When we talk about bias in machine learning, the conversation almost always centers on data. The prevailing intuition is simple: "garbage in, garbage out." If a model outputs harmful stereotypes or brittle logical shortcuts, we assume it is merely reflecting the flaws of its training corpus.
However, a growing body of research reveals that data is only half the story. The architecture of a model—its structural design, attention mechanisms, layer depth, and optimization objectives—acts as an active lens that can amplify, suppress, or fundamentally reshape these biases. Two models trained on the exact same dataset can develop entirely different "bias signatures." Understanding why this variance occurs requires looking past the data and examining the structural mechanics of neural networks.
The Architectural Lens
Why do different architectures react differently to identical data? It comes down to how their internal representational strategies process context.
Consider how a model reads a sentence. An encoder-only model (like BERT) looks at a sequence bidirectionally, while a decoder-only model (like GPT) processes text autoregressively from left to right. This fundamental difference in how context is constructed changes how stereotypes are retrieved. Öztürk et al. demonstrated that decoder-only models can exhibit entirely different, sometimes even anti-stereotypical, tendencies compared to encoder-only models when evaluated on the same cross-linguistic data.
Bidirectional (Encoder-only) ---> Processes entire context simultaneously (High bias retention)
Autoregressive (Decoder-only) ---> Processes context sequentially token-by-token (Altered bias retrieval)
This structural divergence appears across various downstream tasks. For instance, when translating natural language into database queries, encoder-only structures show a significantly higher susceptibility to inheriting social biases from training data than encoder-decoder structures, as observed by Gao et al.. Even when we step outside of transformers entirely, Vanmassenhove et al. showed that phrase-based statistical models, recurrent neural networks (RNNs), and self-attention architectures exhibit distinct profiles of linguistic diversity loss. Each architecture's unique inductive bias dictates exactly how much morphological and lexical variety is stripped away during training.
Structural and Layer-wise Mechanics
Bias is not a uniform fog that settles evenly across a network's weights; it concentrates and fluctuates depending on structural depth.
In model editing, for example, networks often suffer from "lexical bias"—mistaking simple word overlap for true semantic meaning. Rizwan et al. discovered that this lexical bias is highly volatile across a model's depth, typically peaking sharply in the middle layers and fluctuating elsewhere. This makes intervention highly architecture-dependent: a debiasing technique applied to the wrong layer will completely miss the mark.
[Input Layer] ---> [Early Layers: Latent Transfer] ---> [Middle Layers: Lexical Bias Peaks] ---> [Output Layer]
A similar structural sensitivity occurs during knowledge distillation, where a smaller student model learns from a larger teacher. Schrodi et al. found that a student's susceptibility to inheriting a teacher's hidden, latent biases varies wildly depending on the student's architecture. Crucially, early layers play a disproportionate role in this subliminal transfer. Some student structures remain completely immune to the teacher's biases, while others absorb them entirely, proving that the structural recipe itself governs a model's vulnerability to bias propagation.
Idiosyncratic Bias Signatures
Because architecture shapes how information flows, different model families develop highly predictable, idiosyncratic behavioral quirks.
One prominent example is positional bias. Autoregressive models frequently suffer from serial position effects, showing a strong preference for information located at the very beginning (primacy) or the very end (recency) of a prompt. Vosoughi et al. argue that these positional preferences are not superficial prompting artifacts that can be engineered away; they are inherent architectural traits deeply woven into how generative models process sequences.
This manifests clearly in multimodal and evaluation settings:
- Vision-Language Processing: Tian highlights how positional biases fluctuate wildly based on image placement in decoder-only vision-language models.
- Multiple-Choice Selection: When faced with uncertainty, models often "blindly guess" based on option placement. Loginova et al. demonstrated that different neural backbones exhibit highly consistent, unique signatures in selection bias—such as a rigid preference for option "A"—proving that the underlying language backbone heavily dictates how a multimodal system handles uncertainty.
This architectural gravity is so strong that it frequently overrides explicit instructions. When researchers use "sociodemographic prompting" to force a model to adopt a specific persona (e.g., instructing it to respond as a specific age or political demographic), Beck et al. found that the model's default architectural bias consistently outweighs the persona instruction. Models from different families assigned the same persona behave less similarly than a single model family forced into different personas.
Why This Matters for Mitigation
The realization that bias varies by architecture fundamentally changes how we must approach AI safety and fairness. If bias were purely a property of the data, a single data-cleansing pipeline or a universal fine-tuning dataset would suffice.
Instead, mitigation must be architecture-aware. As Cai et al. point out, debiasing strategies that treat the model as a black box and ignore its specific processing mechanics fail because they do not address the unique structural shortcuts the model has internalized. Furthermore, Liu et al. demonstrated that different language model architectures develop entirely distinct sets of spurious correlations even when trained on identical data—one might over-index on geographic identifiers while another fixates on gendered terms.
Ultimately, building fairer models requires us to move past the data-centric paradigm. We must treat network architecture not just as an optimization engine, but as an active participant that dictates how a model perceives, retains, and expresses bias.
Key Foundational Work
The Surface Realisation Task: Recent Developments and Future Plans (2012) — contributed 3 concept elements to this direction.
Addressing Age-Related Bias in Sentiment Analysis (2019) — contributed 1 concept elements to this direction.