PaperlayerView exact revision on arXiv

Source-linked AI summary

WorldTravel: A Realistic Multimodal Travel-Planning Benchmark with Tightly Coupled Constraints

Zexuan Wang, Chenghao Yang, Yingqi Que, Zhenzhu Yang, Huaqing Yuan, Yiwen Wang, Zhengxuan Jiang, Shengjie Fang, Zhenhe Wu, Zhaohui Wang, Zhixin Yao, Jiashuo Liu, Jincheng Ren, Yuzhen Li, Yang Yang, Jiaheng Liu, Jian Yang, Zaiyuan Wang, Ge Zhang, Zhoufutu Wen, Wenhao Huang

arXiv:2602.08367v1cs.CL

TL;DR

Existing travel-planning benchmarks often simplify interdependent constraints and provide structured data, offering only a partial view of planning from fragmented webpage information. WorldTravel pairs real-world travel scenarios with a multimodal web environment. In its evaluation, GPT-5.2 achieves 32.67% feasibility with text-only parameters and 19.33% when extracting parameters from webpages.

  • Problem

    Existing travel-planning benchmarks often use loosely coupled, greedy-solvable constraints and pre-normalized data, providing only a partial view of planning from fragmented webpage information.

    Source: §1 · p. 2

  • Method

    WorldTravel pairs 150 travel scenarios across five European cities with WorldTravel-Webscape, a controlled multimodal environment of over 2,000 rendered webpages for evaluating constraint extraction and itinerary planning.

    Source: §1 · p. 2

  • Results

    Global temporal reasoning remains difficult, and multimodal webpage extraction lowers feasibility: GPT-5.2 achieves 32.67% in text-only settings and 19.33% in multimodal settings.

    Source: §1 · p. 2, §6 · p. 10

  • Takeaways & Limitations

    The findings suggest perception and reasoning are independent bottlenecks, both essential for real-world planning tasks.

    Source: §6 · p. 11

    The evaluation abstracts away secondary real-world noise such as stochastic delays, while retaining temporal coupling and information fragmentation.

    Source: §3.1 · p. 4

Abstract

from arXiv · show

Real-world autonomous planning requires coordinating tightly coupled constraints where a single decision dictates the feasibility of all subsequent actions. However, existing benchmarks predominantly feature loosely coupled constraints solvable through local greedy decisions and rely on idealized data, failing to capture the complexity of extracting parameters from dynamic web environments. We introduce \textbf{WorldTravel}, a benchmark comprising 150 real-world travel scenarios across 5 cities that demand navigating an average of 15+ interdependent temporal and logical constraints. To evaluate agents in realistic deployments, we develop \textbf{WorldTravel-Webscape}, a multi-modal environment featuring over 2,000 rendered webpages where agents must perceive constraint parameters directly from visual layouts to inform their planning. Our evaluation of 10 frontier models reveals a significant performance collapse: even the state-of-the-art GPT-5.2 achieves only 32.67\% feasibility in text-only settings, which plummets to 19.33\% in multi-modal environments. We identify a critical Perception-Action Gap and a Planning Horizon threshold at approximately 10 constraints where model reasoning consistently fails, suggesting that perception and reasoning remain independent bottlenecks. These findings underscore the need for next-generation agents that unify high-fidelity visual perception with long-horizon reasoning to handle brittle real-world logistics.

1 Introduction

WorldTravel frames real-world itinerary planning as an end-to-end operations-research problem in which interdependent temporal and logical decisions can make a plan infeasible. It introduces a benchmark and web environment to test both planning and visual constraint acquisition.

Introduction

  • TravelPlanner [23] and ChinaTravel [17] often simplify tightly coupled challenges into greedily solvable subproblems and provide normalized constraints rather than information embedded in rendered webpages.Real interfaces can contain signals such as sold-out status and tiered pricing, while a single violation can invalidate an itinerary.Source: §1 · pp. 1–2
  • WorldTravel comprises 150 real-world tasks across five European cities and pairs expert-designed rubrics with WorldTravel-Webscape, a multimodal environment of more than 2,000 realistic booking webpages.Agents retrieve parameters from rendered webpages to satisfy hard constraints such as timed entries and travel buffers, alongside soft constraints such as cost and lodging preferences.Source: §1 · p. 2, Figure 1 · p. 2
  • GPT-5.2 achieves 32.67% feasibility with pre-extracted text-only parameters, falling to 19.33% in the multimodal setting, a 13.34 percentage point Perception-Action Gap.The multimodal setting requires models to extract constraint parameters from rendered webpages.Source: §1 · p. 2
  • Performance degrades as constraints increase, with an inflection around 10; even with structured pre-extracted parameters, the best model achieves only 36.67% feasibility.The findings identify visual constraint acquisition and logical reasoning as independent bottlenecks.Source: §1 · pp. 2–3

2 Related Work

Prior work advances LLM agents through decomposition, reasoning, memory, and tools, while travel-planning benchmarks and systems address different aspects of itinerary feasibility, robustness, and personalization. WorldTravel distinguishes strict temporal feasibility constraints from soft decision-correctness constraints.

LLM-Based Agents

  • LLM agents use task decomposition [18], stepwise reasoning [3] [22] [25] [26], long-term memory [11] [30], and tools [14] [16], but effectiveness under multiple constraints remains uncertain [23].Prior work applies LLM-based agents to multi-step tasks and domains including education, healthcare, and personalized dialogue [21] [29] [28] [11].Source: §2 · p. 3

Travel Planning

  • TravelPlanner [23], ChinaTravel [17], and Travel-Sim evaluate multi-constraint itineraries, authentic Chinese scenarios with compositional checks, and feasibility under step-by-step simulation, respectively.Together, these benchmarks target realistic, constraint-driven planning and long-horizon robustness.Source: §2 · p. 3
  • TravelAgent [5], ItiNera [19], and a neuro-symbolic approach [10] combine retrieval, recommendation, planning, and memory; spatial clustering; and SMT constraint solving, respectively.These systems target customized itineraries, spatial coherence, and provable correctness against supplied constraints.Source: §2 · pp. 3–4

Constraint taxonomy

  • Hard constraints govern temporal feasibility, and any violation makes an itinerary infeasible; soft constraints assess cost accounting and lodging selection only for feasible plans.Table 1 labels these categories CH and CS, respectively.Source: Table 1 · p. 4

3 The WorldTravel Benchmark

WorldTravel evaluates itinerary planning as a verifiable constrained-scheduling task, with agents extracting parameters from fragmented visual interfaces. Its Webscape environment combines tightly coupled travel constraints with screenshot-based interaction.

3.1 Task Formulation

  • The task abstracts itinerary planning as selecting activities, times, durations, and discrete options while extracting feasibility parameters from visual interfaces.Temporal decisions are tightly coupled: a fixed entry time can restrict downstream scheduling, sometimes leaving a unique valid itinerary; complexity is quantified by 0–5 temporal anchors. The abstraction removes secondary noise such as stochastic delays while retaining temporal coupling and information fragmentation.Source: §3.1 · pp. 4–5

3.2 Constraint Taxonomy

  • Hard constraints determine temporal feasibility, while soft constraints assess discrete decision correctness within feasible itineraries.Hard constraints include entry slots, minimum visit durations, and arrival buffers; soft constraints include exact cost calculations and selecting the correct hotel. This separation supports verifiable evaluation without conflating scheduling feasibility with choice accuracy.Source: §3.2 · p. 5

3.3 Environment Construction

  • The three-stage construction collects official and practical travel data, synthesizes diverse linked webpages and tasks, then verifies task solvability and quality.Quality control includes author review, independent cross-solving, and external validation, yielding 150 automatically verifiable tasks. Eight APIs return rendered screenshots rather than structured JSON, requiring agents to locate and cross-reference information across pages.Source: §3.3 · p. 5

3.4 Evaluation Protocol

  • Evaluation uses automated constraint verification to distinguish feasible schedules, partial hard-constraint satisfaction, and soft-constraint accuracy among feasible solutions.Feasibility measures whether all hard constraints are met; constraint violation diagnoses partial satisfaction, while optimality among feasible solutions measures correct costs and preference-matching choices.Source: §3.4 · p. 6

3.5 Benchmark Statistics

  • 150 tasks span 5 European cities, with 10–23 tightly coupled constraints per task and an average of 15+ across more than 2,000 webpages.The benchmark covers 36 attractions, 25 restaurants, and 26 hotels; tasks have 0–5 temporal anchors, most often 2 or 3. WorldTravel combines screenshot-based parameter extraction with dense constraint reasoning, unlike benchmarks that offer structured text or screenshots without dense reasoning.Source: §3.5 · pp. 6–7, Table 2 · p. 6

4 Experiments

Across text and screenshot-based evaluations, current models struggle to produce feasible itineraries, with failures tied to tightly coupled temporal constraints and a separate bottleneck in visual parameter extraction.

4.1 Experimental Setup

  • Removing soft constraints from the full 150-task benchmark yields modest Feasibility Rate improvements.This is reported as the Feasibility-Only ablation.Source: Table 4 · p. 8

4.2 Main Results

  • Substantial failures persist on 30 high-coupling tasks even when all hard constraint parameters are provided in structured form.This Gold-θ ablation tests performance with perfect parameter access.Source: Table 5 · p. 8

4.1 Experimental Setup

  • The evaluation tests 10 LLMs in a text setting and 7 VLMs in a multi-modal setting, reporting Feasibility Rate, Constraint Violation, and Optimality | Feasible.Text APIs return structured descriptions, screenshot-based APIs return rendered webpages, and five graduate-educated participants evaluated task difficulty.Source: §4.1 · p. 7, §4.2 · p. 8

4.2 Main Results

  • GPT-5.2 achieves only 32.67% Feasible Rate in the text setting, as tightly coupled scheduling decisions propagate constraints across later activities.The paper attributes this shortfall to temporal constraint satisfaction rather than simple information extraction or instruction-following failures.Source: §4.2 · p. 8
  • Proprietary models achieve Feasible Rates above 15%, while GLM-4.6 and Qwen-3-A235B remain below 10%, with Qwen-3-A235B at 2.67%.GPT-5.2 violates 21.92% of hard constraints on average, compared with 50.47% for Qwen-3-A235B.Source: §4.2 · p. 8
  • GPT-5.2 falls from 32.67% to 19.33% Feasible Rate across text and multi-modal settings, while Claude-Opus-4.5 falls from 21.33% to 14.00%.The multi-modal setting requires extracting parameters from screenshots, including distinguishing sold-out from available slots and combining information across pages.Source: §4.2 · p. 8

5 Discussion

The discussion tests whether parameter access, soft constraints, and constraint load or type explain planning failures. Its results show that planning remains difficult even with supplied parameters, while feasibility drops sharply near 10 constraints and timed-entry slots are especially challenging.

5.1 Planning Remains a Bottleneck Under Perfect Parameter Access

  • With Gold-θ parameters, GPT-5.2 feasibility remained 36.67% (+0 pp), while GPT-5.1 improved only from 13.33% to 16.67% (+3.34 pp).The ablation supplied hard-constraint parameters directly for 30 tasks selected for highest temporal coupling, showing that better parameter access alone did not resolve planning errors.Source: §5.1 · p. 9

5.2 Soft Constraints Have Limited Impact on Hard Constraint Satisfaction

  • GPT-5.2 feasibility rose from 32.67% to 33.33% (+0.66 pp) in text and from 19.33% to 21.33% (+2.0 pp) multimodally when soft constraints were removed.Feasibility nevertheless remained low at 33.33% in Text and 21.33% in Multi-modal, indicating that soft constraints have limited impact on hard-constraint satisfaction.Source: §5.2 · p. 9

5.3 Planning Capability Exhibits a Clear Capacity Limit

  • Hard-constraint satisfaction averaged 42%, 27%, and 40% in text versus 30% (−12 pp), 16% (−11 pp), and 35% (−5 pp) multimodally for dwell time, timed-entry/operational windows, and inter-activity feasibility.Timed-entry slots had the lowest satisfaction in both settings: 27% in text and 16% multimodally, likely because selecting a discrete slot requires precise alignment with downstream activities.Source: §5.3 · p. 10
  • Average feasibility fell from 36.3% at 6–7 constraints to 18.3% at 8–9, then to 3.4% at 10–11 and 3.1% beyond 12.The authors identify a capacity threshold around 10 constraints, beyond which feasibility collapses across model families; multimodal parameter extraction compounds the difficulty.Source: §5.3 · p. 10
  • GPT-5.2 retained over 40% feasibility through 9 constraints but achieved 13.5% at 10–11; GLM-4.6 fell from 28.1% to 0%, and Qwen-3-A235B from 9.4% to 0%.The models' declines illustrate differing resilience around the approximately 10-constraint threshold.Source: §5.3 · p. 10

6 Conclusion · Appendix

WorldTravel evaluates constraint-intensive travel planning across 150 tasks with 15+ tightly coupled temporal constraints, requiring agents to extract webpage parameters and coordinate decisions across an itinerary. Results show low feasibility even for the best model, with planning and perception identified as independent bottlenecks.

6 Conclusion

  • WorldTravel comprises 150 travel-itinerary planning tasks with an average of 15+ tightly coupled temporal constraints.Source: §6 · p. 10
  • Agents must extract parameters from rendered webpage screenshots and coordinate scheduling decisions that propagate across the entire itinerary.Source: §6 · p. 10
  • 32.67% Feasibility Rate for the best model in the text setting drops to 19.33% in the multi-modal setting.Source: §6 · p. 10
  • Even with perfect parameter access, feasibility reaches only 36.67% on high-coupling tasks, identifying planning reasoning as the primary bottleneck.Source: §6 · p. 11
  • Feasibility collapses beyond a planning capacity limit of approximately 10 constraints.Source: §6 · p. 11
  • The findings identify perception and reasoning as two independent bottlenecks, both essential for real-world planning tasks.Source: §6 · p. 11

A Data Collection Specifications … A.3 Data Grounding and Validation

WorldTravel grounds its travel-planning constraints in official factual sources and practical traveler-derived data, then validates both to reflect real-world operating conditions.

A Data Collection Specifications

  • The appendix describes collecting factual and practical data from different sources to construct constraint parameters and task content.The two data categories are presented as inputs to the benchmark’s scenarios.Source: A Data Collection Specifications · p. 14

A.1 Factual Data from Official Sources

  • Official sources provide venue operating hours, tiered attraction fares, and restaurant menu prices.The collected hours include mid-day closure gaps and date-specific closures.Source: §A.1 · p. 14
  • Official sources also supply timed-entry availability, ticket variants and add-on prices, event schedules, and hotel rates, room types, locations, and amenities.Timed-entry data includes both discrete booking slots and their availability status.Source: §A.1 · p. 14

A.2 Practical Data from Travel Platforms

  • Travel platforms and user-generated sources provide recommended visit and dining durations, arrival buffers, optimal visiting periods, and venue logistics.Examples include a 15-minute early arrival recommendation and tour boarding or meeting locations.Source: §A.2 · p. 14
  • Practical data also includes verified-traveler feedback on service, food, ambiance, and useful tips.These qualitative reviews complement operational recommendations and logistical details.Source: §A.2 · p. 14

A.3 Data Grounding and Validation

  • Annotators validate consistency across source types while recording official operating windows separately from practical arrival buffers.For example, a 09:00–18:00 official museum window can be paired with a 30-minute peak-season security buffer reported in reviews.Source: §A.3 · p. 14

B City-Specific Task Design · C Quality control

WorldTravel covers five European cities with attraction-centered, temporally constrained tasks, and uses staged construction, manual annotation, and repeated validation to curate its benchmark. Only tasks that pass three verification rounds are included in the final set of 150.

B City-Specific Task Design

  • Each city’s attraction-centered planning scenarios emphasize temporal feasibility under diverse realistic constraints.Table 6 summarizes representative task focuses for each city; the supplied passage does not specify the individual focuses.Source: B City-Specific Task Design · p. 15, Table 6 · p. 15

C Quality control

  • Quality control spans task construction, solution annotation, and multiple validation rounds.Source: C Quality control · p. 15
  • Each high-coupling core scenario is built as a feasible planning problem, then expanded by varying user profiles and task settings.Variations include language preference, group size, age, budget, ticket types, menu options, and hotel preferences.Source: C Quality control · pp. 15–16
  • Task authors write reference solutions manually, and repeated reviews remove ambiguity and ensure constraints are automatically checkable.Source: C Quality control · p. 16
  • Real-world factors are translated into checkable time and cost constraints, such as treating places with many stairs and no elevators as unsuitable for seniors or children.Source: C Quality control · p. 16
  • 150 tasks across five European cities remain after each task passes three rounds of verification.The rounds involve author review, independent solving by a second annotator, and external review for clarity and answer correctness.Source: C Quality control · p. 16
  • Different task batches receive different date ranges to prevent interference, while webpage edits enforce a unique feasible solution.Edits can adjust remaining ticket numbers, block time slots, or change visit durations.Source: C Quality control · p. 16

D API Specifications … D.5 Design Rationale

WorldTravel exposes travel-planning information as rendered webpage screenshots rather than structured JSON, requiring agents to extract and combine constraints across visual interfaces. Its APIs cover closed city-specific sets of attractions, restaurants, and hotels, along with booking, guide, and transportation information.

D.1 Attraction-Related APIs

  • Attraction APIs provide a closed city-specific attraction list, date-specific ticket pages with availability, schedules, capacity, prices, and add-ons, plus guides with visit-duration and practical advice.Ticket pages distinguish selectable from grayed-out slots; guides include visitor-profile suggestions and crowd-avoidance tips.Source: §D.1 · pp. 16–17

D.2 Restaurant-Related APIs

  • Restaurant APIs provide a closed city-specific list, date-specific reservation pages with opening hours and availability, and dining guides with duration, dish suggestions, menus, and prices.Guide and menu content supports meal planning and alignment with dining preferences.Source: §D.2 · pp. 17–18

D.3 Hotel-Related APIs

  • Hotel APIs return a screenshot listing the city’s hotels, from which agents must select when constructing itineraries.The list enumerates the hotels used in tasks and evaluation and presents image cards with names and brief descriptions.Source: §D.3 · p. 18

D.4 Transportation-Related APIs

  • Transportation APIs return a city-level matrix of supported locations and pairwise travel costs for walking, driving, public transit, and taxi.The matrix supplies inter-location costs across multiple transportation modes.Source: §D.4 · p. 18

D.5 Design Rationale

  • Screenshot outputs aim to mirror human interaction with booking platforms so benchmark performance reflects deployment capability and resists fixed-schema extraction shortcuts.Agents must infer relevant fields from visual layouts and natural-language descriptions.Source: §D.5 · p. 19
  • Agents must locate information in complex layouts, read visual availability cues, and cross-reference tool calls to recover the complete constraint space.The APIs distribute relevant information across booking portals, guide pages, menus, and transportation matrices.Source: §D.5 · p. 19
  • Screenshot-based APIs require visual extraction, cross-page aggregation, and reasoning over available versus unavailable options, avoiding structured-schema shortcuts.The design aims to mirror real booking platforms and support transferable planning behavior rather than benchmark-specific format overfitting.Source: D API Specifications · p. 16, §D.5 · p. 19

E Verification Function Specifications … F.1 Environment Scale

WorldTravel makes itinerary evaluation automated by mapping constraints to verifiable checks, separating temporal feasibility from decision correctness, and documenting the benchmark’s environment scale across five European cities.

E Verification Function Specifications

  • WorldTravel uses a “hard to solve, easy to verify” paradigm, mapping each constraint to an automated boolean check that scores an itinerary 1 if satisfied and 0 otherwise.Constraints are partitioned into hard constraints that determine feasibility and soft constraints that assess decisions within feasible plans.Source: E Verification Function Specifications · p. 19

E.1 Hard Constraint Verification

  • Hard constraints enforce temporal feasibility, so any violation makes an itinerary infeasible; verification requires joint reasoning about entry slots, minimum dwell times, and arrival buffers.Table 7 summarizes the hard-constraint verification functions.Source: §E.1 · p. 19, Table 7 · p. 19

E.2 Soft Constraint Verification

  • Soft constraints assess grounded decisions after feasibility is satisfied, measuring cost-reasoning accuracy and hotel selection.Table 8 summarizes the soft-constraint verification functions.Source: §E.2 · p. 19, Table 8 · p. 20

F Benchmark Statistics

  • Appendix F provides detailed WorldTravel benchmark statistics, including environment scale and constraint distribution.Source: F Benchmark Statistics · p. 20

F.1 Environment Scale

  • F.1 describes a diverse set of real-world attractions, restaurants, and hotels spanning five European cities.Examples range from world-renowned museums and Michelin-starred dining to local markets and boutique accommodations.Source: §F.1 · p. 20
  • The environment-scale table covers daily booking interfaces for 31 days per entity, travel guides, menus, hotel listings, and transportation matrices.Source: Table 9 · p. 20

F.2 Constraint Distribution … 2. Dining Planning Requirements

The appendix reports timed-entry constraint prevalence, describes model-level hard-constraint satisfaction reporting, and specifies the travel-planning rules and examples used in WorldTravel. Its examples show tasks with multiple fixed time anchors, while the prompt requires schedules to respect reservation times, durations, and user constraints.

F.2 Constraint Distribution

  • Over 80% of tasks contain 2 or more timed-entry slots, requiring agents to coordinate multiple time anchors.The distribution is analyzed by total constraint count, hard constraint count, and timed-entry slot count; fixed reservation times reduce scheduling freedom by constraining later activities.Source: §F.2 · pp. 20–21, Table 10 · p. 20

G Constraint Type Satisfaction by Model

H.1 Original System Prompt for Travel Planning Agent

  • The original system prompt defines a professional travel-planner role, tool-use restrictions, constraint hierarchies, and planning requirements for attractions, restaurants, hotels, and transportation.Source: §H.1 · p. 21

Travel Agent Role Setting Tool Usage Restrictions

  • The agent is limited to 30 purposeful tool calls, must not fabricate information or make subjective recommendations, and must prioritize explicit user time requirements over default schedule hours.It must provide a detailed daily itinerary; optional booking-page times and guide-recommended durations take priority over business hours unless the question specifies otherwise.Source: Travel Agent Role Setting Tool Usage… · pp. 21–22

1. Attraction Planning Requirements

  • Attraction planning requires checking city listings, guide durations, and ticket availability, then using an available booking time exactly and following the recommended visit duration.Specified attractions must not be replaced, gray slots are unavailable, and ticket prices are calculated for the travelers and applicable discounts.Source: §1 · pp. 22–23

2. Dining Planning Requirements

  • Dining plans must use restaurant listings, guide notes, and reservation availability, with unavailable gray slots excluded and the chosen start time matching the reservation page.Restaurant stays must follow the guide duration and fit the itinerary; associated rules require key routes and hotel information to be queried through the specified tools.Source: §2 · p. 23
  • The output must be JSON with dated daily schedules, hotel bookends, and item-level timing, route, cost, and transportation details.Ticket costs scale with travelers, hotel costs with rooms capped at two people each, taxi costs use the displayed map price, and bus fares are 3 EUR per person.Source: §2 · p. 23

H Task Examples

  • One task has 4 timed-entry slots across 3 attractions and 1 restaurant and 18 constraints (12 hard, 6 soft); another has 3 slots, including 2 same-day performances at one venue, and 15 constraints (9 hard, 6 soft).Source: §2 · pp. 23–24

H.2 Sample Agent Output … J.1 Experiment 1: Gold-θ Ablation (Berlin City-Specific Constraints)

The appendix illustrates a feasible sample itinerary and webpage-generation prompt, then specifies a Gold-θ ablation that isolates hard-constraint feasibility planning from webpage information extraction. The ablation supplies ground-truth planning parameters and requires an executable itinerary satisfying hard constraints.

H.2 Sample Agent Output

  • Doubao-1.8-Pro’s sample Berlin response is described as temporally feasible and follows the requested JSON schema, including time slots, costs, transportation, and reference images.Source: §H.2 · p. 25

Date: August 20

  • The August 20 sample schedules a hotel stay, attraction visits, restaurant meals, and taxi transfers across the day.Source: Date: August 20 · pp. 25–26

I Example HTML Generation Prompts

I. Basic Information & Header Section

  • The page prompt specifies an attraction hero and experience details, operating hours, ticket prices and quantities, dynamic totals, selectable dates and time slots, unavailable-slot indicators, and a booking button.Source: §I · p. 26

J System Prompts for Ablation Studies

J.1 Experiment 1: Gold-θ Ablation (Berlin City-Specific Constraints)

  • The Gold-θ ablation gives agents ground-truth structured parameters for Berlin attractions, restaurants, and hotels, removing webpage extraction to test pure feasibility planning.Source: §J.1 · p. 27
  • Agents must produce an executable itinerary that uses available reservation slots, respects venue hours, minimum dwell times, and arrival buffers, and includes only query-required activities and mandatory meals.Source: §J.1 · p. 27

J.2 Experiment 2: Temporal Feasibility Only (Hard Constraints Focus) … K.10 Restaurant Travel Notes

This section describes a multimodal experiment focused on temporal feasibility, using WorldTravel-Webscape, a realistic environment of varied rendered travel pages where agents extract dispersed constraints from linked sources and formats to build globally feasible itineraries.

J.2 Experiment 2: Temporal Feasibility Only (Hard Constraints Focus)

  • Experiment 2 tests whether agents can produce globally time-consistent itineraries from webpage screenshots when only temporal hard constraints are evaluated.The evaluation removes cost optimization, ticket preferences, and hotel preferences; constraints include availability windows, dwell times, arrival buffers, reservation slots, and inter-location travel times.Source: §J.2 · p. 28

K The WorldTravel-Webscape

  • WorldTravel-Webscape contains over 2,000 handcrafted webpages across travel domains, requiring agents to navigate multiple page types and integrate extracted details into feasible itineraries.The pages cover attractions, restaurants, hotels, transportation, booking systems, menus, and travel guides.Source: K The WorldTravel-Webscape · p. 34
  • Unlike benchmarks that expose structured attributes, WorldTravel embeds factual constraints, prices, availability, and experiential cues in rendered webpages for agents to perceive and integrate.Agents must extract decision-critical details and use them to form globally consistent itineraries.Source: K The WorldTravel-Webscape · p. 34

K.1 Design Principles of the WorldTravel-Webscape

  • The WorldTravel-Webscape is constructed around three core design principles.Source: §K.1 · p. 34
  • The Webscape embeds task information in varied rendered pages, requiring visual UI-state parsing and cross-page integration rather than reliance on structured fields or a single source.Relevant information is distributed across cards, tables, panels, booking systems, transportation matrices, and travel guides.Source: §K.1 · pp. 34–35

K.2 Attraction Overview Pages

  • Attraction overview cards help agents identify promising landmarks, but hours, ticket types, and reservation availability appear only on linked pages.The cards combine imagery, category tags, and short descriptions.Source: §K.2 · p. 35

K.3 Restaurant Overview Pages

  • Restaurant overview cards show cuisine, location context, and style, while agents must follow links to menus or reservation pages to check fit, times, availability, and prices.Source: §K.3 · p. 35

K.4 Hotel Listing Pages

  • Hotel listings use editorial descriptions of style, ambience, background, amenities, and location, so agents must match accommodations to preferences without preset categories.Source: §K.4 · p. 35

K.5 Transportation Matrix Pages

  • Transportation matrices provide travel times and distances between attractions, restaurants, and hotels across multiple modes for estimating transfers.Source: §K.5 · p. 35

K.6 Attraction Booking Pages; K.7 Restaurant Reservation Pages

  • Attraction booking interfaces encode fixed entry times, tiered ticket options, pricing rules, visitor categories, tours, arrival requirements, and sold-out status; restaurant calendars show meal-session slot availability.Source: §K.6 · p. 35, §K.7 · p. 35

K.8 Attraction Travel Guides

  • Attraction travel guides use blog-style recommendations, and some attractions provide floor plans showing internal spatial layouts.Source: §K.8 · pp. 35–36

K.9 Restaurant Menus; K.10 Restaurant Travel Notes

  • Restaurant menus visually list dishes, beverages, set meals, and time-dependent offerings, while travel notes mix subjective commentary with practical details such as peak hours and recommended visit length.Notes also include best visiting times.Source: §K.9 · p. 36, §K.10 · p. 36

Related papers

Loading 2602.08367v1…