonsense constraints, while personalization evaluates alignment with inferred user preferences.
- Preference-Driven User Behavior Generation: User profiles use Easy, Medium, and Hard difficulty levels that increase preference complexity and behavioral noise.Easy contains one category without noise; Medium combines categories with one noise behavior; Hard covers all categories with a higher noise ratio.
- Dataset Statistics: The full dataset has 9,000 training and 2,400 test instances, while increasing difficulty raises the numbers of actions, preference categories, and involved POIs.Trip durations and group sizes vary at every difficulty level.
4 B2T-Agent
B2T-Agent performs long-horizon travel planning through structured actions that combine tool use, memory, reasoning, and a final result. Its training jointly rewards plan quality, penalizes invalid operations, optimizes with GRPO, and masks externally sourced tokens.
- Structured Action Sequence Generation: B2T-Agent represents each rollout as structured actions followed by a final result, supporting external tools, internal memory, and chain-of-thought reasoning.The sequence is designed for iterative POI retrieval and coherent context management during long-horizon planning.
- Structured Action Sequence Generation: The agent uses dedicated tool calls and key-value memory to retrieve information and preserve user preferences, tool outputs, and interaction history.Memory operations help mitigate context-window saturation from accumulating trajectories and tool responses.
- Multi-Component Reward Modeling: The reward combines an incorrect-action penalty with a gated travel-plan reward for structural correctness, commonsense constraints, and user-preference constraints.The format score must pass before commonsense and preference scores are assessed.
- Multi-Component Reward Modeling: Invalid tool or memory invocations receive Paction = −1, while valid sequences receive Paction = 0.An invalid action is identified from an explicit error message in a tool or memory response.
- Policy Optimization via GRPO: GRPO optimizes the policy model against a reference model while regularizing their divergence with a β-scaled KL term.The query is q, and y denotes a structured action sequence.
- Training with Loss Mask: Loss masking excludes externally sourced tool and memory-response tokens from gradient updates, optimizing only model-generated tokens.The mask is intended to avoid noise from tokens not generated by the model.
5 Experiments
The experiments compare interactive ReAct, SFT, and B2T-Agent systems across travel-planning models and constraint metrics. Results show severe difficulty on harder tasks, while B2T-Agent leads performance, especially for user-preference constraints.
- Experimental Setup: All methods receive full user behavior trajectories and interactively use the environment under a common evaluation setup.The comparison includes ReAct, SFT, and B2T-Agent with Qwen3, Deepseek-V3, GPT-4o, and GPT-4.1.
- Experimental Setup: The evaluation reports Delivery Rate, Micro/Macro Pass Rate for commonsense and user-preference constraints, Final Pass Rate, and LLM Pass Rate.Final measures plans satisfying all constraints, while LLM evaluates flexible plan quality.
- Main Results: 0.5 Final PR is achieved by GPT-4.1 on Hard tasks, showing that existing models struggle as user-preference constraints become more demanding.The paper identifies user-preference constraint satisfaction as the key bottleneck as difficulty increases.
- Main Results: 5.0 Final PR and 74.3 LLM PR are achieved by B2T-Agent-14B on Hard tasks, versus 0.5 and 28.2 for GPT-4.1, respectively.B2T-Agent-14B leads both metrics across Easy, Medium, and Hard splits.
- Main Results: 5.0 UPC Macro PR is achieved by B2T-Agent-14B on Hard tasks, compared with 0.3 for SFT-14B and 0.8 for ReAct with DeepSeek-V3.The result reflects the difficulty of dynamically adapting to user-specific preferences rather than fixed commonsense constraints.
6 Analysis
The analysis tests trajectory benefits, trajectory-complexity effects, component contributions, and cross-benchmark generalization. Behavior trajectories reduce clarification needs, B2T-Agent remains more robust as trajectories grow, its components materially affect performance, and it transfers to TravelPlanner.
- 6.1 RQ1: Impact of Behavior Trajectory: Zero clarification turns produce nearly zero Final PR, while matching trajectory-augmented planning requires approximately 2, 3, and 5 turns on Easy, Medium, and Hard tasks.The results indicate that trajectories improve initial plans and reduce reliance on clarification dialogue.
- 6.2 RQ2: Impact of Trajectory Complexity: Longer trajectories cause sharp metric drops for Qwen3-8B under SFT and ReAct, while GPT-4.1 retains DR but declines in Final PR and LLM PR.The comparison measures DR, Final PR, and LLM PR as the number of trajectory actions increases.
- 6.2 RQ2: Impact of Trajectory Complexity: B2T-Agent performs best overall as trajectory complexity increases, sustaining high DR and stable pass-rate scores.The paper attributes this robustness to reward-driven interaction handling edge cases and user-preference–POI relationships.
- 6.3 RQ3: Ablation Study: Removing memory lowers Final PR and LLM PR by 11.4% and 10.8%, while removing the penalty term lowers DR by 10.2%.Without memory, the agent loses access to user preferences; without penalties, inappropriate actions produce repetitive loops.
- 6.3 RQ3: Ablation Study: Removing loss masking produces the largest ablation decline, reducing DR, Final PR, and LLM PR by 19.0%, 20.2%, and 19.5%, respectively.The paper links this decline to noisy gradients from non-answer tokens.
- 6.4 RQ4: Cross-Benchmark Generalization: On TravelPlanner, Qwen3-8B-B2T-Agent reaches 25.0 CMa and 9.0 FPR, compared with 5.7 and 1.5 for the comparison results.The gains are reported for constraint satisfaction and overall plan quality.
7 Conclusion
The paper introduces Behavior-Aware Travel Planning, Behavior2Trip, and B2T-Agent to infer preferences from behavioral trajectories and generate personalized plans without explicit input. Its experiments report reduced reliance on explicit interaction and more seamless travel planning.
- Conclusion: Behavior2Trip combines user behavior data with B2T-Agent, a reinforcement learning framework using external tools and internal memory to capture user intent.The framework supports autonomous interaction for personalized travel planning.
- Conclusion: B2T-Agent reduces the need for explicit user interaction while enabling more seamless travel planning experiences.This conclusion follows the paper’s reported effectiveness in capturing user intent.
Limitations
B2T-Agent’s reward-driven multi-turn tool interaction and reinforcement learning incur notable computational overhead. The costs arise during both training and inference.
- Limitations: GRPO training requires multiple rollouts per query, substantially increasing computational cost.This is one of the two overhead sources identified by the authors.
- Limitations: Multi-turn tool invocations produce high first-token latency during inference.Reducing this overhead without compromising planning quality remains future work.
Ethics Statement
The study addresses privacy risks in POI and user-preference data by removing explicit personally identifiable information from the POI database before further processing.
- Explicit personally identifiable information is removed from the POI database before subsequent processing.A locally deployed GPT-oss-120B redacts user reviews, business phone numbers, and POI descriptions.
A.1 POI Database Details
The study builds a realistic travel-planning sandbox from diverse POI data, specialized retrieval tools, user-preference fields, and real-world evaluation constraints. The case study contrasts sequential tool use with B2T-Agent’s parallel invocation and memory-based preference retrieval.
- POI Database: The sandbox includes POI types, counts, and associated fields to provide realistic and diverse travel data.The POI information is organized by type and includes the number of POIs and fields for each type.
- Toolset: Nine tools support fine-grained combinations of user-preference information for precise POI retrieval.Each tool has a name, description, and parameter set for querying the large POI database.
- Evaluation Constraints: Thirty-two constraints cover Format, Authenticity, Reasonableness, Personalization, and Information Richness.The constraints are derived from real-world business scenarios and assess travel-plan quality and preference alignment.
- User Preference Profiles: User profiles define travel preferences across five dimensions and 14 attributes aligned with POIs in the database.The fields include historical travel experiences, hotels, attractions, dining, and transportation.
- Agent Operation: B2T-Agent combines parallel tool invocation with memory access to improve retrieval efficiency and preference alignment.Compared with the sequential ReAct baseline, its memory module reads and writes critical user information during planning.
B.2 RQ6: Error Analysis
The error analysis examines 180 sampled trajectories and identifies recurring failures in constraint satisfaction, tool interaction, POI validity, context processing, and output formatting. A case study contrasts sequential execution with B2T-Agent’s parallel tool use and memory module.
- Analysis Setup: 180 sampled trajectories were manually examined, revealing five recurring classes of errors.The samples came from the two best-performing main-experiment models and B2T-Agent, with 30 trajectories sampled per model.
- Constraint Failures: Most plans ignore geographic feasibility, time budgets, or user preferences despite satisfying format and completeness requirements.The resulting plans are often not directly usable because recommended POIs may be unrealistic to visit or misaligned with user needs.
- Tool Interaction Failures: Invalid tool-call formats and parameters can trigger repeated failures because models do not revise their behavior after feedback.These errors occur during step-by-step interaction with external tools or internal memory and can ultimately crash generation.
- POI Validity Errors: Some agents hallucinate POIs absent from the database instead of retrieving them from verified data.This reflects reliance on memorized or incorrect internal representations rather than the available database.
- Context Processing: Contexts often exceed 20k tokens, making it difficult for models to extract relevant signals and generate coherent, relevant plans.Extensive interaction with memory and the environment produces these long contextual sequences.
- Output Formatting: Formatting failures persist despite explicit prompts, hindering automatic evaluation and real-world usability.The errors expose limitations in instruction following and output consistency.