AI Knows Where You Are: Exposure, Bias, and Inference in Multimodal Geolocation with KoreaGEO

arXiv:2506.03371 · cs.CV · Submitted 2025-06-03 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "AI Knows Where You Are".

Jane: The KoreaGEO Bench introduces a fine-grained, multimodal geolocation benchmark specifically designed for Korean street views to systematically evaluate the accuracy, spatial bias,

Tom: First, who's behind it and why it matters.

Title and authors: Tom: Now, let’s look closer at the actual KoreaGEO Bench setup described in "AI Knows Where You Are: Exposure, Bias, and Inference in Multimodal Geolocation with KoreaGEO". It’s not just a dataset; it’s a carefully constructed environment.

Jane: Yeah. The summary talks about how they built this dataset using urban stratification—grouping cities by population density and GRDP to get those four clusters—and then they layered that on top of nine distinct place types.

Lu: They went into detail about the city clustering, explaining that they used KMeans clustering based on population density and GRDP data from KOSIS to derive socio-economic feature vectors for each urban functional unit. That’s how the four clusters came about.

Meng: So it’s not just picking random cities; it’s a systematic way to ensure you get representation across the entire socioeconomic spectrum of Korean cities, from Seoul to other regional areas. That sounds like a necessary step for any good benchmark.

Tom: And then they covered functional diversity by defining those nine place types—Commercial and Recreational Zones, Educational and Cultural Institutions, Residential and Living Areas—to make sure the images aren't just all one type of scene.

Jane: And they made it rich by adding multicontextual annotations, which means tagging the images with a lot more information than just what’s in the picture visually.

Lu: Plus, they simulated real-world privacy risks by generating two styles of Korean captions: one that prohibits place names but requires emotion and hashtags, and another style that explicitly includes real location names or addresses.

Meng: That caption simulation is key because it mimics how people actually post things on social media where location exposure is a constant factor. It’s not just about the image; it's about the text context around it.

Tom: And then they defined their evaluation protocol—that three-path system comparing ten different vision-language models under varying input modalities, which is what makes this benchmark so systematic for diagnostics.

Jane: It really shows that they aren't just throwing images at a model and hoping for the best; they are systematically testing the performance across these specific conditions to see where the weaknesses are.

Lu: The idea here is to enable fine-grained model diagnostics by analyzing geolocation accuracy, bias patterns, and reasoning behavior under these different inputs, which is what they aim to achieve in this paper.

Meng: For deployment, knowing exactly how a model performs in path three versus path one gives us a clearer picture of the trade-offs between precision and privacy exposure that we need to manage.

Tom: And it ties all that back together: KoreaGEO Bench is the first fine-grained, multimodal geolocation benchmark for Korean street views, aiming to systematically evaluate accuracy, spatial bias, and reasoning behavior in real-world social media contexts.

Jane: It sets a high bar for how we need to evaluate these models when location privacy is on the line.

The paper's summary: Tom: Moving into the suggested improvements section of "AI Knows Where You Are: Exposure, Bias, and Inference in Multimodal Geolocation with KoreaGEO". The authors are suggesting ways to make this work even more robust.

Jane: They suggest testing model sensitivity across input modalities explicitly because they found that modality sensitivity has a greater impact on localization and privacy exposure than just the model architecture or size alone.

Lu: That’s a direct recommendation for researchers to pinpoint whether performance degradation is stemming from visual-only reliance or if it’s due to text-based reasoning when assessing those risks, which is crucial for targeted intervention.

Meng: From a production standpoint, this means we need to be able to diagnose if our system breaks because it can't interpret the image well on its own, or because it fails when that text context is missing.

Tom: They also propose using the three-path protocol to systematically analyze geolocation accuracy, bias patterns, and reasoning behavior as a way to enable fine-grained model diagnostics against structural biases in the data.

Jane: That’s how they want us to use this structure not just for scoring, but as a diagnostic tool to find and correct those inherent biases that show up when we look at the data across different clusters.

Lu: They also suggest integrating privacy-aware caption generation by using those two caption styles to simulate real-world exposure, which mimics human-style multimodal communication in social contexts.

Meng: By simulating both functional and high-risk captions, we can train AI systems to learn robust reasoning under varied privacy constraints and compare performance between the two settings directly.

Tom: And the finding that high-risk captions sharply boost accuracy in path three shows us that explicit location names drive precision, which is a key piece of information for future development.

Jane: So they are emphasizing that we need to focus on this input modality as the most important variable when designing systems for reliable geo-inference.

The paper's improvements: Tom: To wrap up "AI Knows Where You Are: Exposure, Bias, and Inference in Multimodal Geolocation with KoreaGEO", the paper shows us that the KoreaGEO Bench is a diagnostic foundation for building more reliable and privacy-aware geo-inference systems.

Jane: It’s important to remember that we have to be aware of the structural biases—like how performance converges heavily on core cities, and how non-core clusters get misclassified as Seoul.

Lu: The KoreaGEO Bench provides a concrete way to diagnose these patterns, exposing structural biases and reasoning patterns that can guide development toward fairness.

Meng: For us in production, this means we need to use this diagnostic foundation to build systems that are explicitly privacy-aware and not just blindly trust the visual input.

Tom: The main conclusion is that the input modality has a greater influence on localization performance and privacy exposure than model scale or architecture, with high-risk captions significantly amplifying precision and risk.

Jane: That’s a clear message: when you build geo-inference systems, you need to design for the context of real social media communication right from the start.

Lu: The study adheres strictly to the Google Street View terms of use, ensuring that all procedures in involving imagery access and usage remain within permissible boundaries.

Meng: It’s a good starting point because it flags where we need to focus our engineering efforts for responsible AI development.

Tom: That’s all for this discussion on "AI Knows Where You Are: Exposure, Bias, and Inference in Multimodal Geolocation with KoreaGEO".

Conclusion: Tom: So we’ve looked at how the KoreaGEO Bench systematically tests vision language models on Korean street views to see where they get inaccurate or biased when people are posting about real places and privacy is involved.

Jane: Right, it's about that three-path evaluation protocol comparing image only, image with a functional caption, and image with a high-risk caption to see how input modality really shifts the performance.

Lu: The authors found that the input modality has a greater influence on localization precision and privacy exposure than model scale or architecture alone.

Meng: So if we’re shipping models, this tells us that what they feed the AI matters more than just making the underlying network bigger, which is good for figuring out how to deploy these things safely.

Lalam: I can confirm that high-risk captions significantly amplify precision and risk when you look at models like o3 and GPT-4o, especially at one kilometer distance <ref:2506.03371#pg2>.

Tom: That’s a big number, a thirty percent boost in accuracy under those high-risk conditions, which is definitely something we need to keep an eye on.

Jane: They also pointed out that touristic and iconic landmarks pose the highest privacy risk at short distances because they are so recognizable visually.

Lu: And they showed a structural bias: the model performs best in core city clusters, Cluster zero while samples from Cluster two and Cluster three get misclassified as Seoul all the time.

Meng: That clustering bias is concerning because it means if we only train on core data, our models will struggle with regional diversity unless we actively fix that representation gap.

Tom: It’s a clear diagnostic tool for building more reliable and privacy-aware systems by exposing those structural biases and reasoning patterns in the KoreaGEO Bench.

Jane: It’s about understanding how these models actually reason when they see text versus just an image, which is a crucial missing piece for developing safer applications.

Lu: The limitation they flag is that the study relies on publicly available street-view imagery, so it doesn't fully account for regional linguistic variations like local dialects or spelling differences.

Meng: That’s a real limitation for deployment; if our system needs to work in a small town in Busan and not Seoul, this dataset might miss the specific challenges there.

Lalam: The KoreaGEO Bench is the first fine-grained, multimodal geolocation benchmark specifically designed for Korean street views to systematically evaluate accuracy, spatial bias, and reasoning behavior.

Tom: It’s a useful diagnostic foundation because it shows us exactly how much text context helps boost precision versus just seeing the image.

Jane: So if you’re building a system that needs to be robust in social media contexts, knowing how input modality impacts risk is essential for designing that system.

Lu: We need to keep looking at these structural biases and reasoning patterns because they show us where the current understanding of geospatial AI is falling short.

Meng: Next up, we’ll look at EPIC-Bench to see if we can apply this fine-grained grounding idea to other embodied environments.

Yonsei University · CISPA Helmholtz Center for Information Security

cs.CV

Submitted: 2025-06-03

Updated: 2026-10-15

Journal ref: Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 83/100

The gist: The KoreaGEO Bench introduces a fine-grained, multimodal geolocation benchmark specifically designed for Korean street views to systematically evaluate the accuracy, spatial bias, and reasoning

Key concepts

KoreaGEO Bench
A fine-grained benchmark built on 1,080 Korean street view images covering four urban clusters and nine place types. It includes two caption styles simulating real-world privacy risks to systematically evaluate geolocation accuracy.
Multi-Level Sampling Strategy
The dataset is structured by grouping cities into four clusters based on population density and GRDP, defining nine place types, and adding contextual details like indoor/outdoor status. This ensures the benchmark covers diverse Korean landscapes and real-world conditions.
Three-Path Evaluation Protocol
Models are tested using three input paths: image only, image plus a functional caption (simulating social media), and image plus a high-risk caption (containing specific location names). This allows researchers to precisely measure how different text inputs affect the model's ability to localize.
Modality Influence
The study found that the input modality—whether it's just an image, a functional caption, or a high-risk caption—has a stronger effect on localization accuracy and privacy exposure than changes in the model's scale or architecture.

Terminology

Summary

The KoreaGEO Bench introduces a fine-grained, multimodal geolocation benchmark specifically designed for Korean street views to systematically evaluate the accuracy, spatial bias, and reasoning behavior of vision-language models in real-world social media contexts. The results reveal that input modality has a greater influence on localization performance and privacy exposure than model scale or architecture, with high-risk captions significantly amplifying precision and risk.

KoreaGEO Bench Construction

The dataset comprises 1,080 high-resolution images sampled across four urban clusters and nine place types, enriched with multicontextual annotations and two styles of Korean captions simulating real-world privacy exposure Our dataset comprises 1,080 high-resolution images sampled across four urban clusters and nine place types, enriched with multicontextual annotations and two styles of Korean captions simulating real-world privacy exposure To address these limitations, we construct KoreaGEO Bench, which contains 1,080 streetview images covering four city clusters and nine place types, with rich annotations

Multi-Level Sampling Strategy

The sampling strategy is designed across three dimensions: urban stratification, functional coverage, and contextual diversity

  1. City Clustering: This involves grouping cities based on population density and Gross Regional Domestic Product (GRDP) using KMeans clustering to derive four clusters representing different development levels and spatial patterns

  2. Place-Type Sampling: Nine high-level semantic categories were defined, including Commercial and Recreational Zones, Educational and Cultural Institutions, and Residential and Living Areas to capture functional diversity across Korean Landscapes

  3. Spatio-Contextual Coverage: The sampling spans all first-level administrative divisions in Korea, covers all four seasons, different times of day, various weather conditions, and includes an indoor/outdoor dimension to acknowledge unintentional privacy exposure through indoor images

Caption Generation Simulation

Two types of Korean captions are generated to simulate real-world privacy risks

  1. Functional Caption Prompt: This prompt prohibits place names and requires the caption to express a mood, include an emotion, and use at least one emoji and 1–3 hashtags without proper nouns

  2. High-Risk Caption Prompt: This prompt requires the inclusion of real location names or addresses in the caption to simulate higher levels of location exposure

Evaluation Protocol

A three-path evaluation protocol is designed to systematically compare ten mainstream VLMs across input modalities

  1. First pathway: Models receive only the image and are asked to predict the corresponding address and geographic coordinates, along with an explanation of its reasoning process

  2. Second path: Models are given the image plus a functional caption that simulates a typical social-media post, requiring prediction and explanation

  3. Third path: Models are given the image together with a high-risk caption that reveals specific location information, requiring only the predicted address and coordinates without an explanation

Key Findings on Exposure and Bias

The results show modality-driven shifts in localization precision and highlight structural prediction biases toward core cities

Exposure Risk Analysis by Input Modality:

Image-only input yields poor performance, while high-risk captions sharply boost accuracy, with o3, GPT-4o, and Gemini surpassing 45% at 1km

Risk Analysis by Place Type:

Touristic and Iconic Landmarks pose the highest privacy risk at distances of 100 meters to 1 kilometer due to their high recognizability driven by distinctive visual features and high public visibility

Bias Analysis:

The model achieves the highest accuracy in high-density core regions (Cluster 0), indicating stable recognition of capital-type urban areas, while a large number of samples from Cluster 2 and Cluster 3 are misclassified as Cluster 0

Reasoning Behavior Analysis

Under the image-only condition, models demonstrate scene understanding by relying on visual cues such as architectural features, street and spatial layout elements, signage, and natural elements

Modality attribution analysis shows that most predictions are judged as mostly visual with some text under the visual with functional caption setting

The KoreaGEO Bench provides a diagnostic foundation for building more reliable and privacy-aware geo-inference systems by exposing structural biases and reasoning patterns The study adheres strictly to the Google Street View terms of use, ensuring that all procedures in involving imagery access and usage remain within permissible boundaries Given that Korea GEO Bench is designed to reveal biases and privacy exposure risks in visionlanguage models, we explicitly state that this dataset is intended solely for the purpose of mitigating algorithmic bias, improving model fairness, and promoting the development of responsible AI systems

The limitations noted include reliance on publicly available street-view imagery which presents clear limitations in spatial coverage and a lack of systematic consideration for regional linguistic features such as local dialects or spelling variations The study concludes by emphasizing that input modality has a greater influence on localization performance and privacy exposure than model scale or architecture, with high-risk captions significantly amplifying precision and risk.

The gist

KoreaGEO Bench introduces a fine-grained, multimodal geolocation benchmark specifically designed for Korean street views to systematically evaluate the accuracy, spatial bias, and reasoning behavior of vision-language models in real-world social media contexts.

How it works

The KoreaGEO Bench is constructed using a multi-level sampling strategy across three dimensions: urban stratification, functional coverage, and contextual diversity This includes city clustering based on population density and GRDP to create four clusters representing different development levels Place types are defined through a three-stage pipeline involving theoretical reference, local alignment using the National Land Planning and Utilization Act, and scene validation Finally, contextual features such as indoor/outdoor status, time of day, and cultural style are manually annotated to capture rich contextual variations in Korean street scenes

Evaluation Protocol

The evaluation employs a three-path protocol to compare ten mainstream VLMs across different input modalities The paths include receiving only the image, receiving the image plus a functional caption, and receiving the image together with a high-risk caption This allows for a fine-grained comparison of how input modality affects geolocation performance and reasoning behavior across various models

Key Analysis

The analysis focuses on exposure risk by comparing performance under visual-only, functional caption, and high-risk caption settings Findings indicate that high-risk captions sharply boost accuracy for models like o3 and GPT-4o, suggesting that explicit textual cues drive precise localization even in weaker models Structural biases are also revealed, as model performance is most stable for core cities (Cluster 0), with frequent misclassifications of non-core city clusters into the capital region

The study concludes that input modality has a greater influence on localization performance and privacy exposure than model scale or architecture, with high-risk captions significantly amplifying precision and risk. The KoreaGEO Bench provides a diagnostic foundation for building more reliable and privacy-aware geo-inference systems.

Improvements for AI systems

  1. Confidence in localization will be enhanced by explicitly testing model sensitivity across input modalities, as modality sensitivity has a greater impact on localization and privacy exposure than model architecture or size alone. This allows developers to pinpoint whether performance degradation stems from visual-only reliance or text-based reasoning when assessing privacy risks.

  2. Fine-grained bias mitigation will be achieved by systematically diagnosing spatial biases using the proposed three-path protocol, which analyze their geolocation accuracy, bias patterns, and reasoning behavior. This enables targeted intervention against structural biases observed in the data, such as how prediction errors converging heavily on Seoul indicate a lack of generalization toward non-core city clusters.

  3. Privacy-aware caption generation will be integrated by using the two caption styles to simulate real-world exposure, thereby mimicking human-style multimodal communication in social contexts. This allows AI systems to learn robust reasoning under varied privacy constraints, specifically by comparing performance between the functional and high-risk captions to understand how explicit location names sharply boost accuracy when used in the third path.

Sources

Related papers