@inproceedings{chen-etal-2026-aeroverifier,
title = "{A}ero{V}erifier: A Typed Spatial Verifier for Frozen Vision-Language Models in Aerial Visual Question Answering",
author = "Chen, Jinfeng and
Yang, Mengfan and
Zhu, Jianyu and
Zhou, Zhicheng and
Liang, Jing",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1125/",
doi = "10.18653/2026.findings-emnlp.1125",
pages = "23826--23839",
ISBN = "979-8-89176-462-0",
abstract = "Drone-based applications such as navigation, inspection, and search depend on vision-language models (VLMs) that can answer spatial visual questions. The VLM faces three challenging tasks: grounding each referenced phrase to a scene object, choosing a viewpoint for spatial relations, and verifying the relation against accurate 3D geometry. In current frozen VLMs, all three tasks happen inside a single forward pass, and only the final answer is visible from outside. As a result, when that answer is wrong, we cannot tell which task failed. To address this, we introduce AeroVerifier, which exposes each of these three tasks as an explicit and checkable interface. Given a question, AeroVerifier compiles it into an executable typed program, runs the program over object-level 3D evidence, and returns the program answer together with stage-wise diagnostic signals to the base VLM as single-step revision feedback. On a 1000-row Open3D-VQA subset, both AeroVerifier{'}s program answer and its revision feedback yield higher answer accuracy than the Qwen2.5-VL-7B and GPT-5.4 API baselines."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="chen-etal-2026-aeroverifier">
<titleInfo>
<title>AeroVerifier: A Typed Spatial Verifier for Frozen Vision-Language Models in Aerial Visual Question Answering</title>
</titleInfo>
<name type="personal">
<namePart type="given">Jinfeng</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mengfan</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jianyu</namePart>
<namePart type="family">Zhu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhicheng</namePart>
<namePart type="family">Zhou</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jing</namePart>
<namePart type="family">Liang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Drone-based applications such as navigation, inspection, and search depend on vision-language models (VLMs) that can answer spatial visual questions. The VLM faces three challenging tasks: grounding each referenced phrase to a scene object, choosing a viewpoint for spatial relations, and verifying the relation against accurate 3D geometry. In current frozen VLMs, all three tasks happen inside a single forward pass, and only the final answer is visible from outside. As a result, when that answer is wrong, we cannot tell which task failed. To address this, we introduce AeroVerifier, which exposes each of these three tasks as an explicit and checkable interface. Given a question, AeroVerifier compiles it into an executable typed program, runs the program over object-level 3D evidence, and returns the program answer together with stage-wise diagnostic signals to the base VLM as single-step revision feedback. On a 1000-row Open3D-VQA subset, both AeroVerifier’s program answer and its revision feedback yield higher answer accuracy than the Qwen2.5-VL-7B and GPT-5.4 API baselines.</abstract>
<identifier type="citekey">chen-etal-2026-aeroverifier</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.1125</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1125/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>23826</start>
<end>23839</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T AeroVerifier: A Typed Spatial Verifier for Frozen Vision-Language Models in Aerial Visual Question Answering
%A Chen, Jinfeng
%A Yang, Mengfan
%A Zhu, Jianyu
%A Zhou, Zhicheng
%A Liang, Jing
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F chen-etal-2026-aeroverifier
%X Drone-based applications such as navigation, inspection, and search depend on vision-language models (VLMs) that can answer spatial visual questions. The VLM faces three challenging tasks: grounding each referenced phrase to a scene object, choosing a viewpoint for spatial relations, and verifying the relation against accurate 3D geometry. In current frozen VLMs, all three tasks happen inside a single forward pass, and only the final answer is visible from outside. As a result, when that answer is wrong, we cannot tell which task failed. To address this, we introduce AeroVerifier, which exposes each of these three tasks as an explicit and checkable interface. Given a question, AeroVerifier compiles it into an executable typed program, runs the program over object-level 3D evidence, and returns the program answer together with stage-wise diagnostic signals to the base VLM as single-step revision feedback. On a 1000-row Open3D-VQA subset, both AeroVerifier’s program answer and its revision feedback yield higher answer accuracy than the Qwen2.5-VL-7B and GPT-5.4 API baselines.
%R 10.18653/2026.findings-emnlp.1125
%U https://aclanthology.org/2026.findings-emnlp.1125/
%U https://doi.org/10.18653/2026.findings-emnlp.1125
%P 23826-23839
Markdown (Informal)
[AeroVerifier: A Typed Spatial Verifier for Frozen Vision-Language Models in Aerial Visual Question Answering](https://aclanthology.org/2026.findings-emnlp.1125/) (Chen et al., Findings 2026)
ACL