Recursive prompting with large language models enables scalable synthetic dataset generation but introduces the risk of bias amplification. We investigate gender bias dynamics across three generations of recursive text generation using three complementary evaluation frameworks: rule-based pattern matching, embedding-based semantic similarity, and downstream task performance. Experiments with three initial bias levels (0.1, 0.3, 0.6) and four mitigation strategies reveal equilibrium dynamics rather than monotonic amplification. The low initial bias amplifies toward the model's inherent bias level (+36%), whereas the high initial bias decays toward it (-26%). Among mitigation methods, contrastive augmentation, which introduces gender-swapped variants, achieves significant downstream bias reduction (98.8% for low initial bias and 91% on average) despite producing higher embedding-based bias scores. This paradox demonstrates that semantic similarity metrics may diverge from behavioral fairness outcomes, highlighting the need for multidimensional evaluation in responsible synthetic data generation.

Coreset selection methods have shown promise in reducing the training data size while maintaining model performance for data-efficient machine learning. However, as many datasets suffer from biases that cause models to learn spurious correlations instead of causal features, it is important to understand whether and how dataset reduction methods may perpetuate, amplify, or mitigate these biases. In this work, we conduct the first comprehensive analysis of the implications of data selection on the spurious bias levels of the selected coresets and the robustness of downstream models trained on them. We use an extensive experimental setting spanning ten different spurious correlations benchmarks, five score metrics to characterize sample importance/ difficulty, and five data selection policies across a broad range of coreset sizes. Thereby, we unravel a series of nontrivial nuances in interactions between sample difficulty and bias alignment, as well as dataset bias and resultant model robustness. For example, we find that selecting coresets using embedding-based sample characterization scores runs a comparatively lower risk of inadvertently exacerbating bias than selecting using characterizations based on learning dynamics. Most importantly, our analysis reveals that although some coreset selection methods could achieve lower bias levels by prioritizing difficult samples, they do not reliably guarantee downstream robustness.

Although value-aligned language models (LMs) appear unbiased in explicit bias evaluations, they often exhibit stereotypes in implicit word association tasks, raising concerns about their fair usage. We investigate the mechanisms behind this discrepancy and find that alignment surprisingly amplifies implicit bias in model outputs. Specifically, we show that aligned LMs, unlike their unaligned counterparts, overlook racial concepts in early internal representations when the context is ambiguous. Not representing race likely fails to activate safety guardrails, leading to unintended biases. Inspired by this insight, we propose a new bias mitigation strategy that works by incentivizing the representation of racial concepts in the early model layers. In contrast to conventional mitigation methods of machine unlearning, our interventions find that steering the model to be more aware of racial concepts effectively mitigates implicit bias. Similar to race blindness in humans, ignoring racial nuances can inadvertently perpetuate subtle biases in LMs.

Diffusion models have demonstrated impressive capabilities in synthesizing diverse content. However, despite their high-quality outputs, these models often perpetuate social biases, including those related to gender and race. These biases can potentially contribute to harmful real-world consequences, reinforcing stereotypes and exacerbating inequalities in various social contexts. While existing research on diffusion bias mitigation has predominantly focused on guiding content generation, it often neglects the intrinsic mechanisms within diffusion models that causally drive biased outputs. In this paper, we investigate the internal processes of diffusion models, identifying specific decision-making mechanisms, termed bias features, embedded within the model architecture. By directly manipulating these features, our method precisely isolates and adjusts the elements responsible for bias generation, permitting granular control over the bias levels in the generated content. Through experiments on both unconditional and conditional diffusion models across various social bias attributes, we demonstrate our method's efficacy in managing generation distribution while preserving image quality. We also dissect the discovered model mechanism, revealing different intrinsic features controlling fine-grained aspects of generation, boosting further research on mechanistic interpretability of diffusion models.

While task-agnostic debiasing provides notable generalizability and reduced reliance on downstream data, its impact on language modeling ability and the risk of relearning social biases from downstream task-specific data remain as the two most significant challenges when debiasing Pretrained Language Models (PLMs). The impact on language modeling ability can be alleviated given a high-quality and long-contextualized debiasing corpus, but there remains a deficiency in understanding the specifics of relearning biases. We empirically ascertain that the effectiveness of task-agnostic debiasing hinges on the quantitative bias level of both the task-specific data used for downstream applications and the debiased model. We empirically show that the lower bound of the bias level of the downstream fine-tuned model can be approximated by the bias level of the debiased model, in most practical cases. To gain more in-depth understanding about how the parameters of PLMs change during fine-tuning due to the forgetting issue of PLMs, we propose a novel framework which can Propagate Socially-fair Debiasing to Downstream Fine-tuning, ProSocialTuning. Our proposed framework can push the fine-tuned model to approach the bias lower bound during downstream fine-tuning, indicating that the ineffectiveness of debiasing can be alleviated by overcoming the forgetting issue through regularizing successfully debiased attention heads based on the PLMs' bias levels from stages of pretraining and debiasing.

Large Language Models (LLMs), now used daily by millions of users, can encode societal biases, exposing their users to representational harms. A large body of scholarship on LLM bias exists but it predominantly adopts a Western-centric frame and attends comparatively less to bias levels and potential harms in the Global South. In this paper, we quantify stereotypical bias in popular LLMs according to an Indian-centric frame and compare bias levels between the Indian and Western contexts. To do this, we develop a novel dataset which we call Indian-BhED (Indian Bias Evaluation Dataset), containing stereotypical and anti-stereotypical examples for caste and religion contexts. We find that the majority of LLMs tested are strongly biased towards stereotypes in the Indian context, especially as compared to the Western context. We finally investigate Instruction Prompting as a simple intervention to mitigate such bias and find that it significantly reduces both stereotypical and anti-stereotypical biases in the majority of cases for GPT-3.5. The findings of this work highlight the need for including more diverse voices when evaluating LLMs.

Most continual learning (CL) algorithms have focused on tackling the stability-plasticity dilemma, that is, the challenge of preventing the forgetting of previous tasks while learning new ones. However, they have overlooked the impact of the knowledge transfer when the dataset in a certain task is biased - namely, when some unintended spurious correlations of the tasks are learned from the biased dataset. In that case, how would they affect learning future tasks or the knowledge already learned from the past tasks? In this work, we carefully design systematic experiments using one synthetic and two real-world datasets to answer the question from our empirical findings. Specifically, we first show through two-task CL experiments that standard CL methods, which are unaware of dataset bias, can transfer biases from one task to another, both forward and backward, and this transfer is exacerbated depending on whether the CL methods focus on the stability or the plasticity. We then present that the bias transfer also exists and even accumulate in longer sequences of tasks. Finally, we propose a simple, yet strong plug-in method for debiasing-aware continual learning, dubbed as Group-class Balanced Greedy Sampling (BGS). As a result, we show that our BGS can always reduce the bias of a CL model, with a slight loss of CL performance at most.

Model explanations such as saliency maps can improve user trust in AI by highlighting important features for a prediction. However, these become distorted and misleading when explaining predictions of images that are subject to systematic error (bias). Furthermore, the distortions persist despite model fine-tuning on images biased by different factors (blur, color temperature, day/night). We present Debiased-CAM to recover explanation faithfulness across various bias types and levels by training a multi-input, multi-task model with auxiliary tasks for explanation and bias level predictions. In simulation studies, the approach not only enhanced prediction accuracy, but also generated highly faithful explanations about these predictions as if the images were unbiased. In user studies, debiased explanations improved user task performance, perceived truthfulness and perceived helpfulness. Debiased training can provide a versatile platform for robust performance and explanation faithfulness for a wide range of applications with data biases.

Bias in Word Embeddings has been a subject of recent interest, along with efforts for its reduction. Current approaches show promising progress towards debiasing single bias dimensions such as gender or race. In this paper, we present a joint multiclass debiasing approach that is capable of debiasing multiple bias dimensions simultaneously. In that direction, we present two approaches, HardWEAT and SoftWEAT, that aim to reduce biases by minimizing the scores of the Word Embeddings Association Test (WEAT). We demonstrate the viability of our methods by debiasing Word Embeddings on three classes of biases (religion, gender and race) in three different publicly available word embeddings and show that our concepts can both reduce or even completely eliminate bias, while maintaining meaningful relationships between vectors in word embeddings. Our work strengthens the foundation for more unbiased neural representations of textual data.