Source-linked AI summary
DisasterRoute: Evidence-Guided Dual-System Routing for Visual Disaster Question Answering
Tan Zhang, Xing Deng, Lu Zhang, Xiaofeng Zhu, Ping Hu
TL;DR
UAV disaster question answering must deliver timely, accurate judgments from potentially incomplete imagery under limited edge resources. Disaster-S1 makes direct option decisions, while DisasterRoute uses evidence-informed routing for selective DisasterVL review. On DisasterBench, selective review raises accuracy while invoking generation for a minority of questions and reducing latency relative to full reasoning.
Problem
UAV disaster question answering needs to balance timely, accurate judgments from potentially incomplete imagery with limited edge resources, while efficient direct decisions and their coordination with generative analysis remain underexplored.
Source: §1 · pp. 1–2Method
Disaster-S1 scores five answer options in one multimodal pass, and DisasterRoute combines confidence, evidence sufficiency, and option–state agreement to route selected questions to DisasterVL review.
Source: §5 · pp. 9–10Results
77.79% to 82.42% accuracy; generation is invoked for 23.68% of questions, with mean latency approximately 69% lower than applying DisasterVL full reasoning to every question.
Source: §5 · p. 10Takeaways & Limitations
These findings support selective coordination of direct decisions and generative review for disaster question answering.
Source: §5 · p. 10High candidate coverage does not ensure successful review.
Source: §4.2 · p. 9
Abstract
from arXiv · showhide
Visual question answering from low-altitude UAV imagery must deliver timely and accurate judgments to support disaster response under limited edge computing resources. Routine questions may be answered through direct decisions, whereas more challenging questions may benefit from step-by-step reasoning. However, existing research primarily focuses on improving models' visual understanding and generative reasoning capabilities, with limited exploration of efficient direct decision-making and its adaptive coordination with generative reasoning. We develop Disaster-S1 based on a vision-language model adapted to the disaster domain. Trained for option discrimination, it outputs a probability distribution over answer options in a single multimodal forward pass for fast decision-making. We further propose DisasterRoute, a dual-system routing framework that combines prediction confidence with disaster evidence to decide whether to accept the initial answer or invoke the disaster reasoning model DisasterVL for review. By allocating inference computation on demand, the framework retains the efficiency of direct decision-making and improves overall question-answering accuracy through selective review. Experiments on DisasterBench show that Disaster-S1 achieves superior accuracy with low latency, while DisasterRoute further improves accuracy through selective review at substantially lower latency than full reasoning.
1 Introduction
DisasterRoute addresses the need for timely, accurate UAV disaster assessments when imagery and edge resources are limited. It pairs fast option decisions with evidence-informed selective review to allocate additional reasoning to selected questions.
1 Introduction
- Visual disaster question answering must balance accuracy and response time because UAV imagery can leave hazard evidence incomplete while edge resources and deadlines are limited.Occlusion, viewpoint changes, and limited coverage can obscure the source, propagation path, or impact zone.Source: §1 · p. 1
- The research gap is efficient direct decision-making and its on-demand coordination with generative analysis, beyond work focused mainly on visual understanding, reasoning, and domain adaptation.Multiple-choice autoregressive answering incurs sequential decoding costs, motivating direct option scoring.Source: §1 · p. 2
- A model’s preference among options alone does not establish whether the image contains the evidence a question requires.The introduction motivates selective review that also considers evidence sufficiency and agreement between candidate explanations and scene state.Source: §1 · p. 2
- Disaster-S1 uses option-discrimination training to produce probabilities over five options in one multimodal pass without autoregressive generation.The model is built on a disaster-domain-adapted vision–language model.Source: §1 · p. 2
- DisasterRoute combines prediction confidence, question-specific evidence sufficiency, and option–state agreement to route answers to direct acceptance or DisasterVL review.When rescue conditions are met, one candidate outside the Top-2 is added to reduce the risk of omitting the correct answer.Source: §1 · p. 2
- The evaluation examines fast decisions and selective review on DisasterBench across accuracy, calibration, routing, candidate coverage, review outcomes, and resource costs.Latency, throughput, and GPU memory are also evaluated.Source: §1 · p. 2
2 Related Work
Prior work extends remote-sensing and disaster analysis toward multimodal interaction and reasoning, and explores direct decisions coordinated with deliberation. DisasterRoute applies these ideas to UAV disaster questions using evidence-aware selective review.
2.1 Multimodal Disaster Understanding
- DisasterBench covers low-altitude UAV questions across causal attribution, hazard propagation, damage analysis, and response decisions, alongside the domain-adapted DisasterVL model.Related work also includes models for grounded satellite-image dialogue, geographic domain alignment, temporal event monitoring, and disaster damage assessment.Source: §2.1 · p. 3
2.2 Jev-Style System 1 Decisions and Selective Review
- Jev-style systems return structured decisions over predefined options, with visual variants supporting image-conditioned candidate decisions and research examining when to escalate to generative reasoning.Prior work also notes that confidence alone may not determine whether a visual answer is trustworthy.Source: §2.2 · p. 3
- DisasterRoute combines confidence, question-specific evidence sufficiency, and option–state agreement to select DisasterVL review, with conditional candidate rescue beyond the initial Top-2.This applies direct decisions and selective review to UAV-based disaster question answering.Source: §2.2 · p. 3
3 Methodology
DisasterRoute combines a fast five-option decision with evidence-state assessment to decide whether to accept the initial answer or send it for constrained DisasterVL review. Only selected questions receive generative review.
3.1 System 1: Single-Pass Disaster Decision
- Disaster-S1 reads five answer-letter logits at the answer position, normalizes them into option probabilities, and reuses the same representation for evidence-state prediction.Validation-selected temperature supplies inference probabilities while preserving the option ranking.Source: §3.1 · p. 4
3.2 Cross-Fitted Evidence States and Routing
- The evidence-state head predicts nine disaster-context fields, including severity, access, exposure, and evidence sufficiency, as fallible hypotheses rather than verified observations.The supervision targets are built from training-split disaster annotations, image–question evidence annotations, and predefined option-semantic rules.Source: §3.2 · pp. 4–5
- Cross-fitting assigns each training example a state prediction from a head not trained on that example’s state labels; a final head is then fitted on the full training split.Validation data determine calibration and routing parameters, while test labels do not participate in state prediction or route construction.Source: §3.2 · p. 5
- Routing compares predicted severity, access, and exposure with option text, then combines confidence, option–state conflict, evidence sufficiency, and candidate rescue to decide whether to review.A rescue candidate outside the Top-2 is activated only when it has sufficient combined support relative to the runner-up and substantially stronger state agreement than the initial answer.Source: §3.2 · p. 5
3.3 Constrained Selective Review
- DisasterVL reviews routed questions using the original image, question, and two or three original-letter candidates; invalid or out-of-set outputs fall back to the initial answer.Review can add an option outside the Top-2 when rescue conditions are met, and reviewed examples require image re-encoding and autoregressive decoding.Source: §3.3 · p. 6
- The state head reuses the System 1 representation, but routed examples incur image re-encoding and autoregressive review that may dominate mean and tail latency.Unreviewed examples keep the System 1 prediction; invalid or out-of-set review outputs also revert to it.Source: §3.3 · p. 6
4 Experiments
DisasterBench experiments assess Disaster-S1 and DisasterRoute for accuracy, routing, and inference efficiency. Selective review raises accuracy over direct decisions while using less mean latency than full reasoning, with additional latency and memory costs over direct inference.
4.1 Experimental Setup
- The test split contains 2,998 five-option image–question pairs; model fitting and routing selection use training or validation data, with test labels reserved for final evaluation.Timing includes decoding, preprocessing, inference, routing, and answer parsing, but excludes model loading and warm-up; latency comparisons are not strictly paired.Source: §4.1 · pp. 6–7
4.2 Results and Analysis
- 77.79% accuracy: S1-Direct exceeds DVL-Logit’s 72.25% at similar mean latency, supporting the benefit of option-discrimination training for direct decisions.The comparison does not isolate reasoning length because prompts, generation limits, and output constraints differ across protocols.Source: §4.2 · p. 7
- 82.42% accuracy: DisasterRoute answers 2,471 questions correctly versus 2,332 for S1-Direct, a 4.64-point gain, while using 1,215.01 ms mean latency versus 3,915.24 ms for DVL-FR.DisasterRoute reviews 710 questions (23.68% GCR); P95 latency is slightly higher than DVL-FR’s, and peak GPU memory rises from 4.37 to 8.55 GiB versus S1-Direct.Source: §4.2 · p. 7, Table 1 · p. 7
- Temperature scaling lowers ECE from 0.16 to 0.03, while TS+RS achieves the highest AUROC (0.8453) and RER (76.19%) at a 10.51% routing rate.These are standalone risk-selection diagnostics, not results for the complete DisasterRoute gate; TS+LC instead has the highest DAA (86.46%) at a 14.74% routing rate.Source: §4.2 · pp. 7–8, Table 2 · p. 8
- Option–state conflict and candidate rescue each reach 100.00% RER, but their candidate coverage differs at 97.31% and 89.12%, respectively.Triggered subsets overlap, and their union comprises the 710 questions reviewed by DisasterVL.Source: §4.2 · p. 8, Table 3 · p. 8
- Among 710 reviewed questions, candidate coverage is 93.52% but final accuracy is 32.96%; direct acceptance covers 2,288 questions at 97.77% accuracy.Review changes 157 predictions and corrects 143 initial errors, for a net gain of 139 correct answers over S1-Direct.Source: §4.2 · p. 8, Table 4 · p. 8
- DisasterRoute reaches 82.42% accuracy at 1.22 s mean latency, compared with 55.34% at 6.29 s for the best of three general-purpose VLM baselines, InternVL3.5-2B.General-purpose baselines were evaluated in one run, while the remaining results came from the three-run main comparison.Source: §4.2 · p. 9, Table 5 · p. 9
5 Conclusion
DisasterRoute coordinates fast option decisions with selective DisasterVL review for UAV-based disaster question answering. On DisasterBench, it improves accuracy over direct decisions while using substantially less mean latency than full reasoning.
5 Conclusion
- Disaster-S1 uses option-discrimination training to produce probabilities for five options in one multimodal forward pass without autoregressive generation.Source: §5 · p. 9
- Accuracy rises from 77.79% to 82.42% with generation for 23.68% of questions, while mean latency falls approximately 69% versus full DisasterVL reasoning.The framework incurs additional latency and memory costs over direct inference.Source: §5 · p. 10
- DisasterRoute routes questions to DisasterVL using prediction confidence, question-specific evidence sufficiency, and option–state agreement, with conditional rescue expanding candidates beyond the initial Top-2.Source: §5 · p. 10
- High candidate coverage does not ensure successful review, and tail latency remains a limitation.Future work targets evidence estimation, candidate comparison, review costs, and generalization across disaster settings.Source: §5 · p. 10