¹

²

1 Introduction↩︎

Vision-language models (VLMs) are increasingly used in applications where fairness matters—from content moderation to image retrieval to assistive tools that describe visual scenes. As these models enter high-stakes settings, auditing them for bias has become a priority. The standard approach is straightforward: show the model an image, ask it a question, and check whether the output reflects stereotypical or harmful associations. If a model describes a doctor as “he” or a nurse as “she” when gender is ambiguous, these types of bias might be flagged [2].

This output-level auditing has driven significant progress. Alignment techniques such as RLHF [3] have made modern VLMs remarkably careful: when asked to describe an image of a worker whose gender is not visible, they generally answer “a person” rather than “a man” or “a woman.” We show that these outputs are only the surface, and the bias remains underneath. A model that produces neutral text may still carry biased representations—associations encoded in the activations of its visual tokens that shape downstream behavior even if they do not appear in the final response. These internal associations matter for at least two reasons. First, VLM embeddings are increasingly used as features for downstream systems (image search, content ranking, hiring tools), where biased representations propagate without ever passing through the model’s language thinking process. Second, output neutrality or clean inputs are a fragile condition: biases suppressed by alignment may resurface under different prompting strategies, not very clear visual inputs, or even fine-tuning, or deployment conditions.

In this paper, we ask a simple question: do VLMs’ internal visual representations carry the same gender associations as their outputs, even when the input images are ambiguous? To answer this, we introduce Lals (Latent Association Leaning Score), a zero-shot metric that measures concept associations at the level of individual visual tokens and layers. Lals builds on recent work showing that visual token activations in VLMs can be projected into the model’s text embedding space, enabling a direct reading of what each image patch “encodes” at any point in the network [4]. By comparing these decoded representations against a gender-balanced reference corpus, Lals produces a continuous score (from male-leaning to female-leaning) for every token at every layer, without any training. Our main findings are:

1. Internal representations and outputs are decoupled when input images are ambiguous. We identify three regimes: stereotypical occupations where internals and outputs agree on male (e.g., firefighter), where both agree on female (e.g., makeup artists), and sometimes where models internally encode female associations but output male (e.g., babysitter). This divergence regime represents a concrete blind spot for output-level auditing.

2. Late layers act as an asymmetric filter. Sweeping Lals across layers reveals that male associations amplify from early to late layers, while female associations peak in the mid-late of the network and are suppressed toward the output. This mechanism might be a potential explanation on why the male default dominates outputs even for occupations that are internally female-associated in non-obvious gender images.

3. Internal associations are shaped by culturally loaded visual cues. A color ablation shows that changing the clothing from blue to pink substantially reduces the internal male signal—not because the model is confused by color, but probably because it has learned the cultural gender associations that colors carry.

2 Related Work↩︎

2.0.0.1 Bias auditing in vision-language models.

Work on VLM bias has overwhelmingly operated at the output level. Early studies documented gender and racial biases in image captioning [5]–[7], and more recent benchmarks evaluate VLMs on occupation–gender defaults, counterfactual image pairs, and stereotype-consistent prompts [8]–[12]. All of these assume that a model’s output is a faithful window into its internal associations. In NLP, this assumption has been challenged: linear probes and embedding-space analyses repeatedly show that demographic biases persist after output-level debiasing [13]–[17]. Extending this line to VLMs remains underexplored. Most representation-level analyses focus on feature quality rather than social bias [18], and the few exceptions either operate on contrastive encoders rather than generative VLMs [19] or use causal mediation to localize bias to the image encoder without quantifying what is encoded at each layer [20]. This work fills this gap: we zero-shot and operate at token-level granularity, identifying which image patches carry biased associations and how they evolve across layers.

2.0.0.2 Interpreting internal representations in vision models.

A growing line of work reads intermediate representations by projecting them into interpretable spaces. LogitLens [21] projects hidden states into the output vocabulary, giving a coarse, word-level reading; TunedLens [22] refines this with learned per-layer affine transforms. Recent work has extended these tools to VLMs: LatentLens [4] shows that visual token activations can be meaningfully projected into the model’s text-embedding space, and [23] use logit lens to trace how object information flows through VLM layers. Lals adapts this projection in a new direction: rather than using it for general-purpose interpretability, we pair it with a structured text reference corpus to quantify demographic associations in a zero-shot, token-level manner. A complementary tradition—activation patching [24] and causal tracing [20], [25]—identifies which components are causally responsible for a behaviour by intervening on activations. These methods locate where a decision is made; our work measures what is encoded at each location. Our layer-sweep analysis connects the two by tracing how gender signal propagates through the network.

3 Approach↩︎

3.1 Lals: Latent Association Leaning Score↩︎

Lals measures the degree to which a visual token’s internal representation is associated with one pole of a concept dimension (e.g., male vs.female). It requires no training and operates at the level of individual tokens and layers.

3.1.0.1 Reference corpus.

We construct two balanced word lists for the target concept. For gender, one list contains male-associated terms (man, father, boy, husband, …) and the other female-associated terms (woman, mother, girl, wife, …), including gendered names and role terms. Each term is embedded using the VLM’s own text encoder, producing a reference database \(\mathcal{D} = \{(\mathbf{e}_i, g_i)\}\) where \(\mathbf{e}_i\) is the text embedding and \(g_i \in \{+1, -1\}\) indicates the concept pole.

3.1.0.2 Visual token projection.

Modern VLMs process images as sequences of visual tokens—patch-level vectors that pass through the same transformer layers as text. At any layer \(\ell\), we extract each visual token’s hidden state \(\mathbf{h}_t^\ell\) and project it into the text embedding space using the LatentLens procedure [4] , yielding a vector \(\mathbf{v}_t^\ell\) that lives in the same space as the reference corpus. This lets us directly compare what each image patch encodes against gendered text concepts.

3.1.0.3 Scoring and aggregation.

For each projected token, we retrieve its \(k\) nearest neighbors from \(\mathcal{D}\) by cosine similarity and compute the gender balance: \[\text{LALS}(t, \ell) \;=\; \frac{1}{k} \sum_{i \in \mathcal{N}_k(\mathbf{v}_t^\ell)} g_i \label{eq:lals95token}\tag{1}\] This produces a score in \([-1, +1]\): fully male-associated, fully female-associated, or balanced. To obtain an image-level score, we aggregate over the top 5% of tokens by absolute magnitude (validated empirically in Fig. 7, appendix), focusing on the patches with the strongest signal: \[\text{LALS}_{\text{image}}(\ell) \;=\; \frac{1}{|\mathcal{T}_{5\%}|} \sum_{t \,\in\, \mathcal{T}_{5\%}} \text{LALS}(t, \ell) \label{eq:lals95image}\tag{2}\] Negative values indicate male-leaning representations, positive values female-leaning, and values near zero no detectable association.

3.1.0.4 Properties.

Lals is zero-shot (no labeled images needed), token-level (revealing which image regions carry the association), layer-level (tracing how associations evolve through the network), and concept-general (swapping the reference corpus audits any attribute expressible as opposing text poles).

3.2 Experimental Setup↩︎

3.2.0.1 Models.

We evaluate four open-weight, instruction-tuned VLMs with different architectures, vision encoders, and vision–language connectors: Qwen2-VL-7B [26], Qwen2.5-VL-7B [27], LLaVA-v1.6-Mistral-7B [28], and InternVL2.5-8B [29]. We report Lals with \(k = 20\) neighbors and top-5% aggregation, unless stated otherwise.

3.2.0.2 Ambiguous-person dataset.

We use Google Gemini 2.5 Flash (image generation mode) [30] to generate images of faceless or obscured figures in occupation-specific settings, where gender cannot be determined from visual cues alone (Figure 2). A human annotator verified every image, discarding any with visible gender markers. The final dataset spans 15 occupations—male-stereotyped (e.g., firefighter, construction worker), female-stereotyped (e.g., nurse, florist), and neutral (e.g., chef, waiter)—with 60 images per occupation unless stated otherwise.

3.2.0.3 Output responses.

To compare internal representations with output behavior, we query each model with two prompt types. Open-ended: “Describe what this person is doing”—testing whether the model spontaneously attributes gender. Forced-choice (FC): “If you had to guess, is this person male or female? Answer in one word”—forcing an explicit commitment. We also run the FC prompt without any image to measure each model’s text-only prior.

4 Results↩︎

4.1 Lals Detects Bias Signal↩︎

Before applying Lals to ambiguous images, we verify that the metric (i) detects genuine gender signal when it is visually present, (ii) produces no spurious signal when people are absent, and (iii) is robust to methodological perturbations.

4.1.0.1 Localization on matched scenes.

We construct matched scene sets in which the same background is shown with no person, a man, a woman, or both, isolating Lals responses to gender-visible individuals while holding scene context constant. Figure 3 illustrates a kitchen scene under all four conditions. With no person present, the heatmap is nearly flat and the net Lals hovers near zero. Adding a man produces a clear male-leaning (blue) cluster localized on the person; adding a woman produces the opposite female-leaning (red) pattern in the corresponding region. When both are present, Lals correctly assigns male and female signal to the respective individuals. The pattern replicates on a construction-site scene (Appendix Fig. 8), and across person-free images (\(N{=}10\)) all net Lals values fall close to zero (mean \(=+0.001\), \(\sigma{=}0.005\)).

4.1.0.2 Controls.

Two additional checks guard against artifacts. Randomly permuting the gender labels in the reference corpus (shuffled database) collapses the signal by 98%, confirming that Lals depends on correct text–embedding alignment rather than on distributional properties of the embedding space. Varying the neighborhood size (\(k \in \{10, 20, 50\}\)) produces stable results (\(<\!15\%\) variation in gender delta), indicating that Lals is not sensitive to the exact number of nearest neighbors.

4.1.0.3 Cross-check with a supervised probe.

As an independent check, we train a logistic regression probe on visible-gender hidden states (\(N{=}200\), 5-fold cross-validation) to predict binary gender from visual token representations. The probe achieves 97% accuracy at layer 4 and 94.5% at layer 16. Applied to ambiguous-occupation images, the probe’s per-image \(P(\text{female})\) correlates with Lals (\(r=0.52\), \(p=0.003\)), confirming that both approaches capture overlapping structure in the representations. The moderate rather than near-perfect correlation is expected: the probe learns a single linear boundary, while Lals aggregates over a broader neighborhood of the embedding space.

4.2 Outputs Collapse Toward Male Under Ambiguity↩︎

To study how gender is represented in ambiguous inputs we first ask what models say when shown gender-ambiguous images. When prompted in open-ended format (“Describe what this person is doing”), all four models reliably produce gender-neutral responses across all 15 occupations: “the person is arranging flowers,” not “the woman is arranging flowers,” or they refuse to attribute gender at all. This is the expected effect of alignment training.

The behavior changes immediately under minimal prompt pressure. With a forced-choice (FC) prompt—“If you had to guess, is this person male or female?”—occupation-dependent defaults emerge sharply (Table 1). Firefighters are classified as male in 100% of images across all four models, which is unsurprising. More strikingly, most female-stereotyped occupations also collapse toward male: hairdresser (92% BLS female) is classified as male 88–96% of the time across all four models, babysitter (93% BLS female) is majority male in all models (72–96%), and preschool teacher (97% BLS female) is majority male in two of four models (LLaVA and InternVL). Even nurse—one of the most strongly female-coded occupations in the U.S.labor force at 87% BLS female [31]—is classified as male by LLaVA. The surface neutrality of open-ended outputs masks biases that become visible the moment a model is forced to commit.

A chain-of-thought variant of the FC prompt (Fig. 4; prompt in App. 8.1) makes the underlying reasoning explicit. For male-stereotyped occupations, the model cites visible cues (high-visibility jacket, drill) to justify a male guess. For moderately female-stereotyped occupations like florist, the model acknowledges the female stereotype in its reasoning—“these jobs are typically associated with women”—yet still concludes male. The override happens in plain sight: the model knows the stereotype and chooses against it in favor of a male default.

4.2.0.1 The pattern is one-sided.

Comparing model outputs against BLS ground-truth labor-force statistics (Table 1), all five male-stereotyped occupations (BLS %F \(<\!30\)) produce 64–100% male FC across every model—unsurprising. But six of the seven female-stereotyped occupations (BLS %F \(>\!70\)) also produce majority-male FC in most or all models: hairdresser (88–96% male), babysitter (72–96% male), maids/cleaning (99% male), preschool teacher (40–74% male), and florist (82–88% male). Only makeup artist consistently surfaces as female. The default direction is always male, never female: no occupation in our study produces majority-female FC against a male labor-force baseline. We refer to this one-sided behavior as male-mode collapse.

This raises the central question of the paper: does the male default reflect what models actually encode about each image, or only what they say? We address this in Section 4.3 by comparing forced-choice outputs against Lals measured directly on the visual token representations.

3pt

4.3 Layer Dynamics Reveal Asymmetric Filtering↩︎

The decoupling documented above raises a mechanistic question: at what point in the network does the female signal disappear? We answer this by computing Lals across layers for all four architectures, averaging trajectories across models within each regime (Fig. 5; per-model sweeps in App. Fig. 10).

4.3.0.1 Three qualitatively different trajectories.

The three regimes identified in Figure 5 show clearly distinct depth profiles. Agreement-male occupations (firefighter, construction worker, etc.) enter the network with strongly male-leaning Lals and maintain that signal end-to-end, with most curves dropping further into male territory at deeper layers (Fig. 5, right). Agreement-female occupations (nurse, makeup artist) are female-leaning from early layers, climb to a peak of \(+50\) to \(+80\%\) normalised Lals around relative depth \(0.7\)–\(0.8\), and then partially decline toward the output but remain clearly female-leaning at the final layer (Fig. 5, left). Divergence occupations (florist, preschool teacher, hairdresser, etc.) follow a qualitatively different trajectory: Lals rises through early layers, plateaus around \(+25\) to \(+40\%\) at mid-network depths, and then collapses sharply toward the final layer—in several cases crossing zero into male-leaning space (Fig. 5, middle). Real images replicate the effect. To rule out an artifact of synthetic image generation, we repeat the FC experiment on a set of real photographs of gender-ambiguous construction workers and nurses (Appendix Fig. 9). Model outputs and per-layer Lals trajectories closely match those on synthetic images (Pearson \(r{=}0.90\) and \(r{=}0.64\)), confirming that male-mode collapse is not an artifact of how we generated the test images.

4.3.0.2 The asymmetry is strictly directional.

Male signal passes through the full depth of the network unattenuated; female signal is the only direction that gets suppressed. No male-stereotyped occupation develops a female association that is subsequently filtered out. This asymmetric filtering connects directly to the forced-choice results: male-stereotyped occupations produce 100% male FC across all models, consistent with a signal preserved end-to-end. Agreement-female occupations partially survive the late-layer compression and reach majority-female FC in most models. But divergence occupations—which carry meaningful mid-layer female signal—never make it out: the late-layer collapse erodes the signal below threshold, and the model outputs male.

4.3.0.3 Cross-architecture consistency.

The three-regime structure replicates across all four architectures (App. Fig. 10). The two Qwen models and InternVL2.5 share a similar qualitative pattern of male amplification and female mid-layer peak followed by late collapse. LLaVA exhibits a milder variant in which female signals compress toward zero in late layers rather than crossing into male territory. InternVL2.5 produces the strongest overall male output bias in our study, consistent with even a small residual male lean at the final layer being sufficient to tip the forced-choice decision.

4.4 Where Does the Bias Come From?↩︎

The asymmetric filtering documented above raises a natural follow-up: where do these internal gender associations originate? We investigate three possible sources (visual content, alignment training, and the language model backbone) through three targeted experiments.

4.4.0.1 Visual cues modulate the signal.

We first test whether Lals responds to specific visual content by manipulating a single visual cue. Taking ambiguous images of construction workers and nurses, we vary only the color of one item of clothing (hat or scrubs), holding pose, scene, and all other cues constant (Fig. 6). Construction workers remain male-leaning across all conditions, but a pink hat reduces the male signal by roughly half; pink scrubs more than double the nurse’s female signal compared to blue scrubs. A single color change shifts the internal gender association by a magnitude comparable to the differences between entire occupation categories. This sensitivity likely reflects genuine structure in human culture: decades of psychological work have shown that pink predicts femininity in clothing, products, and environments so reliably that it functions, in effect, like a gendered pronoun [32]. The models appear to have internalised these social-chromatic associations, likely because pink in human-made environments genuinely co-occurs with female-associated contexts in training data.

4.4.0.2 Pretraining, not alignment, creates the asymmetry.

Having established that visual cues can shape the strength of internal associations, we next ask whether the late-layer suppression of female signal is introduced by instruction tuning—i.e., whether RLHF teaches the model to dampen female associations before generation. We run the same Lals layer sweep on the Qwen2-VL-7B base checkpoint (no instruction tuning) and compare it to the instruct variant (App. Fig. 11). The base model reproduces the same occupation-dependent profiles: nurse and florist are female-leaning at mid-layers, firefighter and construction worker are male-leaning, and the late-layer collapse of female signal appears in both variants—though more mildly in the base model. This suggests that the asymmetric structure is established during pretraining and amplified, rather than created, by alignment.

4.4.0.3 The collapse is specific to the visual pathway.

A remaining possibility is that the gender associations are inherited from the language model: perhaps the word “nurse” or the “color pink” already carries a female prior regardless of the image. We test this by feeding the model occupation names as text-only prompts (no image) and measuring Lals on the resulting text tokens (App. Fig. 11, text-only condition). The text-only dynamics are totally different: for female-stereotyped occupations (nurse, florist, librarian), the female signal amplifies in late layers—the opposite of the collapse observed for images. The vision encoder thus contributes a distinct, image-dependent component that diverges from the text-only baseline. The late-layer female collapse is specific to the visual pathway.

Taken together, these results suggest that internal gender associations are shaped by visual content (with fine-grained modulation by social cues such as color), are established during pretraining rather than alignment (base \(\approx\) instruct), and originate in the vision encoder rather than the language backbone (visual \(\neq\) text-only).

5 Discussion↩︎

When gender is clearly visible, modern VLMs behave well—alignment training has made their outputs largely accurate and appropriate [3]. The problems we document arise specifically when the model cannot tell: a figure in full gear, seen from behind, or too distant to read. In these cases, the model has to guess, and its guesses are not random. For most occupations the model defaults to male—even when its own internal representations lean female. A florist, a nurse, a preschool teacher: all encoded as female-associated inside the network, yet output as male under forced choice for some or most of the VLMs. The bias has not been removed; the model has learned not to express it. Our base-versus-instruct comparison supports this reading: the asymmetric structure is already present in the base model and is amplified rather than created by alignment, echoing NLP findings that output-level debiasing masks (or leaves intact) rather than eliminates representation-level bias [14], [17].

5.0.0.1 Why male?

Why the default direction is consistently male remains an open question. A simple distributional account (training data more often depicts people as male, so “male” is the safer completion when visual evidence is weak) is hard to reconcile with the consistency of the pattern across occupations whose corpora are not uniformly male-dominated. The asymmetry is also consistent with recent text-to-image findings that prompting for “a human” disproportionately produces male figures [33], suggesting the male default may operate in how models interpret images as well as how they generate them. Whether this prior originates in data frequency, the geometry of the embedding space, or some deeper psychologically-binding interaction of the two is an important direction for future work.

5.0.0.2 Why this matters in practice.

Ambiguous inputs are common: surveillance footage [34], blurred or distant figures, workers in protective gear—precisely the cases where downstream systems make decisions and biased priors carry the most risk [35]. Output-level evaluations—the current standard in academic benchmarks [5], [8] and industry red-teaming—will systematically miss the divergence we document, because the model produces neutral or male-default text regardless of what its representations encode. The risk extends beyond text generation: VLM embeddings are increasingly reused as features for image search, content ranking, and automated screening [36], where biased internal representations propagate without ever passing through the language head that alignment controls [37]. In these pipelines, what matters is not what the model would say but what it encodes. Lals provides a zero-shot, label-free tool for auditing at this level, and the reference corpus can be swapped to audit any concept expressible as opposing text poles—extending naturally to race, age, and intersectional attributes.

More broadly, our results suggest that alignment and debiasing are not the same thing. RLHF effectively controls what models say, and for clear images this is often sufficient. But for ambiguous inputs, alignment masks the underlying representations without modifying them. The color ablation illustrates the point: pink functions as a gendered semantic cue in the model’s visual processing, faithfully encoding the social-chromatic associations present in training data [32]. Whether a model that mirrors the gendered semiotics of human visual culture should be considered biased or simply faithful to the world it learned from is a question that extends beyond engineering—and one that representation-level tools like Lals can help inform.

6 Limitations↩︎

Our gender lexicon imposes a binary framework and covers only common English terms [38]; Lals is agnostic to lexicon contents and can in principle accommodate non-binary or intersectional categories, but we have not validated this. A second key question is causality. LALS measures geometric proximity in embedding space, which is consistent with but does not on its own establish a causal link to downstream behaviour. An activation-ablation experiment (Appendix 11) verifies that removing the mid-layer signal along a single gender direction shifts the forced-choice output in the predicted direction, supporting a necessity claim; the symmetric sufficiency test(additive steering to flip a male default to female) and a fuller localisation of the late-layer suppression mechanism remain open and future work.

7 Acknowledgements↩︎

This work was partially funded by Harvard Mind, Brain, Behavior Interfaculty Initiative (https://mbb.harvard.edu/) and Pivotal Research (https://www.pivotal-research.org/). Arnau Marin-Llobet is supported by Coefficient Giving and the RCC-Harvard Fellowship. Simon Henniger’s wor was supported by the Harvard Paulson SEAS Prize Fellowship and the German Academic Fellowship Organization, funded by the German Federal Ministry for Economic Affairs and Energy.

8 Appendix↩︎

This appendix provides implementation details and additional experiments that did not fit within the main paper. All experimental settings—models, prompts, ambiguous-person dataset, and Lals hyperparameters (\(k{=}20\), top-5% token aggregation)—follow Section 3 of the main paper unless stated otherwise.

8.1 Chain-of-Thought Prompt↩︎

For the chain-of-thought condition (Fig. 4, main paper), we use the following prompt with Qwen2-VL-7B-Instruct:

Look carefully at this image. Do you think the person is male or female? Think step by step. First, briefly describe the visual cues you can see (clothing, color, hair, posture, body shape, context, occupation, anything else relevant). Then commit to your best guess.
Use this exact format on two lines:
REASONING: <1-2 sentences listing the cues>
GUESS: <male or female>

8.2 Top-% Token Aggregation↩︎

Figure 7 validates the choice of top-5% aggregation used throughout the paper. We compute Lals on a held-out visible-gender set (Qwen2-VL) and measure two metrics as a function of the top-% of tokens aggregated by \(|\)Lals\(|\): ROC-AUC for predicting visible gender, and sign accuracy (whether the image-level Lals matches the true gender). Both metrics peak between 5–7% and degrade as low-magnitude tokens dilute the signal. We adopt 5% as the default, but the qualitative findings are stable across the 3–15% range.

9 Robustness Checks↩︎

9.1 Localization Replicates Across Scene Types↩︎

Figure 8 reproduces the kitchen-scene gender localization experiment from Section 4.1 in a construction-site setting. The empty scene yields near-zero Lals; inserting a man shifts the signal toward male (blue) and inserting a woman shifts it toward female (red), with the response localized on the inserted figure. This confirms that Lals responds to gender cues in the image rather than to scene context.

9.2 Real Photographs vs.Synthetic Images↩︎

A natural concern is that our findings may be specific to AI-generated images. Figure 9 compares layer-wise Lals trajectories on real photographs to those on our synthetic dataset for construction workers and nurses (Qwen2-VL, \(N{=}10\) per condition). The trajectories are closely aligned: Pearson \(r{=}0.90\) (\(p{=}0.006\)) for construction workers and \(r{=}0.64\) (\(p{=}0.122\)) for nurses. The lower significance for nurses reflects the small sample size (the shape of the curve matches well, but with \(N{=}10\) the correlation test is underpowered). The real-photo nurse set pools nurse and doctor images, both of which wear scrubs and are gender-ambiguous from typical angles.

10 Extended Layer Analyses↩︎

10.1 Per-Architecture Layer Sweep↩︎

Figure 10 shows the full per-architecture layer sweep across all 15 occupations and the four VLMs we evaluate. The qualitative pattern is consistent across architectures: male-leaning occupations enter the network with negative Lals and remain so through the final layer, while female-leaning occupations peak in mid-network depths (layers \(\sim\)​12–16 for the Qwen models; \(\sim\)​14–23 for LLaVA and InternVL) and are attenuated before the output. LLaVA exhibits the mildest collapse, compressing female signals toward zero rather than crossing into male-leaning space, while InternVL2.5 shows the strongest late-layer suppression—consistent with its near-100% male forced-choice rates on most occupations (Table 1, main paper). Despite differences in vision encoders, vision–language connectors, and training data, the male-amplify/female-suppress asymmetry is recovered in every architecture we tested.

10.2 Instruct vs.Base Model↩︎

To test whether the late-layer suppression of female signal is introduced by instruction tuning, we run the same Lals layer sweep on the Qwen2-VL-7B base checkpoint (no RLHF) and compare it to the instruct variant (Fig. 11). Both variants show the same occupation-dependent profiles: female-stereotyped occupations (red) peak around layer 16 and decline toward the output, while male-stereotyped occupations (blue) maintain or amplify their signal end-to-end. The late-layer collapse is milder in the base model—suggesting that instruction tuning amplifies the suppression—but the asymmetric structure is already present before RLHF, indicating that it is established during pretraining rather than introduced by alignment.

11 Causal Intervention: Is the Mid-Layer Signal Necessary?↩︎

The layer sweeps in Section 4.3 show that female-associated occupations carry a mid-network LALS peak that is attenuated before the final layer. These results are correlational: they establish that the signal exists and that the model nevertheless outputs male, but they do not show that the mid-layer signal is part of the causal pathway to the output. To test this, we directly ablate the signal and measure whether the forced-choice output moves.

11.0.0.1 Method.

We perform a single-direction activation intervention on Qwen2-VL-7B-Instruct. (i) We construct a gender direction \(\mathbf{d} \in \mathbb{R}^{d}\) from the model’s own text embeddings as the difference between the mean female-term and mean male-term embedding (same word lists used by LALS, §3.1), normalised to unit length. (ii) At layer 16 (the mid-network peak identified in Fig. 10), we register a forward hook that projects \(\mathbf{d}\) out of every visual-token hidden state during the forward pass: \[\mathbf{h}_t^{(16)} \;\leftarrow\; \mathbf{h}_t^{(16)} \;-\; \alpha\,(\mathbf{h}_t^{(16)} \cdot \mathbf{d})\,\mathbf{d},\] with \(\alpha=1\) (full ablation along \(\mathbf{d}\)). Text tokens are left untouched, and no other layer is modified. (iii) We re-run the model with the hook attached and measure both LALS at layer 16 and the forced-choice output. We evaluate \(N=20\) ambiguous images per occupation across seven occupations spanning the three regimes from Section 4.3.

11.0.0.2 Result.

Figure 12 shows the effect. Ablating the gender direction at layer 16 collapses the internal LALS signal by roughly 60–90% for the female-leaning occupations (nurse, preschool teacher, librarian, florist), and the forced-choice female rate drops in lockstep: nurse \(65 \rightarrow 30\%\), preschool teacher \(60 \rightarrow 40\%\), librarian \(50 \rightarrow 40\%\), pilot \(45 \rightarrow 30\%\). Male-default occupations move negligibly (firefighter stays at \(0\%\) female), consistent with the direction \(\mathbf{d}\) being aligned with female rather than male signal.

11.0.0.3 Interpretation.

Two findings follow. First, the mid-layer gender association LALS detects is causally involved in the model’s output: removing a single direction at a single mid-network layer is sufficient to shift the forced-choice distribution in the predicted direction across four occupations. This rules out the alternative reading in which LALS picks up an epiphenomenal cluster of female-coded contextual vocabulary (scrubs, classroom) that the model ignores when generating. Second, the effect is partial—nurse drops to \(30\%\) female rather than \(0\%\)—which is expected: a one-direction, one-layer ablation cannot eliminate gender information that is distributed across directions, tokens, and layers.

11.0.0.4 Scope and caveats.

The experiment is deliberately narrow and we flag the corresponding limits: (i) Necessity, not sufficiency. The intervention removes signal and moves outputs female \(\rightarrow\) male; we did not run the symmetric additive-steering experiment that would test whether adding \(\mathbf{d}\) flips a male default to female. We therefore make no debiasing or correction claim. (ii) Not a localisation of the late-layer suppression. The hook is placed at layer 16, where the female signal peaks; the experiment tests that this signal matters for the output, not where in the late layers it is filtered out. Identifying the specific components responsible for the late-layer collapse—attention heads, residual-stream subspaces, or specific MLP blocks—is an open question. (iii) Single model, single layer, single \(\alpha\). Results are reported for Qwen2-VL-7B-Instruct at layer 16 with \(\alpha{=}1\); whether the same direction transfers across architectures or whether intermediate \(\alpha\) values trace a smooth dose-response curve is left to future work. (iv) Sample size. \(N{=}20\) per occupation is sufficient to observe the qualitative shift but not to support fine-grained per-occupation significance claims; we report the result as a confirmatory mechanistic check rather than a quantitative effect estimate.

12 Reproducibility: Licenses and Compute↩︎

12.0.0.1 Model licenses.

All four VLMs we evaluate are open-weight and used strictly for forward-pass inference and activation extraction (no fine-tuning, no weight redistribution). Qwen2-VL-7B (both instruct and base), Qwen2.5-VL-7B, and LLaVA-v1.6-Mistral-7B are released under the Apache 2.0 license; InternVL2.5-8B is released under the MIT license. The ambiguous-occupation images were generated with Google Gemini 2.5 Flash under its standard terms of service; the dataset is fully synthetic and contains no real individuals.

12.0.0.2 Compute.

All experiments ran on a single NVIDIA H100 (80 GB) GPU and consumed approximately 25 GPU-hours in total, covering activation extraction, forced-choice and chain-of-thought generation across the four models, the robustness checks, the color ablation, the causal intervention and other experiments. LALS itself is training-free and adds negligible overhead beyond the forward pass.

References↩︎