Project page

Seeing Red, Thinking Bad:
Color Bias in Vision Language Models

Kohsuke Ide1,2, Ryousuke Yamada1,3, Yoshihiro Fukuhara1, Hirokatsu Kataoka1,4, Yutaka Satoh1,2

1AIST  ·  2University of Tsukuba  ·  3FunAI Lab, UTN  ·  4University of Oxford

ide.agi@aist.go.jp

Concept: subtle visual styling biases VLM outputs
Subtle visual styling biases VLM outputs. Identical text content produces different sentiment classifications when positive words are colored green — VLMs can treat ordinary formatting as an effective stealth visual prompt, even though humans regard it as non-instructive decoration.

Abstract

Vision language models (VLMs) are increasingly used in industrial decision-making systems. We study how VLMs interpret text rendered as an image and investigate the influence of visual styling biases. We introduce Stealth Visual Prompts, which subtly change the visual styling of text — such as color and contrast — while preserving semantic content, and use them to systematically measure their impact on VLM analysis. We observe that coloring positive words green consistently shifts sentiment predictions toward positive, so that VLMs often fail to properly account for negative words; this behavior correlates with changes in the latent representations of the vision encoder induced by color variation. We further show that reducing text–background contrast increases reliance on visually salient cues and leads to more incorrect Visual Question Answering (VQA) outputs. Ordinary visual styling of rendered text can guide VLMs' interpretation in ways that diverge from human semantic understanding.

Stealth Visual Prompts

A Stealth Visual Prompt is a controlled perturbation of the visual rendering of text while the underlying string is held fixed. We focus on two complementary, ubiquitous attributes:

Color

Recolors a subset of words using six canonical hues (red, green, blue, yellow, cyan, magenta) at three intensity levels. Color can encode learned semantic associations.

Contrast

Reduces text–background contrast (low-contrast grayscale), optionally with one high-contrast salient span. Contrast directly modulates perceptual accessibility.
Examples of generated stimuli
Generated stimuli. (a) Mixed-sentiment text (Color axis). (b) Structurally separated long text. (c) Saliency Competition (Contrast axis): an incorrect decoy is made visually salient.

Sweep the hue, watch the meaning move

The word string never changes — only its hue. Yet, a Vision encoder (CLIP) places the recolored word at a different location in concept space. Drag the slider to sweep the hue and watch how strongly the word reads as its own meaning (e.g. how warm "warm" looks), live, from real measured data.

Rendered stimulus
warm
Selected word
Pure color (no word)

Key findings

1 · Color prompts induce systematic sentiment biases

Recoloring sentiment-bearing words shifts predicted polarity relative to an all-black baseline. Susceptibility differs sharply across models.

ModelMax Pos. ↑Max Neg. ↓Range
IDEFICS2-8B+0.160−0.3600.520
LLaVA-Mistral-7B+0.030−0.0100.040
LLaVA-Vicuna-7B+0.060−0.0600.120
Qwen2-VL-7B+0.420−0.4800.900
Short-sentence sentiment bias
Short-sentence Sentiment Set: average sentiment bias vs. baseline for four VLMs.

2 · Hue shifts vision-encoder semantic projections

Sweeping a rendered word's hue moves its CLIP image embedding along human-interpretable semantic axes (valence, emotion) — a diagnostic, representation-level correlate of the behavioral bias (evidence consistent with it, not proof of a causal mechanism inside every VLM).

CLIP hue tuning curves
CLIP representation probe: hue-dependent semantic projections of six probe words on four axes.

3 · Contrast prompts increase saliency-driven VQA errors

As the non-salient context becomes harder to read, several models increasingly copy a visually salient but incorrect decoy. We report the Induced Error Rate (fraction of predictions containing the decoy) — not a general VQA accuracy measure; it isolates decoy copying in the Decoy-Salient condition.

Modelg=1g=16g=64g=128g=192g=240
IDEFICS2-8B24%23%24%27%32%36%
LLaVA-Mistral-7B27%25%26%24%24%27%
LLaVA-Vicuna-7B19%19%20%20%22%25%
Qwen2-VL-7B4%4%4%4%5%6%
IER vs contrast
Induced Error Rate vs. grayscale level of the non-salient context (higher ⇒ lower contrast).
Readability transition
A model-dependent readability transition from a minimal VLM-based OCR proxy.

Why it matters

These sensitivities imply a reliability and safety risk for VLM pipelines that ingest documents or UI screenshots: benign or adversarial styling can steer model decisions without changing the underlying text. Practical safeguards include normalizing rendered text before inference, cross-checking image-based answers with OCR-extracted text, and adding style-invariance checks to evaluation suites.

Beyond the paper

Recoloring a word shifts its valence projection in CLIP
The same token "bad" in black vs. green text is encoded with a different valence projection in this diagnostic probe. Across the six canonical hues, the projection changes even though the string is unchanged. CLIP-ViT-L/14-336 measurements for the interactive companion.

Recoloring a word, with the string fixed, still shifts where a vision encoder places it. A companion note reads this through the representational-convergence debate (the Platonic and Umwelt hypotheses), as an interpretation rather than a paper claim. Read the blog →

BibTeX

@inproceedings{ide2026seeingred,
  title     = {Seeing Red, Thinking Bad: Color Bias in Vision Language Models},
  author    = {Ide, Kohsuke and Yamada, Ryousuke and Fukuhara, Yoshihiro
               and Kataoka, Hirokatsu and Satoh, Yutaka},
  year         = {2026},
  howpublished = {Manuscript},
  url          = {https://github.com/KohsukeIde/color-bias-vlm}
}