Multimodal models, meaning systems that read an image and its accompanying text together, are mostly trained on image–text pairs scraped from platforms like Reddit and from web crawls. These sources skew young, male, Western, and wealthy. In that data, the caption usually restates what the photo already shows clearly: a well-lit suburban kitchen labeled "my new kitchen." This proposal argues that the skew teaches models a habit, not just a demographic imbalance. Cross-attention, the layer in which one stream decides how much to draw on the other, comes to treat agreement between image and text as a sign of relevance. In homes and settings the platforms rarely see, the image often fails on specific visual cues. Texture differs, lighting is dark, objects are occluded. The text and image then agree less, and the attention layer quietly discounts whichever stream disagrees. That is often the stream carrying the information the other lacks.
We propose to measure this directly and then to fix it with a narrow change. First, we estimate how much information image and text share, separately for each income and region group, and test whether that shared fraction falls in exactly the groups where models fail. Second, we replace agreement-driven weighting with a gate driven by visual reliability. A small head predicts, from the image alone, whether its texture and lighting cues have degraded, and this head is trained with explicit degradation labels. The fusion layer then sets the weight between its visual and text paths from that prediction rather than from image–text similarity. If the measurement holds and the gate closes more of the income gap than an equally sized learned gate, the field gains a concrete account of part of the "unexplained" bias that arises inside fusion layers. It also gains a fix that does not require collecting balanced data.
Core Claim
The claim has three linked parts, and each can be falsified. First, image–text redundancy is not uniform across demographic groups. The fraction of task information that image and text share is higher in high-income and Western data than in low-income and non-Western data, even after correcting for groups that differ in how often text is present at all. Second, similarity-based cross-attention trained on platform data turns this difference into error. On images whose visual cues have failed, attention mass on visual tokens drops, and mistakes concentrate on examples where one modality held unique information. Third, conditioning the fusion weight on a separately supervised estimate of visual reliability removes a measurable part of the resulting gap. A learned gate of the same size, with no reliability supervision, does not.
The evidence motivating each part comes from separate literatures that have not been joined. Exploring Why Object Recognition Performance Degrades Across Income Levels and Geographies with Factor Annotations annotates 16 visual factors on Dollar Street. It finds that texture is five times more likely to be a distinguishing factor in low-income homes and 1.7 times more likely to be present in a mistake, with texture, occlusion, and darker lighting shared across architectures. That work is purely visual and image-only. We ask what these primitive-level failures do once a text stream is present. FairSSL shows that the "redundancy assumption" behind multimodal self-supervised learning, the premise that views share most task information, breaks in heterogeneous data, and that enforcing alignment amplifies demographic bias. It addresses this with subject-aware regularization of encoders, not with the attention layer. ICYM²I shows that when a modality's presence is selected non-randomly, naive estimates misattribute shared information as unique. It reports chest X-ray unique information falling from about 5% to 1.8% under inverse-probability weighting, a method that reweights examples by how likely their observed modality pattern was. This finding constrains our measurement: any per-group redundancy estimate must be corrected this way, or the first claim could be an artifact of missingness. Finally, Person-Centric Annotations of LAION-400M finds that linear co-occurrence statistics explain 60–70% of gender bias in downstream models and leaves the rest unexplained. Our mechanism is one candidate for part of that remainder.
Research Program
The work has a measurement stage and an intervention stage. The intervention is only worth building if the measurement comes out as predicted.
The measurement stage needs paired image–text data with group labels and visual-factor labels. No existing resource in our material has all three. Dollar Street (The Dollar Street Dataset) supplies images across 63 countries and monthly incomes from about \$27 to \$19,671, and the factor-annotation study supplies primitive labels on 14,000 of those images. It has no text, so we add two text channels for the factor-annotated subset. The first is "platform-style" captions from a captioner trained on web image–text data. These reproduce the redundancy habit by construction. The second is short functional notes written by annotators recruited from the same regions as the households, describing what the object is and how it is used, as a resident might when posting. The contrast between these channels is deliberate. It lets us separate "text happens to be redundant on this platform" from "text carries what the image cannot." We acknowledge that annotator-written notes carry their own biases, and we will report inter-annotator variation per region.
A second testbed uses data that is already multimodal and platform-sourced. This is CrisisMMD, Twitter disaster posts, together with the Reddit sentiment set, both used by Robustness of Fusion-based Multimodal Classifiers to Cross-Modal Content Dilutions. That paper shows topically aligned "dilution" text drops CrisisMMD F1 from 0.734 to 0.564, but it never asks who is harmed. Where per-post event or location metadata permits, we split by region and ask whether dilution damage is asymmetric. The paper's dilution generator lets us inject redundant text in a controlled way.
For each group we estimate the shared and unique task information of image and text with an information-decomposition estimator. Following ICYM²I, we apply inverse-probability weighting for text presence. We then train a standard bidirectional cross-attention fusion model on a web-curated Reddit image–text corpus in the style of RedCaps (RedCaps), with a classification head mapped to Dollar Street categories, and apply it to Dollar Street without in-domain fine-tuning. This is the regime the bias literature worries about: models built on platform data and deployed elsewhere. We record per-example attention mass on visual tokens and relate it to the annotated factors.
The intervention stage changes only the fusion weighting. Let $h_v$ be the text-conditioned visual context, meaning cross-attention with text queries over image keys and values, and let $h_t$ be the text-path representation. The fused output is
$$h = g\,h_v + (1-g)\,h_t, \qquad g = \sigma\big(a\,\hat r(x) + b\big),$$
where $\hat r(x)$ is a visual-reliability score computed from the image $x$ alone, $\sigma$ is the logistic function, and $a, b$ are learned scalars. The departure from the gated cross-attention of An Interpretable Framework for Drug-Target Interaction with Gated Cross Attention lies in where the gate's signal comes from. That method learns its gate from the fused features themselves, so the gate can inherit the agreement habit. Our $\hat r$ is a small head trained with a separate loss to predict primitive degradation. The labels come from synthetic degradations applied to well-lit, high-income training images, namely darkening, blur that removes texture, and occlusion patches, and, in a second variant, from the real factor annotations on a held-out set of countries. The gate therefore has no direct path to the cross-modal similarity score.
Baselines
The naive reference point is the image-only zero-shot CLIP result from the factor-annotation study, which reports 93% top-5 accuracy for high-income and 62% for low-income homes, a 31-point gap. We rerun it on our exact split so that every other number is comparable to it. Alongside it we run simple late fusion, which averages the predictions of independent image and text heads. This shows whether any attention-based fusion beats not fusing deeply at all.
The strongest existing method on this question is learned gated cross-attention in the style of the drug–target interaction model. It is the most direct competitor because it already lets the network down-weight an unreliable stream, which is exactly what we claim is needed. We implement it on the identical backbone with the same parameter count as our gate. It gets the same training corpus, the same synthetic degradation augmentations, and the same hyperparameter search budget, which is a fixed number of trials over learning rate and gate initialization. We cannot cite a reported number for it on Dollar Street or CrisisMMD; no reported result was found on Lacuna for gated cross-attention on these benchmarks. We therefore treat our own tuned run as the bar.
The key ablation isolates the reliability supervision. It uses the same gate architecture with the auxiliary degradation loss removed, so $\hat r$ is learned end to end. If this ablation matches the full method, the proposal's central ingredient does nothing.
We also run three established remedies. The first is shared/unique subspace disentanglement with orthogonality constraints (Disentangling Complementary and Redundant Information in Multimodal Fusion). The second is a texture-debiased visual backbone, which the factor-annotation study reports improves low-income accuracy by 4.1 points. The third is modality dropout. As an upper reference rather than a competitor, we report an image model fine-tuned on Dollar Street itself, which The Dollar Street Dataset reports averages roughly 75% consistently across income levels.
Beating the strongest baseline means three things across five seeds. Low-income top-5 accuracy and worst-group accuracy over income-by-region cells must exceed the tuned learned gate, with confidence intervals that do not overlap. The gain over image-only CLIP must clearly exceed the 4.1 points that texture debiasing delivers. And high-income accuracy must not fall by more than one point. On CrisisMMD, the target is a smaller dilution-induced F1 drop than the reported 23.3%, together with a smaller difference in that drop between regions.
Why not just…?
Why not just balance the training data across demographics? Balancing addresses who appears in the data, while our mechanism concerns how image and text relate within each example. The factor-annotation study rules out class imbalance as the sole driver of the income gap. Racial and Gender Stereotypes Encoded Into CLIP Representations finds strong trait disparities in CLIP even when evaluated on the race- and gender-balanced FairFace set. We nonetheless run group-reweighted training as a baseline. If it matches the reliability gate, the redundancy story is unnecessary and we will say so.
Why not just fine-tune on in-distribution data like Dollar Street? This is a real threat. The roughly 75% income-consistent accuracy of a fine-tuned model shows that in-domain labels largely solve the image-only problem. Our contribution is scoped to the setting where labeled data from underrepresented groups is unavailable at training time, which is the default for platform-trained models. We report fine-tuning as an upper reference. We also run a few-shot variant to measure how many in-domain examples are needed before the gate stops mattering.
Why not just augment training images with darkening and blur? The gate's supervision uses these same degradations, so the gains could come from augmentation alone. Every baseline, including the learned gate, therefore receives identical augmentations. In addition, an augmentation-only model with no gate is run. Only the gap between these and the full method counts as evidence for reliability gating.
Why not just use modality dropout or a learned gate? Modality dropout forces use of each stream during training but does not change how weights are set when both streams are present. ICYM²I warns that non-random missingness distorts what models learn about informativeness. A learned gate is the strongest baseline and is handled above. The ablation without reliability supervision is the experiment that decides this question.
Decisive Checks
The first check comes before any intervention. If the corrected shared-information fraction does not differ between high- and low-income groups, or differs only before inverse-probability correction, the core mechanism is wrong and the intervention stage is dropped. The second check is a correlation. Visual attention mass should fall on images with annotated texture, lighting, or occlusion failures more steeply in low-income than in high-income homes, and errors should concentrate on examples where the functional notes carried information the image did not. The third check is a dose-response test. Applying synthetic darkening to high-income images should reproduce the attention drop, and the reliability gate should keep that drop from turning into errors. Finally, on CrisisMMD, adding dilution text in a controlled way should harm the groups with lower baseline redundancy more. If it does not, the platform-asymmetry part of the claim fails even if the Dollar Street results hold.
References
- Exploring Why Object Recognition Performance Degrades Across Income Levels and Geographies with Factor Annotations
- FairSSL: Fair Multimodal Self-Supervised Learning
- ICYM$^2$I: The illusion of multimodal informativeness under missingness
- Person-Centric Annotations of LAION-400M: Auditing Bias and Its Transfer to Models
- The Dollar Street Dataset: Images Representing the Geographic and Socioeconomic Diversity of the World
- Robustness of Fusion-based Multimodal Classifiers to Cross-Modal Content Dilutions
- RedCaps: Web-curated image-text data created by the people, for the people
- An Interpretable Framework for Drug-Target Interaction with Gated Cross Attention
- Integrating Representation Subspace Mapping with Unimodal Auxiliary Loss for Attention-based Multimodal Emotion Recognition
- Racial and Gender Stereotypes Encoded Into CLIP Representations