Vision-Language Models Suppress Female Representations Under Ambiguous Input

Even when AI models give neutral answers about a person's gender, they often "think" in a biased way. This study found that while models might say "a person" to be polite, their internal processing often identifies female-stereotyped jobs as female. They then "filter out" that information right before speaking. This leads the model to default to male instead.

The Problem

Current auditing for Vision-Language Models (VLMs) relies almost exclusively on output-level monitoring. We check if the model's generated text contains biased descriptors. Alignment techniques like RLHF (Reinforcement Learning from Human Feedback) have successfully taught models to avoid blatant demographic biases. This leads to neutral descriptions like "a person" when gender is unclear .

However, this surface-level neutrality is deceptive. The status quo fails because it assumes the model's output is a faithful window into its internal associations. As the authors demonstrate, a model can be "aligned" to hide its bias without actually removing it from its representations. This creates a massive blind spot for engineers. If you use VLM embeddings (vector representations of images) as features for downstream systems like image search or automated screening, the bias persists. Those pipelines consume the internal signal directly. They bypass the language head that alignment training was designed to control.

How It Works

To expose this hidden signal, the researchers introduce LALS (Latent Association Leaning Score). This is a zero-shot metric (requiring no training data) that peers into the model's "thoughts" at the level of individual visual tokens. The methodology follows a structured pipeline:

1. Text Projection: Using a procedure called LatentLens, the authors project the hidden states of visual tokens into the model's own text-embedding space. This translates "what this image patch looks like" into "what text concepts this patch resembles."
2. Concept Comparison: They build a reference corpus of balanced gendered terms (e.g., "man" vs. "woman"). For every projected visual token, LALS calculates the cosine similarity (a measure of how much two vectors point in the same direction) to these terms.
3. Score Aggregation: The score is the average gender pole ($+1$ or $-1$) of the $k$-nearest neighbors in the embedding space. To avoid noise, they aggregate only the top 5% of tokens with the strongest signal.
4. Layer-wise Sweeping: By calculating LALS at every layer of the transformer (the core neural architecture), they trace how these associations evolve from the initial vision encoder through the deep layers.

This approach differentiates between three regimes. These include occupations where the model is internally and externally consistent, occupations where both are female-leaning, and "divergence" occupations where the model encodes a female association internally but outputs a male default .

Numbers

The authors report a stark "male-mode collapse" when models are pushed with forced-choice prompting. While open-ended queries yield neutral text, forcing a guess reveals heavy bias. Many female-stereotyped occupations result in overwhelming male classifications. For example, hairdressers are classified as male 88–96% of the time. Babysitters are classified as male 72–96% of the time [Table 1]. This occurs across all four evaluated models (Qwen2-VL, Qwen2.5-VL, LLaVA, and InternVL2.5).

Crucially, the paper finds an asymmetric filtering problem. Layer-wise analysis shows that male signals are amplified end-to-end. Conversely, female signals peak in the mid-to-late layers and are then actively suppressed before reaching the output . To prove this wasn't just a correlation, they performed a causal intervention. They ablated (removed) the gender direction at layer 16 in Qwen2-VL-7B. They saw the forced-choice female rate drop in lockstep with the internal LALS signal. For nurses, the female rate dropped from 65% to 30% . This confirms the mid-layer signal is a necessary part of the causal pathway to the output.

What's Missing

The paper is technically rigorous, but there are gaps. First, the LALS metric relies on a binary gender lexicon. The authors suggest it could be adapted for non-binary or intersectional identities. However, they have not validated this. In production, relying on a binary proxy might create a false sense of security regarding other demographic biases.

Second, the authors show that removing the signal changes the output. They do not show that adding the signal can flip a male default to a female one. Third, the mechanism of suppression remains a black box. The paper identifies that the signal is suppressed in late layers. It does not pinpoint where. We do not know if this happens in specific attention heads or certain MLP (Multi-Layer Perceptron) blocks.

Should You Prototype This

Yes, but for auditing, not for fixing.

If you are building downstream services that consume VLM embeddings, you should prototype an LALS-style audit. The paper proves that "neutral" text outputs are a poor proxy for the actual features being embedded. However, do not expect this to be a turnkey solution for debiasing. The authors show that this asymmetry is established during pretraining and merely amplified by alignment. Simply applying more RLHF is unlikely to solve the underlying structural issue. Use LALS to map your risk surface. Realize that the bias is baked into the weights themselves.

Figures from the paper