TriFlow: A Progressive Multi-Agent Framework for Intelligent Trip Planning

TriFlow: A Progressive Multi-Agent Framework for Intelligent Trip Planning Yuxing Chen The University of Sydney Sydney, Australia yche0009@uni.sydney.edu.au Basem Suleiman∗† University of New South Wales Sydney, Australia b.suleiman@unsw.edu.au Qifan Chen The University of Sydney Sydney, Australia qifan.chen@sydney.edu.au Abstract Real-world trip planning requires transforming open-ended user requests into executable itineraries under strict spatial, temporal, and budgetary constraints while aligning with user preferences. Existing LLM-based agents struggle with constraint satisfaction, tool coordination, and efficiency, often producing infeasible or costly plans. To address these limitations, we present TriFlow, a progressive multi-agent framework that unifies structured reason- ing and language-based flexibility through a three-stage pipeline of retrieval, planning, and governance. By this design, TriFlow progres- sively narrows the search space, assembles constraint-consistent itineraries via rule–LLM collaboration, and performs bounded it- erative refinement to ensure global feasibility and personalisation. Evaluations on TravelPlanner and TripTailor benchmarks demon- strated state-of-the-art results, achieving 91.1% and 97% final pass rates, respectively, with over 10 × runtime efficiency improvement compared to current SOTA. CCS Concepts • Information systems →Information systems applications; Decision support systems. Keywords Agent System, Trip Planning, Route Recommendation ACM Reference Format: Yuxing Chen, Basem Suleiman, and Qifan Chen. 2026. TriFlow: A Progressive Multi-Agent Framework for Intelligent Trip Planning. In Proceedings of Proceedings of the ACM Web Conference 2026 (WWW Companion ’26). ACM, New York, NY, USA, 4 pages. https://doi.org/XXXXXXX.XXXXXXX 1 Introduction Real-world trip planning is a complex task that combines natu- ral language understanding, spatiotemporal constraint satisfaction, and multi-objective optimisation. Users often describe their travel goals in open-ended and ambiguous language (e.g., “fun places,” ∗Basem Suleiman is the corresponding author. †Basem Suleiman also with The University of Sydney Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org. WWW Companion ’26, Abu Dhabi, United Arab Emirates © 2026 Copyright held by the owner/author(s). Publication rights licensed to ACM. ACM ISBN 978-1-4503-XXXX-X/26/04 https://doi.org/XXXXXXX.XXXXXXX Retrieve Plan Govern User Query Data Space Query Decomposition Final Itinerary Retrieval Workflow Persistent Query Metadata Task-related Data Itinerary Skeleton Construction Plan Suggestion Suggestion Validation Agent Governance Constraint Checking Itinerary Update Figure 1: Three-stage progressive architecture of TriFlow. “local food”), while executable itineraries must satisfy strict opera- tional, temporal, spatial, and budgetary constraints. Despite rapid progress under the emerging “LLM Agent + Tools” paradigm, exist- ing systems still struggle with real-world deployment. LLM-based agents that rely on tool generation often suffer from hallucinations, violate real-world constraints, and incur high token costs [5, 6]. These limitations hinder both feasibility and personalisation. Recent research has addressed these challenges through both benchmarking and methodological advances. New benchmarks such as TravelPlanner and TripTailor expand task scale, POI cover- age, and verifiable metrics, exposing a persistent gap between hard- constraint satisfaction and human-level experiential quality [5, 6]. Methodologically, neuro-symbolic and optimisation-coupled frame- works translate natural-language intents into computable struc- tures [3], while multi-agent and “generate–verify–retrieve” sys- tems enhance coordination and constraint checking through co- operative reasoning [2]. Personalisation-oriented studies explore lightweight user modelling and preference elicitation to improve alignment [1, 4]. However, these works often evolve independently, lacking an integrated design that jointly ensures orchestration ro- bustness and cost efficiency. To bridge these gaps, we propose TriFlow, a progressive multi-agent framework for intelligent trip planning that unifies structured reasoning and flexible natural lan- guage understanding. TriFlow is built upon three principles: (1) a staged retrieval–planning–governance pipeline that progressively maps user requests into denoised structured intermediates, na

This content is AI-processed based on open access ArXiv data.