cs.CVOct 9, 2026

Benchmarking Answer-Grounding Consistency in Remote Sensing Visual Question Answering

Authors: Ping Guo, Zhiqi Huang, Xinran Li, Linkun Yu, Jingyin Pei, Kamata Seiichiro

Organizations: Graduate School of Information, Production and Systems, Waseda University, Kitakyushu, Japan · School of Artificial Intelligence, The Chinese University of Hong Kong, Shenzhen, China · School of Software, Dalian University of Technology, Dalian, China · Graduate School of Fundamental Science and Engineering, Waseda University, Tokyo, Japan

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

In grounded remote sensing visual question answering, a model produces a textual answer and identifies relevant objects in the image as grounding evidence. Existing evaluations measure answer correctness and localization quality, but do not directly assess answer-grounding consistency. We propose the Answer-Evidence Support Protocol (AESP) to assess this consistency through evidence support: whether the grounding evidence supports the model's own answer. AESP applies task-specific rules to existence, counting, and comparison questions. We further construct the Grounded Remote Sensing Visual Question Answering (G-RSVQA) benchmark, containing 15,027 samples with questions, reference answers, and all related object-level annotations. Experiments on seven grounded vision--language models reveal failures of evidence support across all three question types. Up to 50.82% of correct answers in the evidence evaluation subset are not supported by the provided grounding evidence.

Figures & tables

Figure 1: Conventional evaluation measures answer correctness and localization quality; AESP additionally assesses answer-grounding consistency.
Figure 2: Overview of G-RSVQA construction. The rightmost panel shows a sample and its annotation format.
Figure 3: Question types and data sources of G-RSVQA.
Figure 4: Illustration of AESP evaluation protocol for assessing whether grounding evidence supports a model’s answer. Each question is paired with supporting and non-supporting evidence across existence, counting, and comparison tasks.
Table 1: Evaluation of seven vision–language models on G-RSVQA under the proposed protocol. All values are percentages; the best result in each column is bold.
ExtractorExistenceCountingComparisonOverall
Qwen3-8B [ 16 ]99.3283.0288.8995.75
Qwen3-14B [ 16 ]99.6696.2394.4498.50
Mistral-NeMo-12B [ 11 ]96.2567.9292.5992.00
Table 2: Answer extraction accuracy (%) on 400 VLM responses, evaluated against human reference annotations.
Table 3: Overall evidence support and joint success rates (%) by question type on G-RSVQA.
More in this topic

Remote Sensing VQA

Remote sensing visual question answering (VQA) studies how AI answers natural-language questions about satellite, aerial, and other Earth-observation imagery.

See all 35 papers

Explore similar work

CardsList

  1. TerraAgent: Demonstration-Grounded Multi-Agent Reasoning for Long-Context Remote Sensing Visual Question Answering

    Oct 7, 2026Jingwen Huang, Yiming Fang, Hanxiao Wu +2Remote Sensing VQAVisual Question Answering

    Yunnan University of Finance and Economics, Kunming, China

  2. VISTAQA: Benchmarking Joint Visual Question Answering and Pixel-Level Evidence

    May 20, 2026Mozhgan Nasr Azadani, Yimu Wang, Yongpeng Zhu +5VLM EvaluationVisual Question Answering

    University of Waterloo · Stanford University · NVIDIA

  3. Knowing Whether, Not Just Where: Anchor-Probing Distillation for Open-Set Remote Sensing Visual Grounding

    Oct 9, 2026Zhihao Chang, Lei Zhai, Shuyuan Yang +1Visual GroundingRemote Sensing Visual Grounding