Zhongteng Cai

Yaxuan Wang, Zhongteng Cai, Yujia Bao et al.

The rapid advancement of large language models (LLMs) has led to growing interest in using synthetic data to train future models. However, this creates a self-consuming retraining loop, where models are trained on their own outputs and may cause performance drops and induce emerging biases. In real-world applications, previously deployed LLMs may influence the data they generate, leading to a dynamic system driven by user feedback. For example, if a model continues to underserve users from a group, less query data will be collected from this particular demographic of users. In this study, we introduce the concept of \textbf{S}elf-\textbf{C}onsuming \textbf{P}erformative \textbf{L}oop (\textbf{SCPL}) and investigate the role of synthetic data in shaping bias during these dynamic iterative training processes under controlled performative feedback. This controlled setting is motivated by the inaccessibility of real-world user preference data from dynamic production systems, and enables us to isolate and analyze feedback-driven bias evolution in a principled manner. We focus on two types of loops, including the typical retraining setting and the incremental fine-tuning setting, which is largely underexplored. Through experiments on three real-world tasks, we find that the performative loop increases preference bias and decreases disparate bias. We design a reward-based rejection sampling strategy to mitigate the bias, moving towards more trustworthy self-improving systems.

Zhongteng Cai, Yaxuan Wang, Yang Liu et al.

As synthetic data proliferates across the Internet, it is often reused to train successive generations of generative models. This creates a ``self-consuming loop" that can lead to training instability or \textit{model collapse}. Common strategies to address the issue -- such as accumulating historical training data or injecting fresh real data -- either increase computational cost or require expensive human annotation. In this paper, we empirically analyze the latent space dynamics of self-consuming diffusion models and observe that the low-dimensional structure of latent representations extracted from synthetic data degrade over generations. Based on this insight, we propose \textit{Latent Space Filtering} (LSF), a novel approach that mitigates model collapse by filtering out less realistic synthetic data from mixed datasets. Theoretically, we present a framework that connects latent space degradation to empirical observations. Experimentally, we show that LSF consistently outperforms existing baselines across multiple real-world datasets, effectively mitigating model collapse without increasing training cost or relying on human annotation.

Zhongteng Cai, Mohammad Mahdi Khalili, Xueru Zhang

As machine learning (ML) algorithms are increasingly used in social domains to make predictions about humans, there is a growing concern that these algorithms may exhibit biases against certain social groups. Numerous notions of fairness have been proposed in the literature to measure the unfairness of ML. Among them, one class that receives the most attention is \textit{parity-based}, i.e., achieving fairness by equalizing treatment or outcomes for different social groups. However, achieving parity-based fairness often comes at the cost of lowering model accuracy and is undesirable for many high-stakes domains like healthcare. To avoid inferior accuracy, a line of research focuses on \textit{preference-based} fairness, under which any group of individuals would experience the highest accuracy and collectively prefer the ML outcomes assigned to them if they were given the choice between various sets of outcomes. However, these works assume individual demographic information is known and fully accessible during training. In this paper, we relax this requirement and propose a novel \textit{demographic-agnostic fairness without harm (DAFH)} optimization algorithm, which jointly learns a group classifier that partitions the population into multiple groups and a set of decoupled classifiers associated with these groups. Theoretically, we conduct sample complexity analysis and show that our method can outperform the baselines when demographic information is known and used to train decoupled classifiers. Experiments on both synthetic and real data validate the proposed method.

Zhongteng Cai, Xueru Zhang, Mohammad Mahdi Khalili

Differential privacy mechanisms such as the Gaussian or Laplace mechanism have been widely used in data analytics for preserving individual privacy. However, they are mostly designed for continuous outputs and are unsuitable for scenarios where discrete values are necessary. Although various quantization mechanisms were proposed recently to generate discrete outputs under differential privacy, the outcomes are either biased or have an inferior accuracy-privacy trade-off. In this paper, we propose a family of quantization mechanisms that is unbiased and differentially private. It has a high degree of freedom and we show that some existing mechanisms can be considered as special cases of ours. To find the optimal mechanism, we formulate a linear optimization that can be solved efficiently using linear programming tools. Experiments show that our proposed mechanism can attain a better privacy-accuracy trade-off compared to baselines.