PaperlayerView exact revision on arXiv

Source-linked AI summary

ChinaTravel: An Open-Ended Travel Planning Benchmark with Compositional Constraint Validation for Language Agents

Jie-Jing Shao, Bo-Wen Zhang, Xiao-Wen Yang, Baizhi Chen, Si-Yu Han, Jinghao Pang, Wen-Da Wei, Guohao Cai, Zhenhua Dong, Lan-Zhe Guo, Yu-Feng Li

arXiv:2412.13682v5cs.AIcs.CL

TL;DR

Travel planning combines practical value with complex constraints and often implicit intent. ChinaTravel benchmarks open-ended multi-POI planning with human queries and compositional constraint validation; neuro-symbolic agents achieve higher constraint satisfaction than purely neural methods, while open-world challenges remain.

  • Problem

    Travel intent can be implicit and contextually ambiguous, requiring agents to connect user expressions with verifiable symbolic meanings.

    Source: §3 · p. 5

  • Method

    ChinaTravel combines open-ended human travel queries with a DSL-based framework for evaluating multi-POI itineraries, feasibility, constraint satisfaction, and preference comparison.

    Source: §1 · p. 2

  • Results

    37.0% constraint-satisfaction success for neuro-symbolic agents versus 2.60% for purely neural methods; contextual grounding and unseen concept composition remain challenges.

    Source: §1 · p. 2

  • Takeaways & Limitations

    ChinaTravel provides a standardized evaluation framework and identifies contextual grounding and compositional concept generalization as directions for future work.

    Source: §5 · p. 10

    The initial release focuses on Chinese cities and native Chinese speakers, although its sandbox, DSL verification, and identified open-world challenge are language- and region-agnostic.

    Source: §7 · p. 11

Abstract

from arXiv · show

Travel planning stands out among real-world applications of \emph{Language Agents} because it couples significant practical demand with a rigorous constraint-satisfaction challenge. However, existing benchmarks primarily operate on a slot-filling paradigm, restricting agents to synthetic queries with pre-defined constraint menus, which fails to capture the open-ended nature of natural language interaction, where user requirements are compositional, diverse, and often implicitly expressed. To address this gap, we introduce \emph{ChinaTravel}, with four key contributions: 1) a practical sandbox aligned with the multi-day, multi-POI travel planning, 2) a compositionally generalizable domain-specific language (DSL) for scalable evaluation, covering feasibility, constraint satisfaction, and preference comparison 3) an open-ended dataset that integrates diverse travel requirements and implicit intent from 1154 human participants, and 4) fine-grained analysis reveal the potential of neuro-symbolic agents in travel planning, achieving a 37.0% constraint satisfaction rate on human queries, a 10 \times improvement over purely neural models, yet highlighting significant challenges in compositional generalization. Overall, ChinaTravel provides a foundation for advancing language agents through compositional constraint validation in complex, real-world planning scenarios. Project Page: https://www.lamda.nju.edu.cn/shaojj/ChinaTravel/index.html

1 INTRODUCTION

Travel planning requires agents to integrate information from tools and make interdependent decisions while meeting user constraints, but prior benchmarks largely use fixed-attribute slot filling rather than open-ended compositional requests. ChinaTravel introduces a multi-POI benchmark with a real-world sandbox, compositional DSL, human-authored queries, and evaluations of neuro-symbolic agents.

1 INTRODUCTION

  • Prior travel benchmarks rely on fixed-attribute slot filling and templated queries, limiting evaluation of compositional, open-ended human requirements.The paper also notes omissions of intracity scheduling and fixed rule lists that do not generalize to unseen requirements.Source: §1 · p. 2

ChinaTravel Sandbox

  • ChinaTravel provides a real-world sandbox and tools for planning multi-day, multi-POI itineraries with detailed travel information and spatiotemporal schedules.Agents are expected to produce feasible plans that satisfy logical constraints and preferences.Source: §1 · p. 2, Figure 1 · p. 2

Compositional Constraints Evaluation

  • Its domain-specific language composes travel-attribute concepts across spatial, temporal, cost, and type dimensions for scalable constraint validation and preference ranking.Evaluation covers feasibility, constraint satisfaction, and preference comparison.Source: §1 · p. 2

Open-Ended Travel Dataset

  • Human-authored queries include 154 validation examples and a 1,000-query test subset designed to assess generalization to unseen constraint compositions.The benchmark integrates diverse travel requirements and implicit intent from human participants.Source: §1 · p. 2

Empirical Analysis of Neuro-Symbolic Agents

  • 37.0% constraint satisfaction success for neuro-symbolic agents compares with 2.60% for purely neural methods on human queries.The authors identify open contextual grounding and unseen concept composition as key challenges.Source: §1 · p. 2

1 INTRODUCTION

  • ChinaTravel aims to bridge academic research and practical applications with a benchmark for rigorous constraint-satisfaction evaluation in travel planning.The benchmark focuses on open-ended multi-POI planning with authentic Chinese travel queries.Source: §1 · pp. 2–3

2 CHINATRAVEL BENCHMARK

ChinaTravel combines a travel-information sandbox with DSL-based evaluation of environmental feasibility, compositional logical constraints, and soft preferences. Its construction spans structured travel data, generated queries, automated checks, and human-authored requirements.

2 CHINATRAVEL BENCHMARK

  • ChinaTravel provides a sandbox for multi-day itineraries with multiple POIs, evaluated for environmental feasibility, constraint satisfaction, and preference comparison.The sandbox supplies travel information for 10 Chinese cities, including transport, attractions, restaurants, and hotels.Source: §2 · p. 3, §2.1 · p. 3

2.1 ENVIRONMENT INFORMATION

  • The sandbox checks 25 environmental constraints across six categories, including dietary, accommodation, transport, temporal, spatial, and attraction requirements.Examples include verifying that planned POIs exist in the designated city, transport is viable, and timing is accurate.Source: §2.1 · p. 3

2.2 LOGICAL CONSTRAINT

  • A Python-like DSL composes basic concepts into logical requirements and validates plans, avoiding per-requirement rule engineering.It supports constraints over travel events, including cost and arrival-time requirements.Source: §2.2 · pp. 3–4

2.3 PREFERENCE REQUIREMENT

  • Soft preferences are quantitative objectives, formalized in the DSL as minimization or maximization rather than binary constraint checks.Examples include maximizing attractions visited and minimizing transport time between POIs.Source: §2.3 · p. 4

2.4 BENCHMARK CONSTRUCTION

  • The benchmark combines manually built data and APIs, LLM-generated queries, DSL-based quality checks, and human-authored Human-154 and Human-1000 sets.Quality control retained 154 queries with novel constraints for validation; subsequent collection produced the Human-1000 test set.Source: §2.4 · p. 4

3 BENCHMARK CHARACTERISTICS

ChinaTravel raises the scale and openness of travel planning through dense contextual information and compositional, often implicit requirements. Its analyses show both broad constraint diversity and substantial difficulty with contextual grounding.

Context-Rich Long-Horizon Planning

  • Over 1,200 candidate POIs per query and 4.8K output tokens for five-day plans make ChinaTravel substantially larger than prior travel benchmarks.The benchmark generates 540M contextual tokens; even six-POI downsampling retains 40K tokens, and the authors state traditional single-pass text generation is inadequate.Source: §3 · p. 5, Figure 3 · p. 5

Diversity and Openness of Travel Requirements

  • ChinaTravel spans 6–12 constraints per query and 100 human-test constraint types, versus TravelPlanner’s 10 combinations and bias toward five or fewer constraints.Its constraint co-occurrence follows a long-tail pattern, unlike the relatively uniform frequencies reported for TravelPlanner.Source: §3 · p. 5

Contextual Grounding for Implicit Intent

  • 78.4% of Human1000 DSL statements contain semantic POIs requiring contextual grounding, compared with 5.4% in TravelPlanner.Examples include interpreting “local cuisine” by city and excluding spicy cuisines for children who cannot eat spicy food.Source: §3 · p. 5
  • On authentic semantic POIs, grounding accuracy drops from 94% to 76% for DeepSeek-V3 and from 79% to 53% for GPT-4o.Both models exceed 90% accuracy on TravelPlanner, whose semantic POIs follow simpler synthetic patterns.Source: §3 · p. 6

4 EMPIRICAL STUDY

The study compares training-free neural and neuro-symbolic planners on ChinaTravel, finding stronger constraint satisfaction for NeSy planning while exposing limitations in translation and preference trade-offs.

4 EMPIRICAL STUDY

  • The evaluation compares training-free pure-LLM and NeSy methods using delivery, environmental, logical, final, and conditional logical pass rates.C-LPR evaluates logical satisfaction conditional on environmental feasibility; the method set includes TTG, LLM-modulo, and NeSy Planning.Source: §4 · p. 6

4.2 MAIN RESULTS

  • $2.43 per query for GPT-4o accompanies persistently low FPR among pure LLMs, while DeepSeek reaches 6% EPR and 5% FPR only at the easy level.High delivery rates indicate structurally valid JSON, but near-zero EPR indicates difficulty acquiring and adhering to constraints.Source: §4.2 · p. 7, Table 2 · p. 7
  • 18% valid solutions on easy instances and only 23% of 2-day and 6% of 3-day cases solved within 15 minutes expose TTG's scaling limits.LLM-modulo rectification falls to at most one error per iteration after 3–5 rounds, while smaller models introduce more emergent errors.Source: §4.2 · p. 7, Figure 6 · p. 9
  • 52.6%, 37.0%, and 23.3% FPRs across three subsets show NeSy Planning with DeepSeek outperforming TTG and LLM-modulo, with gains persisting on human-query subsets.The comparison holds even without oracle translation.Source: Table 3 · p. 8, §4.2 · p. 8
  • Constraint translation remains a bottleneck: syntax repair can drop required clauses, and unseen DSL compositions achieve 12% structural alignment versus 93% for seen patterns.Unseen structures account for 84% of cases; the reported overall alignment is 9% when weighted by frequency.Source: §4.2 · pp. 8–9, Figure 7 · p. 9
  • The authors identify constraint grounding as the primary bottleneck and leave robust compositional generalization to future work.They position ChinaTravel as a benchmark for identifying the gap between current research and real-world scenarios, not as a complete solution.Source: §4.2 · p. 9

4.3 ABLATION STUDY WITH PREFERENCE

  • PEQ outperforms BQ in 5/6 single-preference scenarios, but underperforms on P2, minimizing restaurant transport time.The authors attribute the P2 result as likely arising from misinterpretation of complex spatiotemporal constraints.Source: Figure 8 · p. 10, §4.3 · p. 10
  • In multi-preference tests, PEQ's win rates vary from 0.58 on P3 & P4 to 0.10 on P1 & P4, exposing difficulty with rigid cost-sensitive trade-offs.The comparisons pair preferences across 15 test subsets and use pairwise Pareto win rates.Source: Figure 8 · p. 10, §4.3 · p. 10

5 CONCLUSION

ChinaTravel pairs open-ended Chinese travel queries with a DSL-based framework for compositional constraint validation. Its analysis indicates the potential of neuro-symbolic methods while identifying contextual grounding and compositional generalization as open-world challenges.

5 CONCLUSION

  • ChinaTravel provides an open-ended benchmark for multi-day, multi-POI travel planning focused on authentic Chinese needs.Its DSL-based framework evaluates feasibility, constraint satisfaction, and preference comparison; the empirical analysis highlights neuro-symbolic methods’ potential for constraint adherence.Source: §5 · p. 10

7 ETHICS STATEMENT

The ethics statement describes the benchmark’s potential benefits and deployment risks, and notes its initial focus on Chinese-language travel planning. The authors report following the ICLR Code of Ethics and proactively reviewing data collection, curation, evaluation, and release.

Ethics review

  • The authors state that they followed the ICLR Code of Ethics and proactively reviewed data collection, curation, evaluation, and release.Source: §7 · p. 11

Potential Positive impacts

  • The benchmark aims to encourage robust, constraint-aware travel assistants and support cross-disciplinary research, while its positive impacts remain potential rather than established outcomes.The statement identifies improved planning effectiveness and validation for real-world applications as potential benefits.Source: §7 · p. 11

Potential negative impacts

  • Downstream deployment may carry bias and misuse risks, so the authors recommend fairness audits, inclusive data practices, ethical safeguards, and plan verification before adoption.ChinaTravel is released for research purposes only; the statement also recommends warning users that generated plans are suggestions.Source: §7 · p. 11

Language and Regional Scope (Bias Considerations)

  • The initial benchmark centers on Chinese cities and native Chinese speakers; future iterations are intended to extend its language coverage.The authors cite practical constraints including API costs and compute and latency budgets, while stating that core components are language- and region-agnostic.Source: §7 · p. 11

8 REPRODUCIBILITY STATEMENT

The supplementary materials and main paper describe the resources and procedures needed to rerun the work. LLMs were used for data synthesis and sentence polishing, but not research ideation, experimental design, or core method development.

8 REPRODUCIBILITY STATEMENT

  • The anonymous codebase, dataset splits, README, and main-paper and appendix materials document dependencies, run commands, configurations, benchmark design, DSL, baseline, evaluation, and artifact licensing.The materials include the search sketch, pseudocode, prompts, protocol, and metrics.Source: §8 · p. 11

A THE USE OF LARGE LANGUAGE MODELS

  • LLMs assisted with sentence polishing during revision, and the authors reviewed and revised the edits and retain responsibility for the content.The authors state that LLMs did not contribute to research ideation, experimental design, or core method development.Source: A THE USE OF LARGE LANGUAGE MODELS · p. 16
  • LLMs generated the easy and medium synthetic datasets, with the full procedure and human-verification criteria documented in Appendix D.4.Source: A THE USE OF LARGE LANGUAGE MODELS · p. 16

B DISCUSSION OF LIMITATIONS

The discussion identifies limits in language and query coverage and in the methods evaluated. The accompanying sections describe prior work, the sandbox and DSL, and the construction of synthetic and preference data.

B DISCUSSION OF LIMITATIONS

  • The benchmark focuses on Chinese travel planning, and translated queries may not preserve the challenges of understanding Chinese requirements; its human-query count is also limited by collection costs.The authors note that the benchmark may not cover the full range of real-world requirements and call for continued refinement and expansion.Source: B DISCUSSION OF LIMITATIONS · pp. 16–17
  • The evaluation focuses on prompt-based methods and does not train tool-using agents with RL post-training, which the authors defer amid compute and trajectory-synthesis challenges.The stated trajectory-synthesis challenges include coverage and off-policy bias.Source: B DISCUSSION OF LIMITATIONS · p. 17

C DISCUSSION WITH RELATED WORK

  • Prior work spans slot- and tool-based travel benchmarks, single-day urban planning, and neuro-symbolic constraint solving; the cited TravelAgent study evaluates only 20 queries.The discussion also describes a cited SMT-based framework with a 97% pass rate on TravelPlanner.Source: C DISCUSSION WITH RELATED WORK · p. 17

D.1 SANDBOX INFORMATION; D.2 CITY-WISE DISTRIBUTION STATISTICS

  • The sandbox database was assembled from public information about attractions, accommodation, activities, and transportation, while human queries follow a long-tailed city distribution.POI collection was conducted per city to keep distribution scales comparable across urban datasets.Source: §D.1 · p. 18, §D.2 · p. 18

D.3 TUTORIAL OF DSL EXPRESSION

  • The DSL tutorial presents predefined concept functions for expressing logical constraints and preferences.The tutorial is described as a step-by-step guide to using ChinaTravel’s DSL.Source: §D.3 · p. 18

Logical Constraint

  • The Python-like DSL formalizes travel requirements as executable code, and neuro-symbolic agents use it to validate plans and rank candidate itineraries by preferences.The DSL includes concept functions, logical, arithmetic, and comparison operators, plus loops, conditionals, and set operations.Source: Logical Constraint · pp. 21–22

D.4 QUERY SYNTHESIS

  • Synthetic queries are generated from randomized travel-query skeletons and stratified into easy queries with one extra constraint and medium queries with three to five constraints.LLM paraphrasing varies the natural-language expression of requirements.Source: §D.4 · p. 22, §D.5 · p. 22

D.5 DIVERSITY OF SYNTHETIC DATA AND BIAS MITIGATION

  • The synthetic-data design diversifies constraint combinations across temporal, spatial, and cost dimensions, then uses human validation to filter unrealistic queries.Preference sub-datasets are built from six common preferences identified in user surveys, with corresponding DSL expressions provided.Source: §D.5 · p. 22, §D.6 · p. 27

D.7 BENCHMARK DIFFICULTY AND APPLICABILITY

ChinaTravel’s Human subset remains difficult for the baseline NeSy solution, while its DSL is designed to extend incrementally to new travel concepts and combine them with existing constraints.

D.7 BENCHMARK DIFFICULTY AND APPLICABILITY

  • 60.6% and 46.7% FPR on Easy and Medium subsets, respectively, provide initial testing and refinement points for developers.The Human subset is described as extremely difficult because it requires open-language reasoning and unseen concept composition.Source: §D.7 · p. 29

D.8 DSL EXTENSION

  • The DSL separates reusable compositional operators from pluggable domain predicates, so adding concepts requires exposing new attributes rather than overhauling the framework.The described extension adds an attribute to the sandbox and a helper function or predicate to the DSL library.Source: §D.8 · pp. 29, 31
  • New concepts can combine with temporal and structural constraints, such as choosing the most scenic attraction on day 1 or avoiding COVID-risk venues on specified days.ChinaTravel is designed to represent and automatically check these open-ended compositional constraints.Source: §D.8 · p. 31

E DISCUSSION WITH TRAVELPLANNER

ChinaTravel adds specific POIs, realistic time and transport requirements, and open-ended logical constraints beyond TravelPlanner’s fixed constraint types. Reported benchmark comparisons show current LLM and neuro-symbolic systems struggle on ChinaTravel, while the adapted NeSy planner uses constraint extraction and stepwise search.

E DISCUSSION WITH TRAVELPLANNER

  • ChinaTravel adds constraints such as return times and intercity transport costs to TravelPlanner’s predefined constraint types, creating open-language and unseen-composition challenges.Its constraint set also includes many more cuisine, attraction, hotel-feature, and transport types, plus specific POI names.Source: E DISCUSSION WITH TRAVELPLANNER · p. 31
  • ChinaTravel supports named POIs and more realistic transport and time management, while its medium queries combine 2–5 constraints and its harder examples add unseen logic and colloquial expressions.The paper presents these features as making the benchmark more comprehensive and realistic for evaluating travel-planning agents.Source: E DISCUSSION WITH TRAVELPLANNER · pp. 31–32
  • 0% GPT-4o, 1.29% TTG, and 2.59% LLM-Modulo success on ChinaTravel contrast with TTG’s 91.7% on TravelPlanner.Pure LLM performance also falls to near-zero on ChinaTravel, according to the comparison.Source: Table 14 · p. 32, E DISCUSSION WITH TRAVELPLANNER · p. 32

F NESY PLANNING

  • The adapted NeSy planner extracts logical constraints from natural language, then uses depth-first search to add activities one by one and validate the resulting plan.It uses rule-based activity selection and ranking; meals and attraction visits are set to 90 minutes, subject to closing times.Source: F NESY PLANNING · p. 32
  • The symbolic sketch has a five-minute per-query runtime limit, excluding LLM inference; if no validated plan is found, the system returns its best available intermediate plan.The returned plan may satisfy environmental constraints while achieving the highest number of validation-code successes, or may be partial if no environment-compliant plan is identified.Source: F NESY PLANNING · pp. 32–33

G EVALUATION METRIC IN COMPETITION

The paper defines metrics for delivery, environmental and logical validity, and final feasibility, then examines reasoning models and multi-preference planning. Preference-enhanced queries lead on most tested combinations, but conflicting preferences remain difficult to balance.

G EVALUATION METRIC IN COMPETITION

  • Delivery Rate concerns formatted plan delivery, Environmental Pass Rate measures use of sandbox information, Logical Pass Rate measures personalized requirements, and Final Pass Rate counts feasible plans meeting all constraints.For NeSy planning, delivery also reflects obtaining effective solutions within a limited search time.Source: G EVALUATION METRIC IN COMPETITION · pp. 33–34

H.3 MULTI-PREFERENCE COMPARISON

  • 10/15 preference combinations favor PEQ, with an aggregated ranking of 1.43 versus BQ’s 1.56.The comparisons pair six preferences in 15 subsets, each containing queries with two preference requirements.Source: Table 15 · p. 34, §H.3 · p. 35

H.1 RESULTS WITH LARGE REASONING MODEL

  • DeepSeek-R1 does not consistently outperform R1-Distill-Qwen-7B, while 7–8B LLMs already show some DSL-translating ability.The authors suggest over-thinking may weaken instruction-following in long contexts, and note that modest resources could support cost-effective post-training research.Source: §H.1 · p. 34, Table 16 · p. 35

H.2 RESULTS ON MEDIUM SET

  • The medium set covers queries with 3–5 logical requirements, a middle complexity tier between simple queries and highly complex open-ended scenarios.The paper says these medium-complexity results were omitted from the main text for organizational coherence.Source: Table 17 · p. 35, §H.2 · p. 35

H.3 MULTI-PREFERENCE COMPARISON

  • 9/15 combinations show PEQ improving one preference at the expense of another; in P0↑+P1↓, its aggregated ranking is 1.55 versus BQ’s 1.44.The authors report that conflicting objectives can negate the benefits of natural-language guidance.Source: §H.3 · p. 36

H.4 ANALYSIS OF PURE-LLM METHODS

The analyses expose persistent weaknesses in pure-LLM travel planning and examine which NeSy components contribute to performance. Results also show that English does not remove the challenge and that performance degrades as constraints combine more concepts.

H.4 ANALYSIS OF PURE-LLM METHODS

  • Pure-LLM methods have near-zero constraint-satisfaction success; Reflexion yielded 0% FPR and prohibitive token-heavy costs in preliminary evaluations.These costs made large-scale evaluation infeasible under the authors’ resource constraints.Source: §H.4 · p. 36
  • Reported pure-LLM failure modes include invalid API calls, repetitive planning loops, reasoning-action mismatches, and omissions of essential details such as return transport.The authors say these limitations underscore the inadequacy of pure-LLM approaches for long-horizon, constraint-rich travel planning.Source: §H.4 · p. 36

H.5 THE ABLATION STUDY OF THE PROPOSED NESY PLANNING

  • Oracle-translation comparisons identify NL2DSL translation as critical but a current bottleneck, while DSL-guided backtracking search refines plans more effectively than post-hoc constraint correction.The authors attribute the translation bottleneck to unseen concept composition.Source: §H.5 · p. 37
  • Random POI ranking lowers FPR on Easy-300 and Human-154, while random-ranking NeSy variants still outperform pure-LLM agents.The reported FPR pairs are 74.0 → 30.3 and 52.6 → 25.6 on Easy-300, and 45.4 → 38.3 and 37.0 → 31.8 on Human-154.Source: Table 18 · p. 37, §H.5 · p. 37

H.6 RESULTS ON ENGLISH SETTING

  • In the English setting, pure-LLM performance on long-horizon planning remains near 0%, and DSL translation remains an essential grounding bottleneck for NeSy methods.The authors conclude that the fundamental challenge is language-independent.Source: §H.6 · p. 37, Table 19 · p. 37

H.7 ANALYSIS OF THE CHALLENGE ON COMPOSITION COMPLEXITY

  • Agent performance degrades significantly as the number of composed DSL concepts increases from one to five, affecting syntax structure and semantic understanding.The analysis measures POI Reasoning and Syntax Generation matching rates as composition complexity increases.Source: §H.7 · p. 38, Table 20 · p. 38

I STATEMENTS ABOUT SCIENTIFIC ARTIFACTS

J.7 DSL ANNOTATION FOR HUMAN DATA

Human-data DSL annotations underwent multiple rounds of model-assisted and human verification, a process the authors describe as time-consuming and limiting dataset growth. The accompanying sections report collection timing and adapt TTG’s MILP formulation for ChinaTravel.

J.7 DSL ANNOTATION FOR HUMAN DATA

  • DSL annotations were generated initially with GPT-4o, revised by five annotators, then verified and corrected in two rounds by three benchmark developers.Queries that could not be verified within the sandbox were excluded.Source: §J.7 · p. 40
  • The authors say DSL expertise, sandbox debugging, and meticulous verification make human-data annotation time-consuming and constrain dataset growth.They characterize this rigorous process as supporting the quality and reliability of the human data.Source: §J.7 · p. 41

J.8 TEMPORAL COVERAGE OF HUMAN DATA

  • Human-154 was collected from August to October 2024, while Human-1000 collection ran from November 2024 to April 2025 to capture temporal diversity.The authors cite festivals, seasonal travel, and changing weather or peak periods as examples represented in the longer collection window.Source: §J.8 · p. 41

K.1 CONSTRAINTS FORMULATION

  • ChinaTravel adapts TTG’s MILP formulation for solver-based optimization, with its parameters, variables, and constraints specified in accompanying tables.Source: §K.1 · p. 41

K.2 EXPERIMENTAL SETTING

The experimental setting limits TTG to a reduced POI set because full-sandbox constraint growth is computationally prohibitive. Even under this reduction, TTG yields few valid solutions and solves only a small share of longer itineraries within the allotted time.

K.2 EXPERIMENTAL SETTING

  • The full sandbox would produce over 10B constraints, so TTG uses 22 POIs and one-hour time steps, leaving a scale of 320k × days constraints and 36k × days variables.LLMs select the POIs from the sandbox to preserve flexibility across queries.Source: §K.2 · p. 41, Table 25 · p. 53, Table 24 · p. 53
  • With a relaxed 15-minute search limit, TTG produced 18% valid solutions on easy instances but an 8% final pass rate; it found solutions for 23% of two-day and 6% of three-day itineraries.The authors attribute the lower final pass rate to the solver’s pruning heuristics.Source: §K.2 · p. 41
Loading 2412.13682v5…