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

Despite substantial progress across downstream applications, vision-language models (VLMs) remain susceptible to language bias, often prioritizing linguistic cues over visual evidence and consequently producing incorrect predictions. Prior studies have proposed various approaches to understanding and mitigating language bias in VLMs, yet their findings often conflict due to the difficulty of tracing how language bias propagates within black-box VLMs. Building on the word completion task, we trace how language bias propagates through VLM inference by (1) proposing a diagnostic framework that decomposes the inference process into four distinct yet interdependent stages to trace the propagation of language bias; and (2) examining how two key factors underlying language bias, i.e., linguistic priors and cross-modal coverage, evolve across these stages and ultimately give rise to incorrect predictions. The linguistic prior captures the strength of statistical bias induced by the language model component of a VLM and represents the origin of language bias, whereas cross-modal coverage measures the extent to which linguistic cues cover the visual content. By decomposing inference into four stages and characterizing the interplay between linguistic priors and cross-modal coverage across these stages, we propose a systematic framework for tracing the propagation of language bias throughout the inference process; and uncover the underlying mechanism of language bias by revealing the interplay between linguistic priors and cross-modal coverage.

When you ask a vision-language model (VLM) a question with an image, it answers your question based on the image. But to act means to choose: saying one thing means choosing to respond not say another. This becomes an issue when the image contains multiple features that equally relate to the question at hand: how would VLMs choose to respond? This problem is precisely studied by the classical Stroop task in psychology: among a set of deliberately designed images whose text and color conflicts (e.g., \textcolor{blue}{RED}), participants are instructed to report the color. We adapt this psychological paradigm to a benchmark and prompted a range of contrastive and generative VLMs on their behavior. Across models, we observe a persistent bias toward the feature of text in Stroop images: when text and color disagree, models preferentially generate responses centered to the text rather than color. We then opened the black box and examined the source of this bias and why certain features influences the model's behavior more than others. In the embedding space, we found that textual cues are represented more strongly and separably than color cues. This representational asymmetry predicts intervention success: steering models toward text-related directions is substantially easier than steering them toward color-related directions. Using a classical psychological paradigm, we demonstrate, at both behavioral and representational levels, that distinct features within the same image can exert unequal influence on VLM behavior. Our findings suggest that VLMs assign different degrees of saliency to different features within the same image, highlighting the need to understand how cue-specific saliency shapes their performance in widely used downstream applications.

When a Vision-Language Model (VLM) sees a blue banana and answers "yellow", is the problem of perception or arbitration? We explore the question in ten VLMs with various sizes and reveal an Encoding-Grounding Dissociation: models that fail to report what they see (and thus provide a wrong answer) still encode the visual evidence as strongly as models that provide the correct answer. Using Multimodal Arbitration Crossover (MAC) analysis with layer-by-layer Logit Lens probing, we track the competition between visual and prior signals across every layer of each model. We show that visual attributes can be linearly decodable from early layers (AUC > 0.86). The accuracy remains nearly identical for both successful and failed samples. However, the gap in the final-layer logit - not the strength of encoding - better predicts grounding outcomes with a correlation of $ρ=$ 0.847. After having studied when VLMs base their answers on image clues rather than prior knowledge, we want to understand the causal relationships. We establish causality through full-sequence activation patching. The standard last-token interventions in LLM interpretability do not affect VLMs. In contrast, replacing the full token sequence at layers identified by MAC alters 60 to 84% of outputs. Partial-token decomposition shows that image tokens carry almost all of the causal impact, while text tokens have none. Scaling addresses the remaining architectural differences to achieve perfect retention. Moving from diagnosis to intervention, we show that training-free activation steering - both linear and sparse autoencoder-guided - in early layers can improve visual grounding by up to +3.8% with degrading performance in some setups. Overall, these findings lead to a clear conclusion: VLMs already see well, but the challenge is acting on what they see. Targeted interventions can help to bridge this gap.