AlterAtlas: Shifting Travel Planning from AI Generation to Validation via Persona-Driven Simulations

AI summary

Overview

• Research area: Human-Computer Interaction (HCI), specifically human-AI interaction, LLM-based planning agents, and persona-based simulation.
• Technical level: Intermediate. The paper is readable without deep machine learning background, but understanding the system requires familiarity with LLM agents, retrieval, geospatial grounding, and standard HCI study methodology.
• Scope: The paper introduces AlterAtlas, an interactive web system that reframes AI travel planning from one-shot itinerary generation into iterative, persona-driven simulation and validation, and evaluates it through a formative study (N=7), an expert rating study of 51 matched itinerary pairs, and a within-subjects user study (N=11).

What This Paper Is About

Existing AI travel assistants generate a finished itinerary from a prompt and treat the plan as the endpoint, leaving complex user constraints and possible plan failures unexamined. The authors argue this ignores how people actually plan: by setting protected anchors, mentally simulating how fatigue, hunger, and other states accumulate across a route, comparing alternatives, and iteratively revising. AlterAtlas addresses this by simulating how different traveler personas would experience a candidate itinerary, so users can validate and revise plans rather than just accept generated output. The paper is also framed as moving beyond generation alone: simulation is presented as an inspectable interaction layer for validation, comparison, and revision.

Key Contributions

1. Design implications from a formative study (N=7) describing how people plan travel and where existing AI tools fall short in supporting iterative workflows. These produced four design considerations: support priority-structured itineraries (D1), externalize route-level user states and constraints (D2), enable progressive refinement and comparison (D3), and surface evidence and simulation traces for plan inspection (D4).
2. A persona-based travel plan simulator that converts editable personas into typed simulation variables with update rules and dependencies, grounds itineraries in geospatial data (POI metadata, street view imagery, elevation, distance, travel time sampled every 400 meters), and generates step-level simulated perceptions and persona states.
3. AlterAtlas, an interactive travel planning system implementing a four-stage workflow (trip and persona setup, POI discovery and prioritization, candidate itinerary generation, and simulation-based validation and revision), shifting AI travel planning from one-shot generation to simulation-mediated validation.
4. Findings from an expert-driven study and a within-subjects user study (N=11) showing that simulation-guided revision improves expert-rated plan-persona alignment and that participants used the system to surface hidden constraints, compare alternatives, and build trust in final plans.

Main Findings

• Generated plans were largely executable in both conditions. Experts rated initial and simulation-updated itineraries as broadly feasible (Initial: M=8.098, SD=1.188; Updated: M=8.235, SD=1.335; Overall: M=8.167, SD=1.253). Evaluators still noted recurring failure modes including POI-ordering inconsistencies and routing inefficiencies.
• Simulation-guided revision improved personalization alignment. Updated plans scored higher on alignment than paired initial versions (Initial: M=7.784, SD=0.945; Updated: M=8.275, SD=0.918; overall M=8.029, SD=0.954). A Wilcoxon signed-rank test found a significant improvement (p=0.00201), with an average alignment change of +0.492 (SD=1.027; N change>0 = 31).
• Typical improvements were concrete and localized. Reordering stops to satisfy time-sensitive goals such as watching the sunset, removing stops to reduce routing burden, selecting POIs matching stated preferences, and adjusting time allocation to better reflect real user actions at each location.
• The same itinerary produced different revision suggestions under different personas. The paper states that validation becomes a function of who the plan is for rather than only what the route contains.
• Formative study: planning is structured around anchors and priorities. All seven participants (F1–F7) began with must-go anchors such as attractions, reservations, hotels, or appointments, and treated intermediate stops as flexible and largely interchangeable as long as anchors were preserved.
• Formative study: users mentally simulate accumulated constraints. Participants described checking restroom availability when bringing kids (F7), stress from crossings or parking (F5), fatigue from long walks or hills (F2), whether drives would be pretty and diverse (F4), and whether meal timing would fit shifting hunger (F5), with group planning making this harder (F1, F6, F7).
• Formative study: users want iterative refinement and preserved alternatives. Participants (F3, F4, F5, F6, F7) treated plans as provisional and wanted coarse summaries for filtering plus fine-grained control without rebuilding from scratch; one participant (F5) said "you need constant user input."
• Formative study: planning builds confidence. All participants noted verification and trust-building as integral; current AI tools were criticized for abstracting away reasoning (F5: "I don't want it to be a loading screen and give me the plan") and for requiring manual validation (F1: "If I do not get enough references I have to do my own homework and I hate that").
• User study results are only partially reported in the available content. The paper states that the within-subjects study (N=11) showed AlterAtlas empowered users to uncover hidden constraints, fluidly compare alternatives, and build trust in their final plans; the detailed survey, System Usability Scale, and interview findings are not included in the truncated content provided.
• Latency and cost. Across 100 generated plans, average end-to-end latency was 184.780 seconds (SD=25.643) with average token usage of 209,578.200 (SD=41,344.920). Of this, 146.740 seconds (SD=26.898) of latency and 167,552.600 tokens (SD=39,562.630) are fully blocking; the remainder is non-blocking. The authors note these values are highly variable and depend on tooling availability, model availability, user input, and plan length.
• Persona diversity in the expert study. The 51 synthetic personas had a mean pairwise cosine distance of 0.55 between persona text embeddings and a mean pairwise Bray-Curtis distance of 0.35 over preference field type distribution.
• Expert agreement. Two evaluators resolved disagreements by discussion to reach consensus ratings, with Cohen's kappa = 0.86.

Methodology in Plain English

The authors first interviewed seven frequent travel planners for 30 minutes each, analyzed recordings and notes with codebook thematic analysis by two researchers, and derived four design considerations. They then built AlterAtlas as a web application (Vue.js front end, FastAPI backend). Travelers are described in natural language and parsed into structured persona fields across 12 preference categories: travel style, pace, budget, safety, sustainability, accessibility, social, food, vibe, activity, culture, and cost. For each planning run, an LLM generates typed simulation variables (bounded integers, booleans, or enums), each with update rules describing how it changes across itinerary steps and dependencies on route or user-state factors; hunger, for example, is an integer from 0-100 that increases with travel time and decreases after meals in proportion to meal size, depending on elapsed time, fatigue, and distance walked. Fatigue and hunger are initialized by default, and users can view and edit any variable.

A discovery agent with Google Places and web search tools recommends POIs with justifications and linked evidence; users sort them into "must-go" and "want-to-go." A consolidation agent then builds itineraries that preserve must-go locations, prioritize want-to-go locations when possible, and add supporting stops, using a mapping tool and latitude-longitude waypoints for rerouting. Without any prioritized POIs, this agent functions similarly to existing travel planning agents such as TravelAgent.

For each candidate itinerary, users run one simulation per persona. Each step is grounded with geospatial data, and paths between POIs are treated as nodes sampled every 400 meters for street view images, elevation, distance, and travel time. A simulation agent walks the itinerary node by node, updating persona variables per the rules, producing a localized perception and updated persona state at each node, then an itinerary judgment scoring overall enjoyment plus suggested improvements. Users can scrub through time, compare traces, accept or reject edits, and rerun.

Evaluation used two studies. The first generated 51 synthetic personas with gemini-3-pro-preview (3-7 randomly generated preference fields, screened for realism and locality relevance), produced one initial day itinerary per persona with no POI prioritization as the baseline, then applied the simulation verification and revision pipeline and accepted all suggested improvements, yielding 51 pairs (102 total plans). Two local-area experts (Age: 23, 1M/1F, local experience of 10 years and 7 years) independently rated each plan on 1-10 Likert scales for feasibility and alignment, blind to condition, using a PDF with persona, route map, plan description, time per stop, and stop-level details; the process took approximately 9 minutes per evaluator per plan. The second study was a within-subjects user study with 11 participants (Age: 23-29; 4M/7F; P1–P11), compensated $30 for a 90-minute remote session, comparing AlterAtlas against each participant's self-enhanced AI workflow (shown a demo of Google Gemini 3.1 Pro with Google Maps grounding) across two 8+ hour walking-only day itinerary tasks in Chicago, Los Angeles, or New York City, with at least two travelers or traveler states required. Post-survey items used 7-point Likert scales plus the System Usability Scale.

Implementation details: Google Maps services provide place metadata, street view imagery, elevation, distance, and travel-time estimates; Tavily provides web search; POI discovery and consolidation agents use PydanticAI tuned through manual researcher validation; other LLM modules use DSPy optimized with BootstrapFewShot on a manually labeled dataset of persona-travel plan pairs (tool-calling agents were not DSPy optimized due to tool API limitations and cost). All components use gemini-3.0-flash-preview except images, which are converted to text descriptions with gemini-2.5-flash-lite for context and latency management. Outputs are validated against typed schemas with Pydantic validation and DSPy Refine.

Why This Matters

Impact on research. The paper reframes the design target of AI planning tools: instead of optimizing one-shot generation quality, it proposes simulation as an inspectable interaction layer that externalizes hidden constraints and makes validation a function of who the plan is for. It connects personas, geospatial grounding, and simulation-driven HCI into a single interaction model, and it links simulation variables to route context rather than evaluating single POIs in isolation, which prior work such as Accessibility Scout is described as doing. The expert study doubles as an ablation comparing a one-shot planning agent to the same agent plus simulation-based verification.

Real-world applications:

• Consumer travel planning apps that let travelers compare alternative itineraries and see how a companion, child, or parent would experience each one.
• Accessibility- and comfort-aware trip planning for users sensitive to walking distance, elevation, crowding, or restroom availability.
• Group travel coordination where conflicting needs, such as a traveler with dietary restrictions and a companion who dislikes long walks, must be balanced.
• Destination marketing, tourism boards, and hospitality platforms that want to preview how different visitor segments would experience a proposed route.

Industry relevance. Commercial tools named in the paper, including Layla, Booking.com AI Agents, TripGenie, and Gemini for Google Maps, currently emphasize generating or editing complete itineraries from high-level requests. AlterAtlas points to a complementary product direction: validation, contingency planning, and explanation as the paid value rather than raw generation. The reported cost profile (average 209,578.200 tokens and 184.780 seconds end-to-end per plan, with 146.740 seconds blocking) also indicates the practical engineering constraints any productized version would face.

Future Directions

• Extend beyond single-day, walking-only itineraries. The system deliberately focuses on route-level and micro-level constraint modeling for single-day, walking-only plans; multi-day trips and other transport modes remain open.
• Move from expert judgment of plans to field deployment. The expert study explicitly evaluates matched plan pairs rather than deployed trip outcomes, so whether simulation-guided revision changes actual trip satisfaction is untested.
• Validate and calibrate the simulation variables themselves. Simulation variable schemas are LLM-generated, and the paper does not report accuracy of simulated states against real traveler experiences; establishing ground truth for variables like fatigue, hunger, and stress would strengthen the approach.
• Reduce latency and cost. Blocking latency and token usage are substantial and variable, and the authors note dependence on tooling and model availability, leaving efficiency and caching strategies as open engineering questions.
• Scale persona evaluation and group planning. Testing how the approach behaves with many simultaneous personas, and whether participants' self-authored personas reproduce their real preferences, are unresolved questions.

Target Audience

This paper is most useful to HCI and human-AI interaction researchers working on planning agents, LLM-based assistants, and simulation-driven evaluation; to travel and hospitality product designers and developers building itinerary tools; to researchers in accessibility and personalized spatial assessment; and to practitioners interested in educational or urban-planning applications of persona simulation. Readers wanting implementation specifics will find the LLM stack, agent frameworks, and latency measurements directly applicable, while readers focused on design will find the formative study findings and the four design considerations the most transferable content.

Authors’ abstract

Travel planning requires balancing interacting goals and constraints across time and space. Current AI travel tools provide limited support for encoding these constraints and understanding how generated travel plans may fail users. We present AlterAtlas, an interactive travel planning system that supports high-fidelity itinerary validation and revision through persona-based simulations grounded in geospatial information. AlterAtlas models travelers as editable personas, generates candidate itineraries from prioritized places of interest, and simulates how different personas would experience each plan. Simulations expose route-level tradeoffs, temporal user states (e.g., fatigue, hunger), and mismatches between plans and user preferences to allow users to iteratively refine both itineraries and user personas. An expert evaluation of 51 paired itineraries demonstrates that simulation-guided revisions significantly improve plan-persona alignment. Furthermore, a within-subjects study (N=11) reveals that AlterAtlas empowers users to uncover hidden constraints, fluidly compare alternatives, and build trust in their final plans. Our results suggest that simulation-based validation is a powerful, transparent interaction layer for AI-assisted travel planning.