Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

GTPred: Benchmarking MLLMs for Interpretable Geo-localization and Time-of-capture Prediction

Jinnao Li, Zijian Chen, Tingzhu Chen, Changbo Wang

TL;DR

GTPred introduces a geo-temporal benchmark to evaluate multi-modal large language models on jointly predicting capture year and hierarchical location from images. It provides 370 samples spanning 120+ years, with a hierarchical Country→State→City→Place annotation and a novel evaluation protocol that combines year intervals with hierarchical location matching and reasoning traces scored by an LLM judge. Extensive offline experiments across 15 MLLMs reveal strong perceptual abilities but limited world knowledge and reasoning, with temporal information significantly boosting geo-localization accuracy. The work highlights the need for better temporal reasoning and world knowledge integration in geo-localization systems, offering a new benchmark and analysis framework for future improvements.

Abstract

Geo-localization aims to infer the geographic location where an image was captured using observable visual evidence. Traditional methods achieve impressive results through large-scale training on massive image corpora. With the emergence of multi-modal large language models (MLLMs), recent studies have explored their applications in geo-localization, benefiting from improved accuracy and interpretability. However, existing benchmarks largely ignore the temporal information inherent in images, which can further constrain the location. To bridge this gap, we introduce GTPred, a novel benchmark for geo-temporal prediction. GTPred comprises 370 globally distributed images spanning over 120 years. We evaluate MLLM predictions by jointly considering year and hierarchical location sequence matching, and further assess intermediate reasoning chains using meticulously annotated ground-truth reasoning processes. Experiments on 8 proprietary and 7 open-source MLLMs show that, despite strong visual perception, current models remain limited in world knowledge and geo-temporal reasoning. Results also demonstrate that incorporating temporal information significantly enhances location inference performance.

GTPred: Benchmarking MLLMs for Interpretable Geo-localization and Time-of-capture Prediction

TL;DR

GTPred introduces a geo-temporal benchmark to evaluate multi-modal large language models on jointly predicting capture year and hierarchical location from images. It provides 370 samples spanning 120+ years, with a hierarchical Country→State→City→Place annotation and a novel evaluation protocol that combines year intervals with hierarchical location matching and reasoning traces scored by an LLM judge. Extensive offline experiments across 15 MLLMs reveal strong perceptual abilities but limited world knowledge and reasoning, with temporal information significantly boosting geo-localization accuracy. The work highlights the need for better temporal reasoning and world knowledge integration in geo-localization systems, offering a new benchmark and analysis framework for future improvements.

Abstract

Geo-localization aims to infer the geographic location where an image was captured using observable visual evidence. Traditional methods achieve impressive results through large-scale training on massive image corpora. With the emergence of multi-modal large language models (MLLMs), recent studies have explored their applications in geo-localization, benefiting from improved accuracy and interpretability. However, existing benchmarks largely ignore the temporal information inherent in images, which can further constrain the location. To bridge this gap, we introduce GTPred, a novel benchmark for geo-temporal prediction. GTPred comprises 370 globally distributed images spanning over 120 years. We evaluate MLLM predictions by jointly considering year and hierarchical location sequence matching, and further assess intermediate reasoning chains using meticulously annotated ground-truth reasoning processes. Experiments on 8 proprietary and 7 open-source MLLMs show that, despite strong visual perception, current models remain limited in world knowledge and geo-temporal reasoning. Results also demonstrate that incorporating temporal information significantly enhances location inference performance.
Paper Structure (15 sections, 2 equations, 4 figures, 3 tables)

This paper contains 15 sections, 2 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Geo-localization Benchmarks
  4. Time-of-capture Prediction
  5. The GTPred Benchmark
  6. Overview of GTPred
  7. Data Curation Process
  8. Evaluation Protocol
  9. Experiments
  10. Experimental Setup
  11. Comparisons between MLLMs
  12. Scaling Behavior
  13. Case Study
  14. Ablation Study
  15. Conclusion

Figures (4)

  • Figure 1: Comparisons of the reasoning process between GPT-5.1 and the ground truth for a GTPred sample. The ground truth leverages the predicted year as a cue for reasoning about the location.
  • Figure 2: Overview of the GTPred. (a) Word clouds of the reasoning processes for year (upper) and location (lower). (b) Geographical distribution of the images in the GTPred. (c) Temporal distribution of the images in the GTPred.
  • Figure 3: Scaling behavior of different model families on our GTPred.
  • Figure 4: Examples of successful and failure cases from Qwen3-VL-4B and Qwen3-VL-32B. The results of Qwen3-VL-4B and Qwen3-VL-32B are separated by dashed lines. The green box indicates the signage in the image.