SIGNPOST-Bench: Benchmarking Text-Vision Conflict Resolution in Multimodal Large Language Models

Analysis

Why This Paper Matters

Multimodal large language models (MLLMs) are increasingly deployed in real-world scenarios where visual and textual information must be integrated. However, existing benchmarks often treat these modalities as complementary, rarely probing how models arbitrate when they directly conflict. SIGNPOST-Bench addresses this gap by introducing a controlled counterfactual benchmark that systematically manipulates scene text to create conflicts with visual content. This is crucial because real-world scenes often contain misleading or adversarial text (e.g., street signs, billboards) that can derail model predictions.

The paper's focus on visual geolocation is particularly insightful. Geolocation provides a continuous, measurable output (distance error) that can be precisely tracked under different text interventions. This allows for fine-grained analysis of how conflicting text shifts model predictions toward injected targets. The finding that adversarial text increases median error by 4.8-fold underscores the severity of this vulnerability, which has direct implications for safety-critical applications like autonomous navigation and augmented reality.

Technical Contributions

• Counterfactual quintuplet design: Each image is transformed into Original, Blank, Similar, Random, and Adversarial variants, enabling paired comparisons that isolate the effect of text content.
• Localized text interventions: Synthetic text is inserted while preserving non-textual content, ensuring that any performance change is attributable to the text-vision conflict.
• Geolocation as a diagnostic: The use of geographic distance as a continuous metric provides a sensitive measure of model arbitration between visual and textual cues.
• Comprehensive evaluation: 20 MLLMs from seven providers are assessed, offering a broad view of current model behavior.

Results

The results are striking. Adversarial text increases median localization error from 282 km to 1,347 km, a 4.8-fold degradation. Across models, 6.5-20.1% of geocodable adversarial predictions land within 50 km of the injected target, indicating that models often follow the misleading text. Every evaluated model exhibits a positive mean paired reduction in target distance from Blank to Adversarial, confirming that conflicting text consistently pulls predictions toward the injected location. Importantly, clean-input performance does not predict robustness to conflict, suggesting that this is a distinct failure mode requiring dedicated evaluation.

Significance

SIGNPOST-Bench establishes a new paradigm for evaluating multimodal conflict resolution. By providing a controlled, scalable benchmark, it enables researchers to systematically study how MLLMs handle contradictory evidence. This has immediate implications for improving model robustness and safety, as well as for understanding the underlying mechanisms of multimodal reasoning. The benchmark's design can be adapted to other tasks beyond geolocation, potentially becoming a standard tool for assessing multimodal trustworthiness. As MLLMs become more integrated into daily life, benchmarks like this are essential for identifying and mitigating vulnerabilities before they cause real-world harm.