‹ back to 2026-07-16

HoloGeo: Mitigating Landmark Bias in Geo-localization via Evidence-Driven Reasoning

arXiv cs.CV Multimodal & Generative Pengcheng Zhou, Xuanyu Liu, Yanchen Yin, Bobo Li, Shengqiong Wu, Mong-Li Lee, Wynne Hsu 2026-07-16

TL;DR - HoloGeo is a research framework that tackles "landmark bias" in Vision-Language Models used for image geo-localization, where models over-rely on famous landmarks and misjudge location; it introduces metrics, benchmarks, and evidence-driven reasoning to make geospatial predictions more robust.

Defines two new metrics — Bias Intensity (BI) and Bias Harmfulness (BH) — to quantify how landmarks distort VLM reasoning, plus a benchmark, LandmarkBias-3K.
Proposes HoloGeo, an evidence-driven reasoning framework trained on BF-30k, a dataset annotated with structured multi-evidence, bias-free reasoning chains.
Uses multi-dimensional rewards to encourage balanced attention across diverse visual cues rather than fixating on landmarks.
Reports that HoloGeo maintains strong performance on IM2GPS3K and YFCC4k while outperforming existing open-source VLMs on LandmarkBias-3K (specific numbers not provided in the abstract).

view merged work →