Bias Amplification: Language Models as Increasingly Biased Media

As Large Language Models (LLMs) become increasingly integrated into various facets of society, a significant portion of online text consequently become synthetic. This raises concerns about bias amplification, a phenomenon where models trained on synthetic data amplify the pre-existing biases over successive training iterations. Previous literature seldom discusses bias amplification as an independent issue from model collapse. In this work, we address the gap in understanding the bias amplification of LLMs with four main contributions. Firstly, we propose a theoretical framework, defining the necessary and sufficient conditions for its occurrence, and emphasizing that it occurs independently of model collapse. Using statistical simulations with weighted maximum likelihood estimation, we demonstrate the framework and show how bias amplification arises without the sampling and functional form issues that typically drive model collapse. Secondly, we conduct experiments with GPT-2 to empirically demonstrate bias amplification, specifically examining open-ended generational political bias with a benchmark we developed. We observe that GPT-2 exhibits a right-leaning bias in sentence continuation tasks and that the bias progressively increases with iterative fine-tuning on synthetic data generated by previous iterations. Thirdly, we explore three potential mitigation strategies: Overfitting, Preservation, and Accumulation. We find that both Preservation and Accumulation effectively mitigate bias amplification and model collapse. Finally, using novel mechanistic interpretation techniques, we demonstrate that in the GPT-2 experiments, bias amplification and model collapse are driven by distinct sets of neurons, which aligns with our theoretical framework.

Assessing Bias in Metric Models for LLM Open-Ended Generation Bias Benchmarks

Open-generation bias benchmarks evaluate social biases in Large Language Models (LLMs) by analyzing their outputs. However, the classifiers used in analysis often have inherent biases, leading to unfair conclusions. This study examines such biases in open-generation benchmarks like BOLD and SAGED. Using the MGSD dataset, we conduct two experiments. The first uses counterfactuals to measure prediction variations across demographic groups by altering stereotype-related prefixes. The second applies explainability tools (SHAP) to validate that the observed biases stem from these counterfactuals. Results reveal unequal treatment of demographic descriptors, calling for more robust bias metric models.

SAGED: A Holistic Bias-Benchmarking Pipeline for Language Models with Customisable Fairness Calibration

The development of unbiased large language models is widely recognized as crucial, yet existing benchmarks fall short in detecting biases due to limited scope, contamination, and lack of a fairness baseline. SAGED(-Bias) is the first holistic benchmarking pipeline to address these problems. The pipeline encompasses five core stages: scraping materials, assembling benchmarks, generating responses, extracting numeric features, and diagnosing with disparity metrics. SAGED includes metrics for max disparity, such as impact ratio, and bias concentration, such as Max Z-scores. Noticing that assessment tool bias and contextual bias in prompts can distort evaluation, SAGED implements counterfactual branching and baseline calibration for mitigation. For demonstration, we use SAGED on G20 Countries with popular 8b-level models including Gemma2, Llama3.1, Mistral, and Qwen2. With sentiment analysis, we find that while Mistral and Qwen2 show lower max disparity and higher bias concentration than Gemma2 and Llama3.1, all models are notably biased against countries like Russia and (except for Qwen2) China. With further experiments to have models role-playing U.S. (vice-/former-) presidents, we see bias amplifies and shifts in heterogeneous directions. Moreover, we see Qwen2 and Mistral not engage in role-playing, while Llama3.1 and Gemma2 role-play Trump notably more intensively than Biden and Harris, indicating role-playing performance bias in these models.

JobFair: A Framework for Benchmarking Gender Hiring Bias in Large Language Models

This paper presents a novel framework for benchmarking hierarchical gender hiring bias in Large Language Models (LLMs) for resume scoring, revealing significant issues of reverse bias and overdebiasing. Our contributions are fourfold: First, we introduce a framework using a real, anonymized resume dataset from the Healthcare, Finance, and Construction industries, meticulously used to avoid confounding factors. It evaluates gender hiring biases across hierarchical levels, including Level bias, Spread bias, Taste-based bias, and Statistical bias. This framework can be generalized to other social traits and tasks easily. Second, we propose novel statistical and computational hiring bias metrics based on a counterfactual approach, including Rank After Scoring (RAS), Rank-based Impact Ratio, Permutation Test-Based Metrics, and Fixed Effects Model-based Metrics. These metrics, rooted in labor economics, NLP, and law, enable holistic evaluation of hiring biases. Third, we analyze hiring biases in ten state-of-the-art LLMs. Six out of ten LLMs show significant biases against males in healthcare and finance. An industry-effect regression reveals that the healthcare industry is the most biased against males. GPT-4o and GPT-3.5 are the most biased models, showing significant bias in all three industries. Conversely, Gemini-1.5-Pro, Llama3-8b-Instruct, and Llama3-70b-Instruct are the least biased. The hiring bias of all LLMs, except for Llama3-8b-Instruct and Claude-3-Sonnet, remains consistent regardless of random expansion or reduction of resume content. Finally, we offer a user-friendly demo to facilitate adoption and practical application of the framework.

Multi-hop Commonsense Knowledge Injection Framework for Zero-Shot Commonsense Question Answering

Commonsense question answering (QA) research requires machines to answer questions based on commonsense knowledge. However, this research requires expensive labor costs to annotate data as the basis of research, and models that rely on fine-tuning paradigms only apply to specific tasks, rather than learn a general commonsense reasoning ability. As a more robust method, zero-shot commonsense question answering shows a good prospect. The current zero-shot framework tries to convert triples in commonsense knowledge graphs (KGs) into QA-form samples as the pre-trained data source to incorporate commonsense knowledge into the model. However, this method ignores the multi-hop relationship in the KG, which is also an important central problem in commonsense reasoning. In this paper, we propose a novel multi-hop commonsense knowledge injection framework. Specifically, it explores multi-hop reasoning paradigm in KGs that conform to linguistic logic, and we further propose two multi-hop QA generation methods based on KGs. Then, we utilize contrastive learning to pre-train the model with the synthetic QA dataset to inject multi-hop commonsense knowledge. Extensive experiments on five commonsense question answering benchmarks demonstrate that our framework achieves state-of-art performance.

Once and for All: Scheduling Multiple Users Using Statistical CSI under Fixed Wireless Access

Conventional multi-user scheduling schemes are designed based on instantaneous channel state information (CSI), indicating that decisions must be made every transmission time interval (TTI) which lasts at most several milliseconds. Only quite simple approaches can be exploited under this stringent time constraint, resulting in less than satisfactory scheduling performance. In this paper, we investigate the scheduling problem of a fixed wireless access (FWA) network using only statistical CSI. Thanks to their fixed positions, user terminals in FWA can easily provide reliable large-scale CSI lasting tens or even hundreds of TTIs. Inspired by this appealing fact, we propose an \emph{`once-and-for-all'} scheduling approach, i.e. given multiple TTIs sharing identical statistical CSI, only a single high-quality scheduling decision lasting across all TTIs shall be taken rather than repeatedly making low-quality decisions every TTI. The proposed scheduling design is essentially a mixed-integer non-smooth non-convex stochastic problem with the objective of maximizing the weighted sum rate as well as the number of active users. We firstly replace the indicator functions in the considered problem by well-chosen sigmoid functions to tackle the non-smoothness. Via leveraging deterministic equivalent technique, we then convert the original stochastic problem into an approximated deterministic one, followed by linear relaxation of the integer constraints. However, the converted problem is still highly non-convex due to implicit equation constraints introduced by deterministic equivalent. To address this issue, we employ implicit optimization technique so that the gradient can be derived explicitly, with which we propose an algorithm design based on accelerated Frank-Wolfe method. Numerical results verify the effectiveness of our proposed scheduling scheme over state-of-the-art.

Simple Yet Surprisingly Effective Training Strategies for LSTMs in Sensor-Based Human Activity Recognition

Human Activity Recognition (HAR) is one of the core research areas in mobile and wearable computing. With the application of deep learning (DL) techniques such as CNN, recognizing periodic or static activities (e.g, walking, lying, cycling, etc.) has become a well studied problem. What remains a major challenge though is the sporadic activity recognition (SAR) problem, where activities of interest tend to be non periodic, and occur less frequently when compared with the often large amount of irrelevant background activities. Recent works suggested that sequential DL models (such as LSTMs) have great potential for modeling nonperiodic behaviours, and in this paper we studied some LSTM training strategies for SAR. Specifically, we proposed two simple yet effective LSTM variants, namely delay model and inverse model, for two SAR scenarios (with and without time critical requirement). For time critical SAR, the delay model can effectively exploit predefined delay intervals (within tolerance) in form of contextual information for improved performance. For regular SAR task, the second proposed, inverse model can learn patterns from the time series in an inverse manner, which can be complementary to the forward model (i.e.,LSTM), and combining both can boost the performance. These two LSTM variants are very practical, and they can be deemed as training strategies without alteration of the LSTM fundamentals. We also studied some additional LSTM training strategies, which can further improve the accuracy. We evaluated our models on two SAR and one non-SAR datasets, and the promising results demonstrated the effectiveness of our approaches in HAR applications.

Decentralized Linear MMSE Equalizer Under Colored Noise for Massive MIMO Systems

Conventional uplink equalization in massive MIMO systems relies on a centralized baseband processing architecture. However, as the number of base station antennas increases, centralized baseband processing architectures encounter two bottlenecks, i.e., the tremendous data interconnection and the high-dimensional computation. To tackle these obstacles, decentralized baseband processing was proposed for uplink equalization, but only applicable to the scenarios with unpractical white Gaussian noise assumption. This paper presents an uplink linear minimum mean-square error (L-MMSE) equalization method in the daisy chain decentralized baseband processing architecture under colored noise assumption. The optimized L-MMSE equalizer is derived by exploiting the block coordinate descent method, which shows near-optimal performance both in theoretical and simulation while significantly mitigating the bottlenecks.