Multilingual Agent System for Inclusive Wildfire Evacuation Guidance
Shruti Kulkarni, Lynn Tong, Aditi Namboodiripad, Chelyah Miller, Helen Lin, Peeyush Patel, Bogdan Bistriceanu, Diane Myung-kyung Woodbridge
cs.CY · cs.AI
Quick verdict
BEACON: sound multilingual wildfire evacuation system design, but all evidence is qualitative—no metrics, baselines, or user studies back the claims.
Summary
The paper presents BEACON, an end-to-end mobile system for wildfire evacuation guidance aimed at the 25.7 million US residents with limited English proficiency, given that over 80% of wireless emergency alerts and evacuation orders are issued only in English. The system combines an iOS client with a Python FastAPI backend that ingests fire perimeters, evacuation status, and shelter data from Watch Duty, hourly weather from NOAA's HRRR model, and points of interest from Overpass, cached in BigQuery. Fire danger is scored with the Hourly Wildfire Potential index over a 9x9 HRRR-kernel, and an XGBoost classifier both categorizes danger (safe/elevated/high/extreme) and dynamically schedules reassessment. Evacuation routes are computed with OpenRouteService under hard GeoJSON exclusion constraints covering fire polygons, evacuation zones, and hazard points, with optional drop-by waypoints to shelters. A six-block-prompt LLM agent manages routes, checklists, and language; UI language follows the user's most recent language, backed by Qdrant cross-session memory. Evaluation is entirely qualitative: in Colorado (July-September 2025, nine fires over 1,000 acres), the authors report reliable language detection, routes avoiding all active fire polygons where commercial APIs did not, first-turn context awareness, and six graceful-degradation scenarios. No quantitative metrics, baselines, sample sizes, latency figures, user study, or code release are provided.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 19 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- BEACON automatically detects the user's language from the first chat message and switches the entire application UI accordingly, honoring mid-session switches without configuration (III-A).
- Its polygon-avoidant routing pipeline (OpenRouteService with hard GeoJSON constraints over unioned fire perimeters, evacuation zones, and hazard points) returned routes avoiding every active fire polygon in a Colorado test where baseline commercial routing APIs crossed excluded regions (III-B).
- A six-block session prompt plus Qdrant cross-session memory gives the agent full context (danger category, loaded route, household profile, conversation summary) from the very first user turn (II-D, III-C).
- The system degrades gracefully across six enumerated failure modes while preserving escalation to 911 (III-D).
- An XGBoost classifier categorizes danger into four levels and dynamically schedules danger reassessment based on fire progress and trends (II-B) — claimed but not evaluated.
Strengths
Well-documented equity motivation
The Introduction grounds the problem in concrete sources: GAO-26-107680 reports over 80% of wireless emergency alerts and county-level evacuation orders are English-only, against 25.7M US residents with limited English proficiency (Census ACS [3]). The language feature covers chat, full UI, and route instructions, including mid-session switching persisted to the database (II-D, III-A).
Technically grounded danger and routing core
Danger scoring reuses the peer-reviewed Hourly Wildfire Potential index (Eq. 1, five HRRR surface fields, James et al. [4]) with a 9x9 (~27 km x 27 km) smoothing kernel (Eq. 2) rather than a bespoke heuristic; routing delegates to OpenRouteService with GeoJSON hard constraints over a clearly defined exclusion set A (union of fire perimeters, evacuation areas, HWP-flagged cells, hazardous points), giving a genuine avoidance guarantee rather than a soft penalty (II-C).
Explicit degradation engineering
Honest preliminary framing
Weaknesses
No quantitative evaluation whatsoever
The XGBoost cadence model is asserted, never evaluated
Anecdotal baseline comparison
Low reproducibility
Safety-validation gap for a life-safety loop
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This system paper introduces BEACON (Broadscale Evacuation Agentic Coordination, Outreach & Navigation), an iOS + FastAPI platform combining wildfire danger assessment, polygon-avoidant evacuation routing, personalized checklists, and a multilingual LLM agent. It explicitly targets the ~25.7 million US residents with limited English proficiency (LEP), against the backdrop that over 80% of wireless emergency alerts and county-level evacuation orders are English-only (GAO-26-107680 [2]) and the fire season has lengthened by 84 days since the 1970s. Central claims: (1) automatic per-user language detection and whole-UI language switching without configuration, honoring mid-session switches (III-A); (2) routing that provably avoids unioned fire perimeters, evacuation zones, and hazard points, where commercial APIs did not in preliminary tests (III-B); (3) full first-turn context via a six-block session prompt plus Qdrant cross-session memory (II-D, III-C); (4) graceful degradation across six enumerated failure modes with 911 escalation preserved (III-D); and (5) XGBoost-driven dynamic scheduling of danger reassessment (II-B). All results are self-described as 'preliminary' and 'qualitative observations' (III).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the BEACON wildfire evacuation system?
BEACON (Broadscale Evacuation Agentic Coordination, Outreach & Navigation) is a mobile app prototype from University of San Francisco researchers that combines wildfire danger assessment, evacuation routes that avoid fire perimeters, personalized checklists, and a multilingual LLM chat agent. It targets US residents with limited English proficiency, since over 80% of wireless emergency alerts and evacuation orders are issued only in English.
Which languages does BEACON support?
The paper does not fix an explicit language list for the agent: it detects the user's language from chat and switches the entire app UI accordingly, including mid-session changes. For routing, OpenRouteService natively supports 23 languages; beyond those, routes are computed in English and the turn-by-turn instructions are translated via Claude so the user can continue in their own language.
Was BEACON tested in real wildfires, and what were the results?
Testing used Colorado data from July to September 2025 covering nine fires of more than 1,000 acres each, obtained from the WiDS datathon on Kaggle. Reported results are qualitative only: reliable language detection, routes avoiding all active fire polygons (where commercial routing APIs crossed excluded regions in preliminary tests), context-aware responses from the first message, and six graceful-degradation scenarios. No quantitative metrics, latency figures, or user studies are given.
Where does BEACON get its fire and weather data?
Fire perimeters, evacuation order status, and shelter information come from Watch Duty, a non-profit wildfire tracker; weather inputs come from NOAA's High-Resolution Rapid Refresh (HRRR) model, used to compute an Hourly Wildfire Potential score; and points of interest such as pharmacies and shelters come from the Overpass API. All ingested data is cached in Google BigQuery.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-3 are visual; only their captions appear in the extracted text, so the rendered route/map examples (Fig. 2) and chat/checklist screenshots (Fig. 3) could not be inspected.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.