TRIPPULSE: Multi-Agent Travel Planning with Review-Grounded Reasoning

Priyanshu Karmakar; Borru Vijay Sai; Shubhojit Mallick; Abhik Jana; Shreya Ghosh; Manish Gupta

Source-linked AI summary

TRIPPULSE: Multi-Agent Travel Planning with Review-Grounded Reasoning

Priyanshu Karmakar, Borru Vijay Sai, Shubhojit Mallick, Abhik Jana, Shreya Ghosh, Manish Gupta

arXiv:2608.30924v1Aug 31, 2026cs.CL

TL;DR

Travel itinerary generation must satisfy strict spatio-temporal constraints while capturing experiential preferences that structured attributes miss. TRIPPULSE addresses this gap with localized domain agents, review-derived signals, and hybrid scheduling, and experiments report improved constraint satisfaction, reliability, personalization, and experiential quality across proprietary and open-source models.

Problem
Travel planning must jointly satisfy strict spatio-temporal constraints and nuanced preferences, while structured attributes omit experiential signals available in reviews.
Method
TRIPPULSE decomposes planning across localized agents, integrates review-derived pros and cons, and uses a deterministic scheduling backend alongside LLM-based reasoning.
Results
Experiments on augmented TripCraft demonstrate improved constraint satisfaction and planning reliability while enabling effective use of proprietary and open-source language models.
Takeaways & Limitations
Review-grounded agentic planning supports more personalized, experientially grounded itineraries while preserving structural feasibility.
Takeaways & Limitations
The framework adds inference latency and orchestration overhead, relies on potentially noisy or biased review processing, and is evaluated mainly on U.S. travel scenarios.

Abstract

from arXiv · show

Travel itinerary generation requires balancing strict spatio-temporal constraints with human preferences. Existing LLM-based planners mainly rely on structured attributes and pre- defined traveler personas, but real travel deci- sions are often shaped by reviews that reveal experiential factors such as comfort, safety, ser- vice quality, ambiance, crowding, and hidden risks absent from structured databases. Incor- porating such review information is therefore critical to realistic, user-centric itinerary gen- eration. We propose TRIPPULSE1, a multi- agent framework for review-grounded travel planning. Instead of relying on a monolithic planner (and face context and reasoning bot- tlenecks), TRIPPULSE2 decomposes itinerary generation into specialized agents (each op- erating over localized contexts) for accom- modations, transportation, meals, attractions, and events, coordinated through a global or- chestrator with scheduling mechanisms that enforce temporal and budget feasibility. We augment TRIPCRAFT with 100K+ real-world reviews and introduce Review-Grounded Per- sona Alignment (RGPA), an LLM-as-a-Judge metric for evaluating alignment with human- centric travel experiences. Experiments across multiple trip durations and diverse proprietary and open-source models show that TRIPPULSE maintains strong constraint satisfaction while generating more personalized and experien- tially grounded itineraries.

1 Introduction

TRIPPULSE addresses travel planning as a joint problem of strict spatio-temporal feasibility and nuanced experiential preferences. It uses review-grounded multi-agent planning to coordinate specialized reasoning while preserving constraint satisfaction.

Travel itineraries must coordinate transportation, hotels, and activities under budget and scheduling constraints while optimizing safety, comfort, atmosphere, and suitability.
Existing planners commonly use monolithic prompting for entity selection, preference reasoning, scheduling, and constraint satisfaction, which can produce hallucinations and temporal or spatial violations as complexity grows.
Structured attributes omit experiential signals from reviews, while directly adding large, noisy reviews to monolithic models increases context load and reduces planning reliability.
TRIPPULSE assigns accommodations, transportation, meals, attractions, and events to domain-specific agents operating on localized contexts coordinated by a global orchestrator.
The framework extends TripCraft with 100K+ Airbnb and TripAdvisor reviews distilled into concise pros and cons, and introduces RGPA for experiential alignment evaluation.

2 Related Work

Prior travel-planning systems use LLM reasoning, task decomposition, retrieval, or formal solvers to address multiconstraint itinerary generation. TRIPPULSE targets their limited handling of review-grounded experiential preferences by combining agentic reasoning with deterministic scheduling.

LLM planning methods use chain-of-thought, Tree-of-Thought, and self-reflection, while multi-agent systems decompose tasks for efficiency.
LLMs still struggle with long-horizon constraint satisfaction, and PDDL grounding remains challenging for preference-driven travel planning.
Recent agentic travel systems address reasoning overload but remain limited to structured attributes without review-grounded entity selection mechanisms.
SMT-based hybrid systems improve feasibility but miss semantic nuances such as comfort and safety without extensive manual ontologies.
TRIPPULSE combines agentic reasoning with a deterministic backend to address both constraint feasibility and review-derived experiential attributes.

3 TRIPPULSE

TRIPPULSE decomposes itinerary generation into localized domain agents coordinated by a global orchestrator, then schedules selected entities under temporal and budget constraints. Its review-processing and hybrid scheduling components separate semantic selection from rigid feasibility execution.

3.1 Problem Formulation: The system defines an itinerary as a time-ordered sequence of selected travel entities with scheduled start and end times.
3.1 Problem Formulation: TRIPPULSE aims to maximize review-grounded quality while satisfying structural, temporal, budget, and local constraints.
3.2 Architecture: Five domain-specific agents, a Global Orchestrator, and a scheduling backend form the system’s multi-stage itinerary-generation pipeline.
3.2 Orchestration: The orchestrator coordinates agents, allocates budget, enforces duration and city-transition constraints, and passes selected entities to either an LLM-based or deterministic scheduler.
3.3 Dataset Construction: Review data from 140 U.S. cities is mapped one-to-one to TripCraft entities and transformed into structured pros and cons from up to five reviews per entity.
3.6 Scheduling: The deterministic scheduler replaces generative scheduling to guarantee full constraint satisfaction while reducing scheduling latency and computational overhead.

4 Experimental Setup

The evaluation tests TRIPPULSE across augmented TripCraft itineraries of multiple durations and model types using constraint, temporal, and review-grounded experience metrics. RGPA compares review-grounded and non-review itineraries across human-centric dimensions.

4.1 Dataset and Models: The augmented TripCraft benchmark supplies structured travel queries, deterministic transportation schedules, and hard constraint rules.
4.1 Dataset and Models: Experiments generate 3-day, 5-day, and 7-day itineraries to test scalability under increasing scheduling complexity.
4.2 Metrics: TRIPPULSE is evaluated with constraint satisfaction, temporal consistency, and review-grounded experience metrics.
4.2 Metrics: RGPA uses an LLM-as-a-Judge setup to assess whether review integration improves preference matching, safety, and overall travel quality.
4.2 Metrics: The judge evaluates persona alignment, experiential quality, and overall travel satisfaction, then aggregates pairwise preferences into a review-grounded win rate.
4.2 Metrics: RGPA produces dimension scores from 1 to 10 averaged across itinerary pairs, providing a broader evaluation of personalized travel quality than constraint metrics alone.

5 Experimental Results and Analysis

TRIPPULSE is evaluated against monolithic and agentic baselines on constraint satisfaction, review-grounded itinerary quality, evaluator robustness, and extraction quality. Results indicate that decomposition, deterministic scheduling, and review-derived attributes improve feasibility and persona alignment.

Benchmark Performance: TRIPPULSE compares LLM-based and deterministic scheduling backends with monolithic and agentic baselines on temporal, spatial, persona, and ordering metrics.Table 2 defines Tm, Ta, normalized Ta, Ss, Sp, and So for the comparison.
Benchmark Performance: Agentic decomposition enables smaller open-source models, including Qwen 2.5 and Phi-4, to maintain high feasibility across multiple trip durations.The results are contrasted with monolithic approaches that experience reasoning overload and with the agentic baseline.
Review Integration: Review integration significantly improves persona alignment by incorporating qualitative attributes into itinerary evaluation across models and trip durations.The comparison uses RGPA metrics for itineraries with and without review-derived Pros and Cons.
Robustness and Quality Validation: 81.00%, 65.49%, and 66.60% are the Qwen3-32B Win Rates for review-grounded 3-, 5-, and 7-day itineraries, respectively.These relative preferences remain consistent with GPT-5 evaluation despite differences in absolute score calibration.
Robustness and Quality Validation: 77.5% of manually evaluated 3-day itinerary pairs were judged equal to or better than the baseline, closely matching the 80.52% GPT-5 Win Rate.The reported gains primarily involved persona-aware selection of culturally relevant attractions, highly rated restaurants, and premium venues.
Robustness and Quality Validation: Pros and Cons extraction covered 2,369 accommodations, 3,791 restaurants, and 4,897 attractions, with empty-Pro rates below 1% but accommodation empty-Cons rates of 53.14%.The passage also reports high extraction precision and sentiment accuracy, particularly for positive signals.
Scheduling Analysis: The deterministic scheduler mitigates duration hallucinations and chronological inversions by enforcing meal gaps, transit buffers, and logical ordering.Generative scheduling produced zero-duration stays and a 3:30 AM lunch reservation, while the deterministic backend mapped selected Pros and Cons onto a mathematically verified skeleton.

6 Discussion

The discussion attributes TRIPPULSE’s constraint performance to localized multi-agent reasoning and its experiential alignment to review-derived Pros and Cons. It presents deterministic scheduling as a complementary mechanism for strict feasibility and computational efficiency.

Impact of Agentic Decomposition: Localized specialist agents reduce reasoning overload and entity hallucinations, improving constraint satisfaction over monolithic prompting.The design evaluates entities within bounded contexts, such as restaurants in one destination city.
Impact of Agentic Decomposition: Bounded contexts enable Qwen 2.5 and Phi-4 to achieve planning stability and constraint adherence competitive with massive proprietary models.The discussion identifies bounded-context design as the primary driver of this outcome.
Effect of Review Integration: Review-derived Pros and Cons improve experiential metrics and connect nuanced persona requests, such as a cozy atmosphere, to suitable entities.The same attributes can penalize locations flagged for hygiene or safety risks.
LLM vs. Algorithmic Scheduling: The deterministic scheduler guarantees strict temporal and budget satisfaction while reducing computational overhead through fewer LLM calls.The comparison frames this against the LLM scheduler’s higher semantic flexibility.

7 Conclusion

TRIPPULSE combines specialized travel-planning agents, review-derived experience signals, and centralized coordination to support feasible, personalized itineraries. Experiments on augmented TripCraft show improved constraint satisfaction and planning reliability across proprietary and open-source language models.

TRIPPULSE decomposes itinerary generation across accommodations, transportation, dining, attractions, and events, coordinated by a central Global Orchestrator.Review-derived Pros and Cons are integrated into planning, alongside the RGPA metric for experience quality, personalization, and risk avoidance.
The framework integrates review-derived Pros and Cons to evaluate experience quality, personalization, and risk avoidance through RGPA.
Experimental results on the augmented TripCraft benchmark show improved constraint satisfaction and planning reliability.
TRIPPULSE supports both proprietary and open-source language models in complex travel planning.

8 Limitations

TRIPPULSE improves constraint satisfaction and experiential quality, but its multi-agent design, review-processing pipeline, and benchmark scope impose important boundaries. The main trade-offs concern latency, review-data reliability, and generalization beyond U.S. travel settings.

The distributed, sequential architecture adds inference latency and orchestration overhead compared with single-pass monolithic generation.This creates a trade-off between semantic reasoning quality and real-time execution speed.
Automated Pros-and-Cons extraction is sensitive to review noise, manipulation such as review bombing, and domain biases in scraped datasets.
Evaluation on TripCraft primarily covers U.S. travel scenarios and structured transit networks.The benchmark may not fully capture global travel regions with less formalized tourism infrastructure.

9 Ethical Considerations

The benchmark augmentation uses publicly available user reviews without direct author recruitment or consent. Privacy protections remove personally identifiable information and usernames, while qualitative attributes are aggregated at the entity level.

The review data came from publicly available platforms, and the authors did not directly recruit or interact with its authors.
Individual consent was not obtained from the review authors.
Personally identifiable information and reviewer usernames were stripped during preprocessing.
The NLP pipeline aggregates qualitative attributes at the entity level rather than profiling individual reviewers.

A.2 Travel Planning Benchmarks and Systems

Prior travel-planning systems emphasize structured constraints, while review-aware methods add experiential information but face context and formalization challenges. TRIPPULSE addresses this gap through distilled review signals, specialized agents, and deterministic constraint enforcement.

Travel Planning Benchmarks and Systems: TravelPlanner and TripCraft formulate itinerary generation as multiconstraint reasoning over fine-grained spatio-temporal constraints.TP-RAG uses retrieval-augmented frameworks to guide point-of-interest sequences, while these systems primarily rely on structural databases.
LLM-as-a-Judge Evaluation: Deterministic constraint checks and spatial-temporal metrics measure structural feasibility and routing efficiency but miss subjective qualities such as ambiance, service quality, and safety.
LLM-as-a-Judge Evaluation: RGPA equips an LLM judge with distilled user reviews to connect quantitative TripCraft scores with human-centric experiential quality.
Review-Aware Recommendation Systems: Review-aware recommendation methods extract latent features, aspects, sentiments, or salient attributes from review text to represent preferences and item properties.
Review-Aware Recommendation Systems: TRIPPULSE uses distilled Pros and Cons as reasoning primitives inside constraint-aware entity-selection loops, rather than merely ranking entities.
Formal Planning and PDDL Grounding: Specialized domain modules create localized context windows, while deterministic controllers enforce budget, time, and distance constraints without discarding review nuance.

B Implementation Details and Hyperparameters

TRIPPULSE uses reproducible model configurations, formal query extraction, and deterministic constraint verification to support itinerary generation under strict formatting and feasibility requirements.

Model Setup: The evaluation includes GPT-5 through an API and open-weight models deployed locally with Hugging Face Transformers in PyTorch.Local models use bfloat16 precision with automated device mapping.
Hyperparameters: Generation uses top_p = 0.9 and a maximum of 2000 new tokens across models.
Hyperparameters: Open-weight models use temperature T = 0.6 and repetition penalty 1.05, with prompts and ChatML templates enforcing parseable JSON.
Query Representation: Travel queries are mapped into the formal tuple q = (cs, Cd, W, k, B, Clocal, π) through query extraction.
Query Representation: A 5-day example encodes Denver as the source, Iowa as the destination state, two cities, three travelers, and a $5,400 budget.
Query Representation: A 7-day example encodes San Antonio, three Tennessee cities, one traveler, a $7,150 budget, no flights, visitor-allowed accommodations, and sports and arts events.
Constraint Verification: The temporal evaluation implementation patches the original TripCraft verifier so POI visits overlapping intercity departures are correctly flagged as physically impossible.

E.2 Original Formulation and Limitation of Temporal Attraction Score

The original Temporal Attraction Score combines visit-duration and attraction-count likelihoods, but its Poisson component artificially compresses realistic itineraries into low scores. Normalizing that component restores an intuitive [0, 1] range.

E.2 Original Formulation and Limitation of Temporal Attraction Score: The original score evaluates realistic attraction time using a joint probabilistic model of visit duration and daily attraction count.
E.2 Original Formulation and Limitation of Temporal Attraction Score: The duration component uses a Gaussian distribution for the actual time spent at each attraction.
E.2 Original Formulation and Limitation of Temporal Attraction Score: In this model, d_i is actual duration, μ_i^d is expected duration, and σ_d captures duration variability.
E.2 Original Formulation and Limitation of Temporal Attraction Score: The daily attraction count follows a Poisson distribution whose λ represents the expected count for a traveler persona.
E.2 Original Formulation and Limitation of Temporal Attraction Score: The final Temporal Attraction Score averages the joint duration-and-count likelihood across attractions.
E.2 Original Formulation and Limitation of Temporal Attraction Score: For a typical laid-back traveler with λ ≈1.11, the Poisson term peaks near 0.36, artificially capping an otherwise perfect score around 0.35.
E.2 Original Formulation and Limitation of Temporal Attraction Score: The low cap can make near-optimal itineraries appear to perform poorly because the Poisson probability is strictly less than 1.
E.3 Normalized Improvement: Normalizing the Poisson term by its persona-specific maximum preserves relative count likelihoods while allowing a realistic itinerary to reach 1.0.

F.1 Failures of the Generative LLM Scheduler

The generative LLM scheduler produces severe duration, chronological, transit, and circadian errors when assigning timestamps and buffers. A deterministic backend resolves the same selected entities into feasible schedules, supporting a hybrid division between semantic selection and programmatic scheduling.

Generative scheduler failures: Generative LLM scheduling produces severe chronological and logical errors when assigning specific timestamps and transit buffers.These failures include invalid durations, overlapping activities, and infeasible timing.
Generative scheduler failures: Zero-duration accommodations and overlapping hotel and restaurant activities violate positive-duration and transit-buffer requirements.The LLM scheduled hotel stays from 10:42 to 10:42 and 12:00 to 12:00, then overlapped a pizzeria visit with the hotel stay.
Generative scheduler failures: A 03:30–04:30 lunch reservation demonstrates a circadian violation despite standard daylight constraints.The generated itinerary scheduled lunch at Ray’s in the City during the early morning.
Deterministic scheduling: The deterministic scheduler corrected arrival timing, inserted transit buffers, and maintained chronological order and appropriate meal windows.It scheduled hotel drop-off from 10:12 to 10:42, a Fox Theatre visit from 11:12 to 13:24, daytime lunch, and evening dinner.
Hybrid architecture: The case study concludes that semantic entity selection is better handled by multi-agent reasoning, while temporal-window assignment is best handled deterministically.TRIPPULSE uses strictly formatted JSON output for algorithmic scheduling and scoring.
Review-grounded experiential improvements: Review-grounded case studies report improved cultural alignment, scenic immersion, dining authenticity, comfort, and risk avoidance.The review-aware planner replaced generic or crowded venues with culturally immersive and peaceful alternatives, and filtered venues associated with hygiene or service issues.