Search papers, labs, and topics across Lattice.
This paper introduces SIGNPOST-Bench, a novel benchmark designed to evaluate how multimodal large language models (MLLMs) resolve conflicts between textual and visual cues in real-world scenes. By creating a controlled set of counterfactual image variants, the study reveals significant increases in localization error when models encounter adversarial text, with median errors rising from 282 km to 1,347 km. The findings highlight the nuanced effects of different types of text interventions on model predictions, underscoring the importance of robust evaluation frameworks for MLLMs in multimodal contexts.
Adversarial text can increase localization errors in MLLMs by nearly five times, revealing critical vulnerabilities in multimodal reasoning.
Multimodal large language models (MLLMs) make grounded predictions in real-world scenes by combining visual and textual cues, yet existing benchmarks rarely reveal how they arbitrate between these evidence sources when they conflict. We introduce SIGNPOST-Bench, a controlled counterfactual benchmark for evaluating text-vision conflict resolution. Each source image is transformed into a counterfactual quintuplet of Original, Blank, Similar, Random, and Adversarial variants. Synthetic, localized scene-text interventions are designed to preserve non-textual content, enabling paired measurements of changes in localization performance and directed shifts toward geographic targets introduced by conflicting text. SIGNPOST-Bench contains 5,111 counterfactual groups and 25,555 image variants from four datasets. We evaluate 20 MLLMs from seven providers. Compared with Original images, Adversarial variants raise median localization error from 282 km to 1,347 km, a 4.8-fold increase. Among geocodable adversarial samples, 6.5-20.1% of predictions lie less than 50 km from the injected target across models, and every evaluated model exhibits a positive mean paired reduction in target distance from Blank to Adversarial. Compatible, unrelated, and conflicting text replacements produce distinct effects on model predictions, while clean-input localization performance does not fully predict robustness to conflicting text. These results establish visual geolocation as a continuous diagnostic of scene-text arbitration and provide a controlled framework for evaluating how MLLMs resolve conflicting multimodal evidence.
