Modern machine learning follows a "pre-train then fine-tune" paradigm. While this allows models to acquire vast general knowledge, it also forces them to inherit the sociocultural prejudices embedded in uncurated web-scale data. Researchers have recently identified a troubling pattern: even when a model is successfully "debiased" during its initial training, these biases often return—or even intensify—the moment the model is adapted to a specific downstream task.
The Persistence of Catastrophic Inheritance
When we fine-tune a foundation model, we aren't starting from a blank slate. Instead, the model brings a "bias floor" established during pre-training. Yupeng Chang and colleagues describe this as Catastrophic Inheritance (Bias-Aware Low-Rank Adaptation), where the adaptation process acts as a bottleneck that magnifies existing flaws rather than smoothing them over.
This inheritance is remarkably "sticky." Even if the downstream dataset is perfectly balanced or explicitly cleaned, the model may still favor name-based stereotypes or geographic biases learned months earlier. As Hadi Salman demonstrates in When does Bias Transfer in Transfer Learning?, pre-existing associations persist even when the target domain provides no reason to use them.
Why Fine-Tuning Re-activates Bias
The central tension in model adaptation is the trade-off between task performance and fairness. Fine-tuning is an optimization process: the model searches for the most efficient way to minimize loss on a new dataset. Often, the "easiest" way to achieve high accuracy is to exploit stereotypical shortcuts present in the training labels.
Masahiro Kaneko and others argue that debiasing is often superficial (Debiasing Isn't Enough!). They found that models quickly re-learn social biases because the fine-tuning process lacks the constraints to prevent the model from re-acquiring associations it was previously told to ignore. This leads to Bias Resurgence: a phenomenon where a "safe" model reverts to its biased state to better fit the distributional patterns of a specific task.
This can be viewed as a specialized form of catastrophic forgetting. As Guangliang Liu explains, models "forget" their fairness constraints as they optimize for new objectives (Towards Understanding Task-agnostic Debiasing). Internal mechanisms, such as specific attention heads that were neutralized during debiasing, can shift back toward biased states during fine-tuning.
The Fragility of Parameter-Efficient Adaptation
To save on computation, many practitioners use Parameter-Efficient Fine-Tuning (PEFT) methods like LoRA or Adapters. While these methods are efficient, they are not immune to bias. In fact, because PEFT methods often rely on the frozen internal knowledge of the base model, they can inadvertently anchor the model to its original, biased representations.
Sumit Agarwal notes that updating all weights (full fine-tuning) can sometimes purge biases more effectively than PEFT, but even then, the original "source-domain" biases often remain (PEFTDebias). Conversely, Anne Lauscher suggests that modular approaches like ADELE (Sustainable Modular Debiasing) can help by "patching" models with lightweight debiasing layers that are preserved even when the model is trained on new tasks.
Moving Toward Integrated Mitigation
The recurring evidence across these 73 papers suggests that fairness is not a static property that can be "fixed" once. Instead, it is a dynamic state that must be actively maintained throughout the model's lifecycle.
Recent work has shifted from post-hoc debiasing (fixing the model after training) to "debias-while-tuning." Methods like Causal-Debias, proposed by Fan Zhou, use causal invariant learning to force models to ignore "non-causal" demographic attributes while focusing on the actual task (Causal-Debias). Similarly, Xiangjue Dong introduced Co²PT, which uses counterfactual sentence pairs during the adaptation phase to ensure that fairness constraints are treated as a primary objective rather than a secondary concern (Co²PT).
By integrating debiasing directly into the adaptation process, researchers hope to break the cycle of inheritance and resurgence, ensuring that models remain fair even as they become more specialized.