Source-linked AI summary
Words or Vision: Do Vision-Language Models Have Blind Faith in Text?
Ailin Deng, Tri Cao, Zhirui Chen, Bryan Hooi
TL;DR
VLMs perform strongly on vision-centric tasks, but their behavior when visual and textual inputs conflict is underexplored. The paper evaluates controlled text variations across vision-centric settings, identifies a strong bias toward text, examines its drivers, and tests text-augmentation fine-tuning as mitigation. The findings indicate that text bias can substantially reduce robustness under corrupted text, while fine-tuning helps but does not eliminate the broader challenge.
Problem
VLMs’ handling of inconsistencies between visual and textual inputs is underexplored, despite the need for safe and reliable multimodal applications.
Method
The paper introduces matched, corrupted, and irrelevant text variations, evaluates VLM modality preferences and influencing factors, and explores supervised fine-tuning with text augmentation plus theoretical analysis.
Results
VLMs disproportionately trust text over visual evidence during inconsistencies, causing significant performance drops under corrupted text; text bias is shaped by prompts, model size, relevance, token order, and modality certainty.
Takeaways & Limitations
Text augmentation can reduce text bias, but scaling and prompt changes alone do not resolve the issue, leaving multimodal robustness challenging.
Takeaways & Limitations
The supervised-fine-tuning seed data comes from the VQAv2 validation split, separate from the benchmark evaluation data.
Abstract
from arXiv · showhide
Vision-Language Models (VLMs) excel in integrating visual and textual information for vision-centric tasks, but their handling of inconsistencies between modalities is underexplored. We investigate VLMs' modality preferences when faced with visual data and varied textual inputs in vision-centered settings. By introducing textual variations to four vision-centric tasks and evaluating ten Vision-Language Models (VLMs), we discover a \emph{``blind faith in text''} phenomenon: VLMs disproportionately trust textual data over visual data when inconsistencies arise, leading to significant performance drops under corrupted text and raising safety concerns. We analyze factors influencing this text bias, including instruction prompts, language model size, text relevance, token order, and the interplay between visual and textual certainty. While certain factors, such as scaling up the language model size, slightly mitigate text bias, others like token order can exacerbate it due to positional biases inherited from language models. To address this issue, we explore supervised fine-tuning with text augmentation and demonstrate its effectiveness in reducing text bias. Additionally, we provide a theoretical analysis suggesting that the blind faith in text phenomenon may stem from an imbalance of pure text and multi-modal data during training. Our findings highlight the need for balanced training and careful consideration of modality interactions in VLMs to enhance their robustness and reliability in handling multi-modal data inconsistencies.
1. Introduction
VLMs are increasingly used in complex multimodal settings where visual and textual inputs can conflict, yet their behavior under such inconsistencies is underexplored. This study finds that VLMs often place disproportionate trust in text and examines the factors behind this bias and ways to reduce it.
- Motivation: VLMs’ handling of inconsistent visual and textual inputs remains underexplored despite their use in context-rich multimodal applications.Understanding these tendencies is presented as important for safe and reliable real-world deployment.
- Research Questions: The study investigates modality preference, robustness to corrupted text, and factors influencing VLM behavior under cross-modal inconsistencies.The authors construct a comprehensive benchmark to address these questions.
- Core Finding: VLMs exhibit “blind faith in text,” overly trusting textual data even when it contradicts visual evidence.This bias causes significant performance degradation when text is corrupted and raises potential safety concerns.
- Influencing Factors: Instruction prompts modestly adjust text bias, larger language models slightly mitigate it with saturation, and relevant text increases textual influence.These effects are examined alongside token order and uni-modal certainty.
- Influencing Factors: Putting text tokens before image tokens exacerbates text bias, possibly because of positional biases inherited from language models.The interplay between visual and textual certainty also influences modality preference.
- Mitigation and Explanation: Supervised fine-tuning with text augmentation reduces text bias, while theoretical analysis suggests training imbalance between pure-text and multimodal data may contribute to the phenomenon.The proposed explanation reflects VLMs’ reliance on language models primarily trained on textual data.
2. Preliminaries
The paper formalizes multimodal inputs, controlled text variations, model behavior under modality disagreement, and metrics for quantifying text preference and performance changes. Its setup distinguishes matching, corrupted, and irrelevant text while evaluating how combined inputs compare with unimodal answers.
- Multimodal Setup: Each sample contains an image I, textual information T, and question Q, enabling answers from image-only, text-only, or combined inputs.The model is parameterized by θ and each sample has a corresponding ground-truth answer Y.
- Text Variations: The benchmark constructs matching, corrupted, and irrelevant text variants to test helpful information, misleading information, and distraction handling.Matching text also prevents models from succeeding by simply rejecting all text.
- Behavior Categories: For inconsistent cases, behavior is classified as following the image answer, following the text answer, or producing another answer.Empirical analysis retains cases where image-only and text-only answers differ under exact match.
- Metrics: Text Preference Ratio (TPR) quantifies how often a model chooses text over visual information when the modalities disagree.A higher TPR indicates stronger text bias.
- Metrics: Macro Accuracy averages model accuracy across problem sets with matching, corrupted, and irrelevant text variations.Normalized Accuracy measures performance under a text variation relative to accuracy on the base problem set.
- Text Construction: Figure 2’s prompt generates matched and corrupted text from an image, question, and ground-truth answer, with the question and answer inserted for each sample.The generated descriptions are extracted as the matched and corrupted texts.
3. Empirical Analysis
Across four vision-centric domains, the empirical analysis evaluates how VLMs behave with matched, corrupted, and irrelevant text, finding a strong tendency to follow relevant text even when it conflicts with visual evidence. This bias substantially harms performance under corruption and varies with model type, scale, prompting, relevance, and certainty.
- Setup: The benchmark evaluates VLMs across general VQA, document VQA, math reasoning, and brand recognition using matched, corrupted, and irrelevant text.Matched text tests useful information, corrupted text tests misleading information, and irrelevant text tests distraction resistance.
- Blind Faith in Text: Relevant text attracts strong text preference even when incorrect, with high preference ratios across models and particularly strong bias among open models.Claude Haiku reaches 87% and 83% text preference under match and corruption in VQAv2, respectively; Qwen2-VL-7B reaches 29% under corruption versus 13% under match.
- Performance Impact: Corrupted text sharply reduces performance: Qwen2-VL-7B falls to 59%, 63%, and 52% of its original accuracy on VQAv2, DocVQA, and MathVista.Proprietary models decline less overall, but efficient proprietary variants also show significant drops.
- Performance Impact: In brand recognition, corrupted HTML reduces most open models’ accuracy by nearly 50% compared with original performance, while proprietary models show slight resilience.The task combines HTML strings with webpage screenshots, making misleading injected text relevant to practical safety concerns.
4. Investigated Solutions
The paper evaluates prompting and supervised fine-tuning with text augmentation as ways to reduce VLM text bias. Fine-tuning improves robustness to corrupted text, generalizes across datasets, and benefits from sufficient data while retaining useful language behavior.
- Supervised Finetuning (SFT): Instruction prompting improves Macro accuracy by only 1–2%, whereas supervised fine-tuning provides stronger gains under text corruption.The comparison uses an instruction baseline and fine-tuned models.
- Supervised Finetuning (SFT): Corruption accuracy increases from 28.69% to 71.25% after supervised fine-tuning while overall Macro accuracy is maintained.This is reported for in-distribution evaluation.
- Generalization: Fine-tuned models improve across DocVQA, MathVista, and Brand Recognition, but gains are smallest on MathVista under greater distribution shift.The evaluation compares performance beyond VQAv2.
- Effect of Text-Only Data: Fine-tuning reduces text bias and creates a match–corruption performance gap of up to 40%.The ablation keeps total training-data volume constant across experiments.
- Effect of Data Volume: As SFT data increases, reliance on text in corruption cases decreases from 58% to 25% while remaining relatively steady in match cases.Text-only data also helps preserve core language capabilities; removing it can cause indiscriminate rejection of text.
5. Theoretical Analysis
The theoretical analysis models training with pure-text and multi-modal samples and relates their expected losses to the balance between the two data sources. Under stated assumptions, an imbalance favoring pure-text data can make models favor text over images.
- Training-data formulation: The analysis considers N pure-text samples and M multi-modal samples drawn independently from Dtxt and Dmul.The learned parameters are obtained through empirical risk minimization.
- Theoretical result: The expected pure-text and multi-modal losses are weighted by M/(N+M) and N/(N+M), respectively, for the cross-modal error term.The theorem also involves approximation errors and a covering-number quantity under bounded-loss assumptions.
- Interpretation: When N ≫ M, the analysis indicates larger expected multi-modal loss and smaller pure-text loss, potentially producing a preference for text over images.The paper connects this imbalance to reliance on pretrained language models.
6. Related Work
Prior VLM benchmarks mainly evaluate vision-centric capabilities, while robustness research studies perturbations within language or vision. This work shifts attention to textual variations as a source of cross-modal inconsistency.
- Evaluation on VLMs: Existing VLM benchmarks are primarily vision-centric and usually treat text only as question input without additional context.This limits evaluation of how models handle inconsistent visual and textual information.
- Benchmarks with Input Perturbation: Text perturbation research evaluates robustness to distractions or misleading context in natural-language tasks, while computer-vision work commonly perturbs image inputs.The paper instead examines text variations in VLMs that perform vision-centric tasks.
7. Conclusion and Discussion
The paper concludes that VLMs often rely on text over visual input when modalities conflict, and that substantial robustness challenges remain. Its theoretical and implementation analyses formalize VLM processing and training while motivating further work on reliable cross-modal interaction.
- Conclusion and Discussion: VLMs’ blind faith in text causes performance drops and potential safety risks when visual and textual inputs are inconsistent.The conclusion identifies instructions, model size, text relevance, token order, and modality certainty as influencing factors.
- Conclusion and Discussion: Scaling model size and prompt changes alone do not resolve text bias, whereas supervised fine-tuning with text augmentation helps reduce it.The conclusion frames the remaining challenge as balancing robustness and effectiveness in cross-modal settings.
- Details of Theoretical Analysis: The theoretical appendix represents a VLM as a function mapping sequences of d-dimensional text or image features of length τ to an output space.For simplicity, the output space is set to R.
- Structure: The transformer formulation uses layer-wise blocks, normalization, nonlinear activation, softmax attention, and a scalar output derived from the final-layer representation.The appendix states that [G]τ denotes the τ-th row of a matrix G.
A.3. Results
Theorem A.5 states a high-probability bound for VLMs that separates approximation, cross-modal, and generalization errors. The bound includes a covering-number-related quantity and suppresses global constants and logarithmic factors.
- With probability at least 1 −δ, Theorem A.5 provides a bound for every VLM parameter θ ∈Θ.The parameter set satisfies Assumptions A.1–A.4, and the model has L layers.
- The bound decomposes error into approximation, cross-modal, and generalization components.
- Cvlm is related to the covering number of the VLM hypothesis class.
- The notation ≲ hides global constants and logarithmic factors involving quantities other than N, M, and τ.
A.4. Proof of Theorem A.5
The appendix develops Theorem A.5 using covering-number-based lemmas and intermediate high-probability inequalities. It also documents experimental text variations and the phishing-webpage brand-recognition setup.
- A.4. Proof of Theorem A.5: The proof uses lemmas for bounded function classes, Lipschitz losses, and covering-number estimates.Lemma A.6 invokes prior results for real-valued function classes, while Lemma A.7 concerns VLM parameter classes.
- A.4. Proof of Theorem A.5: The argument combines equations 5, 8, and 9, then applies the empirical-risk-minimizer definition to obtain the theorem’s bound.
- Experimental setup: The experimental setup illustrates matching, corrupted, and irrelevant textual variations across datasets.
- Brand recognition: Brand-recognition experiments use TR-OP phishing webpages containing screenshots and HTML that may match or conflict with the displayed brand.HTML is stripped of tags and truncated to 5,000 characters.
B.3. Evaluation
The evaluation follows dataset-specific protocols and adds formatting prompts for some datasets to reduce open-ended responses. MathVista instead uses direct GPT-based output evaluation.
- Dataset-specific evaluation protocols use additional formatting prompts for certain datasets to reduce open-ended answers.
- MathVista omits formatting prompts and evaluates outputs directly with GPT.
C. Experimental Results
The experimental results compare textual matching, corruption, and irrelevance across four datasets using accuracy, normalized accuracy, TPR, and macro accuracy. Additional comparisons cover base models, prompts, and SFT for LLaVA-NeXT-7B and Qwen2-VL-7B.
- Accuracy, normalized accuracy, Text Preference Ratio, and macro accuracy quantify performance across matching, corruption, and irrelevance conditions.Macro accuracy averages the three condition-specific accuracies for comparison with base accuracy.
- The evaluation includes DocVQA, MathVista, and Brand Recognition examples alongside the broader four-dataset results.
- Results for base models, instructional prompts, and supervised fine-tuning are reported across four datasets.