Sirun Li, Minghao Liu, Ling Dai, Yong Li, Haoxin Lyu, Junting Zhou, Fan Zhang
4 min
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.
As multimodal large language models (MLLMs) increasingly rely on both visual context and scene text to perform grounded reasoning, it remains unclear how they arbitrate between these signals when they provide contradictory information. This study investigates whether MLLMs prioritize visual evidence, textual evidence, or a combination of both when performing visual geolocation, a task that provides a continuous coordinate space to measure the magnitude and direction of model errors.
The authors introduce SIGNPOST-Bench, a controlled counterfactual benchmark consisting of 5,111 groups of images. Each group contains five variants of a source image: Original (unmodified), Blank (text removed), Similar (compatible text), Random (unrelated text), and Adversarial (conflicting geographic text). By using localized synthetic edits that preserve the surrounding visual context, the researchers can isolate the influence of text on model predictions. They evaluated 20 MLLMs from seven providers to measure localization error, the shift in predictions toward injected geographic targets, and the overall robustness of the models to textual interference.
The study reveals that conflicting scene text substantially impairs model performance. Compared to original images, adversarial variants increased the median localization error by 4.8 times, rising from 282 km to 1,347 km. Furthermore, the models exhibited a clear bias toward the injected text: in 6.5% to 20.1% of adversarial cases, the model's prediction fell within 50 km of the injected target. The researchers also found that a model's performance on clean, original images does not reliably predict its robustness to conflicting text, suggesting that current benchmarks focusing on aggregate accuracy fail to capture how models resolve multimodal conflicts.
This work demonstrates that even advanced MLLMs struggle to maintain grounded, vision-centric reasoning when presented with misleading textual cues. By establishing visual geolocation as a diagnostic tool, the authors provide a framework for developers to evaluate and improve the reliability of MLLMs in real-world scenarios where visual and textual evidence may conflict, such as in navigation or urban monitoring tasks.
Alex: So a model that's otherwise quite capable at geography can still be steered in the wrong direction by a single fake sign?
Sam: That's the core finding. And what makes it particularly worth paying attention to is that standard performance tests don't predict it. A model can score well on ordinary geography tasks and still be highly susceptible to this kind of conflict. The researchers call this "conflict robustness," and they're treating it as a distinct capability—separate from general accuracy. Recognizing a scene is one skill; correctly weighing a scene against contradictory text is another.
Alex: It does make you think about where else this kind of conflict might appear. Are there limitations the authors flag?
Sam: A couple worth noting. First, the images are synthetically edited—signs are swapped in digitally—and that process can introduce subtle visual artifacts that a real photograph wouldn't have. It's possible the models are responding to those artifacts in some way, not just the text itself. Second, these findings are specific to geographic reasoning. Whether the same pattern holds in other domains—medical imaging, for instance, where a label might contradict what a scan actually shows—is an open question the paper doesn't answer.
Alex: So the work points to a real gap, but there's more to understand about how far it extends.
Sam: Exactly. By using geographic coordinates as a measuring tool, the researchers have found a precise, continuous way to track how much a model "trusts" text over visual evidence. That kind of diagnostic is a meaningful step toward understanding how these systems actually reason—and where that reasoning can go wrong.
Alex: Thanks for listening to ResearchPod.