💬 NLP

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

This paper introduces SIGNPOST-Bench, a controlled counterfactual benchmark comprising 25,555 image variants to evaluate how multimodal large language models resolve conflicts between visual and textual cues, revealing that adversarial text interventions significantly degrade geolocation accuracy and expose distinct arbitration behaviors across 20 evaluated models.

Original authors: Sirun Li, Minghao Liu, Ling Dai, Yong Li, Haoxin Lyu, Junting Zhou, Fan Zhang

Published 2026-08-06

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Sirun Li, Minghao Liu, Ling Dai, Yong Li, Haoxin Lyu, Junting Zhou, Fan Zhang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery: "Where was this photo taken?" You have two clues to help you. The first is the visual scene: the architecture, the trees, the cars, and the way the light hits the buildings. The second is the text written on signs, storefronts, or billboards in the picture. Usually, these clues agree. A sign saying "Paris" sits on a building that looks like it's in Paris. But what happens when they disagree? What if the photo shows a snowy, wooden cabin in the mountains, but a bright neon sign in the window screams "Welcome to the Sahara Desert"?

This is the world of Multimodal Large Language Models (MLLMs). These are super-smart AI computers that can "see" images and "read" text at the same time. Scientists have been teaching them to combine these clues to make guesses about the world. But there's a big question: When the visual clues and the text clues fight each other, which one does the AI trust? Does it ignore the confusing sign and look at the snow? Or does it blindly believe the sign and forget the snow? Until now, we didn't have a good way to test this specific kind of confusion.

Enter SIGNPOST-Bench, a new experiment designed by researchers to play a game of "trick the AI." They took thousands of real-world photos and created a special set of "counterfactual" versions for each one. Think of it like a photo editing workshop where they take one original picture and make four new copies:

The Blank: They erased the text completely.
The Similar: They swapped the sign for a different one that still makes sense for the location (e.g., changing "Welcome to Michigan" to "Welcome to Detroit").
The Random: They put up a sign that has nothing to do with the place (e.g., "Welcome to the Moon").
The Adversarial: This is the tricky one. They replaced the sign with a lie that points to a specific, different location (e.g., changing "Welcome to Michigan" to "Welcome to New York").

The researchers then asked 20 different AI models from seven major tech companies to guess the location for all five versions of every photo. They wanted to see if the AI could spot the lie or if it would get tricked by the new sign.

The results were quite revealing. When the AI looked at the original photos, it was pretty good at guessing the location. But when they introduced the Adversarial lies, the models got confused. On average, the AI's guess became 4.8 times less accurate. Instead of being off by about 282 kilometers, the AI was now off by 1,347 kilometers.

Even more interesting, the AI didn't just get worse; it got misled. When the sign said "New York," the AI didn't just guess randomly; it actually moved its guess closer to New York. In fact, for every single model tested, the presence of the conflicting text pulled the answer toward the lie. About 6.5% to 20.1% of the time, the AI's guess ended up less than 50 kilometers away from the fake location the sign was pointing to.

The study also found that being "smart" at reading signs or seeing pictures didn't guarantee being good at spotting a conflict. Some models that were excellent at guessing locations with clean, honest photos were actually worse at ignoring the lies than some smaller models. It turns out that being able to read a sign doesn't mean you know when to ignore it if it contradicts the rest of the picture.

Finally, the researchers tried to "coach" the AI, asking it specifically to look out for conflicts. While this helped the AI realize something was wrong about 49% of the time, it didn't actually help it guess the correct location any better. The AI knew it was confused, but it still couldn't figure out how to solve the puzzle.

In short, this paper shows that current AI models are still very easily tricked when text and images disagree. They tend to trust the text too much, even when it's a blatant lie. The researchers built this new "SIGNPOST-Bench" test to help future AI developers understand exactly where their models are failing, so they can teach them to be better detectives who know when to trust their eyes and when to question the signs.

Technical Summary: SIGNPOST-Bench

Problem Statement

Multimodal Large Language Models (MLLMs) increasingly rely on both visual context and scene text to make grounded predictions in real-world scenarios. While existing benchmarks evaluate OCR, visual recognition, and general multimodal capabilities, they rarely isolate how models arbitrate between visual and textual evidence when these sources conflict. In natural scenes, architectural or terrain cues may suggest one location while signs or addresses suggest another. Current evaluation methods often report only final discrete outcomes (e.g., correct/incorrect) without revealing which evidence source drove the decision or measuring the magnitude and direction of prediction shifts in a shared output space. Consequently, strong performance on aligned inputs does not guarantee a model's ability to reliably choose between competing visual and textual cues.

Methodology

The authors introduce SIGNPOST-Bench (Scene Image Geo-localization with Noisy Perturbation on Observed Sign Text), a controlled counterfactual benchmark designed to evaluate text–vision conflict resolution through visual geolocation.

Benchmark Construction

The core unit of the benchmark is a counterfactual quintuplet derived from a single source image. Each quintuplet consists of five variants generated via synthetic, localized scene-text interventions designed to preserve surrounding non-textual content:

Original: The unmodified source image with native scene text.
Blank: Selected scene-text spans are removed, serving as a text-ablated baseline.
Similar: Text is replaced with alternatives compatible with the ground-truth geographic context or language (e.g., "Welkom to Flanders" instead of "Welkom to Holland").
Random: Text is replaced with unrelated, readable text without a designated geographic target.
Adversarial: Text is replaced with a geographically conflicting cue. If the text is geocodable, it defines an injected target ( $y_{trap}$ ) distinct from the ground truth.

The benchmark comprises 5,111 counterfactual groups and 25,555 image variants drawn from four datasets: IM2GPS3K, YFCC4K, Google Street View, and Baidu Street View. Scene text is categorized into a taxonomy based on geographic identifiability: T1 (Portable), T2 (Cultural), and T3 (Geo-Specific).

Evaluation Protocol

The study evaluates 20 MLLMs from seven providers (Google, OpenAI, Anthropic, Alibaba, Moonshot, ByteDance, xAI) across all five conditions. The evaluation employs three primary diagnostic metrics:

Weighted Localization Accuracy (WLA): Measures prediction quality relative to the ground truth using exponential decay based on geodesic distance.
Text Bias Score (TBS): Measures the paired change in localization error caused by edited text relative to the Blank baseline ( $TBS = D_{edited} - D_{blank}$ ). A positive TBS indicates the text degraded performance.
Trap-Fit Rate (TFR): Measures the fraction of Adversarial predictions falling within a 50 km radius of the injected target, indicating if the model was "trapped" by the conflicting text.

Additionally, a Multimodal Conflict Robustness Score (MCRS) is proposed, decomposing performance into a Capability Score (C) (performance on Original/Blank) and a Conflict Robustness Score (R) (performance retention under Random/Adversarial text, penalized by error increase and trap-following).

Key Results

Performance Degradation: Conflicting scene text substantially degrades localization. Across all models, the median localization error increased 4.8-fold (from 282 km to 1,347 km) when comparing Original images to Adversarial variants. Mean WLA dropped by 36.6%.
Directed Shifts: The conflict does not merely increase random error; it induces directed shifts. For geocodable adversarial samples, 6.5% to 20.1% of predictions across models fell within 50 km of the injected target. Every evaluated model exhibited a positive mean paired Trap Distance Reduction (TDR), confirming that predictions shifted toward the injected geographic target rather than just becoming less accurate.
Distinct Effects of Text Types:
- Similar replacements (compatible text) generally reduced error compared to the Blank baseline, showing models utilize semantic alignment.
- Random and Adversarial replacements increased error significantly.
- T3 (Geo-Specific) text provided the highest performance on aligned inputs but also induced the largest degradation (42.2% proportional drop) under conflict, suggesting a trade-off between utility and vulnerability.
Capability vs. Robustness: Clean-input localization performance (Capability Score) does not fully predict robustness to conflicting text. Models with high WLA on Original images did not necessarily maintain high WLA under Adversarial conditions. For instance, Qwen3-VL-30B showed higher robustness than Seed-2.0-Pro despite lower original capability.
Conflict Awareness: Targeted probing experiments with Gemini-2.5-Flash and GPT-4o-mini revealed that explicit conflict detection accuracy (CDA) does not reliably correlate with improved localization performance. Even when models identified the conflict, they often still followed the injected text.

Significance and Claims

The paper claims three primary contributions:

Formulation: It formulates text–vision conflict resolution in a continuous coordinate space, enabling the measurement of both localization degradation and directed shifts toward injected geographic targets, a capability lacking in discrete-task benchmarks.
Benchmark Construction: It provides SIGNPOST-Bench, a large-scale, controlled framework with 25,555 image variants that allows for within-scene comparisons of native text, text removal, and compatible, unrelated, and conflicting replacements.
Empirical Findings: It establishes that conflicting scene text is a systematic failure mode for current MLLMs. The results demonstrate that models are susceptible to being "trapped" by conflicting text, shifting predictions toward injected targets. Crucially, the paper argues that clean-input localization performance is not a sufficient proxy for conflict robustness, highlighting a gap in current model alignment where explicit place names in text can override less direct visual evidence.

The authors position SIGNPOST-Bench as a reproducible framework for diagnosing how MLLMs arbitrate between conflicting multimodal evidence, moving beyond aggregate accuracy to analyze specific failure modes in scene-text interpretation.