Bias in large language models (LLMs) remains a persistent challenge, manifesting in stereotyping and unfair treatment across social groups. While prior research has primarily focused on individual models, the rise of multi-agent systems (MAS), where multiple LLMs collaborate and communicate, introduces new and largely unexplored dynamics in bias emergence and propagation. In this work, we present a comprehensive study of stereotypical bias in MAS, examining how internal specialization, underlying LLMs and inter-agent communication protocols influence bias robustness, propagation, and amplification. We simulate social contexts where agents represent different social groups and evaluate system behavior under various interaction and adversarial scenarios. Experiments on three bias benchmarks reveal that MAS are generally less robust than single-agent systems, with bias often emerging early through in-group favoritism. However, cooperative and debate-based communication can mitigate bias amplification, while more robust underlying LLMs improve overall system stability. Our findings highlight critical factors shaping fairness and resilience in multi-agent LLM systems.

Recent advances in large-scale generative language models have shown that reasoning capabilities can significantly improve model performance across a variety of tasks. However, the impact of reasoning on a model's ability to mitigate stereotypical responses remains largely underexplored. In this work, we investigate the crucial relationship between a model's reasoning ability and fairness, and ask whether improved reasoning capabilities can mitigate harmful stereotypical responses, especially those arising due to shallow or flawed reasoning. We conduct a comprehensive evaluation of multiple open-source LLMs, and find that larger models with stronger reasoning abilities exhibit substantially lower stereotypical bias on existing fairness benchmarks. Building on this insight, we introduce ReGiFT -- Reasoning Guided Fine-Tuning, a novel approach that extracts structured reasoning traces from advanced reasoning models and infuses them into models that lack such capabilities. We use only general-purpose reasoning and do not require any fairness-specific supervision for bias mitigation. Notably, we see that models fine-tuned using ReGiFT not only improve fairness relative to their non-reasoning counterparts but also outperform advanced reasoning models on fairness benchmarks. We also analyze how variations in the correctness of the reasoning traces and their length influence model fairness and their overall performance. Our findings highlight that enhancing reasoning capabilities is an effective, fairness-agnostic strategy for mitigating stereotypical bias caused by reasoning flaws.

Recent advances in generative AI have enabled visual content creation through text-to-image (T2I) generation. However, despite their creative potential, T2I models often replicate and amplify societal stereotypes -- particularly those related to gender, race, and culture -- raising important ethical concerns. This paper proposes a theory-driven bias detection rubric and a Social Stereotype Index (SSI) to systematically evaluate social biases in T2I outputs. We audited three major T2I model outputs -- DALL-E-3, Midjourney-6.1, and Stability AI Core -- using 100 queries across three categories -- geocultural, occupational, and adjectival. Our analysis reveals that initial outputs are prone to include stereotypical visual cues, including gendered professions, cultural markers, and western beauty norms. To address this, we adopted our rubric to conduct targeted prompt refinement using LLMs, which significantly reduced bias -- SSI dropped by 61% for geocultural, 69% for occupational, and 51% for adjectival queries. We complemented our quantitative analysis through a user study examining perceptions, awareness, and preferences around AI-generated biased imagery. Our findings reveal a key tension -- although prompt refinement can mitigate stereotypes, it can limit contextual alignment. Interestingly, users often perceived stereotypical images to be more aligned with their expectations. We discuss the need to balance ethical debiasing with contextual relevance and call for T2I systems that support global diversity and inclusivity while not compromising the reflection of real-world social complexity.

As large language models (LLMs) have been deployed in various real-world settings, concerns about the harm they may propagate have grown. Various jailbreaking techniques have been developed to expose the vulnerabilities of these models and improve their safety. This work reveals that many state-of-the-art LLMs are vulnerable to malicious requests hidden behind scientific language. Specifically, our experiments with GPT4o, GPT4o-mini, GPT-4, LLama3-405B-Instruct, Llama3-70B-Instruct, Cohere, Gemini models demonstrate that, the models' biases and toxicity substantially increase when prompted with requests that deliberately misinterpret social science and psychological studies as evidence supporting the benefits of stereotypical biases. Alarmingly, these models can also be manipulated to generate fabricated scientific arguments claiming that biases are beneficial, which can be used by ill-intended actors to systematically jailbreak these strong LLMs. Our analysis studies various factors that contribute to the models' vulnerabilities to malicious requests in academic language. Mentioning author names and venues enhances the persuasiveness of models, and the bias scores increase as dialogues progress. Our findings call for a more careful investigation on the use of scientific data for training LLMs.

With the introduction of (large) language models, there has been significant concern about the unintended bias such models may inherit from their training data. A number of studies have shown that such models propagate gender stereotypes, as well as geographical and racial bias, among other biases. While existing works tackle this issue by preprocessing data and debiasing embeddings, the proposed methods require a lot of computational resources and annotation effort while being limited to certain types of biases. To address these issues, we introduce REFINE-LM, a debiasing method that uses reinforcement learning to handle different types of biases without any fine-tuning. By training a simple model on top of the word probability distribution of a LM, our bias agnostic reinforcement learning method enables model debiasing without human annotations or significant computational resources. Experiments conducted on a wide range of models, including several LMs, show that our method (i) significantly reduces stereotypical biases while preserving LMs performance; (ii) is applicable to different types of biases, generalizing across contexts such as gender, ethnicity, religion, and nationality-based biases; and (iii) it is not expensive to train.

Differential Privacy (DP) can be applied to raw text by exploiting the spatial arrangement of words in an embedding space. We investigate the implications of such text privatization on Language Models (LMs) and their tendency towards stereotypical associations. Since previous studies documented that linguistic proficiency correlates with stereotypical bias, one could assume that techniques for text privatization, which are known to degrade language modeling capabilities, would cancel out undesirable biases. By testing BERT models trained on texts containing biased statements primed with varying degrees of privacy, our study reveals that while stereotypical bias generally diminishes when privacy is tightened, text privatization does not uniformly equate to diminishing bias across all social domains. This highlights the need for careful diagnosis of bias in LMs that undergo text privatization.

Large Multimodal Models (LMMs) have achieved great success recently, demonstrating a strong capability to understand multimodal information and to interact with human users. Despite the progress made, the challenge of detecting high-risk interactions in multimodal settings, and in particular in speech modality, remains largely unexplored. Conventional research on risk for speech modality primarily emphasises the content (e.g., what is captured as transcription). However, in speech-based interactions, paralinguistic cues in audio can significantly alter the intended meaning behind utterances. In this work, we propose a speech-specific risk taxonomy, covering 8 risk categories under hostility (malicious sarcasm and threats), malicious imitation (age, gender, ethnicity), and stereotypical biases (age, gender, ethnicity). Based on the taxonomy, we create a small-scale dataset for evaluating current LMMs capability in detecting these categories of risk. We observe even the latest models remain ineffective to detect various paralinguistic-specific risks in speech (e.g., Gemini 1.5 Pro is performing only slightly above random baseline). Warning: this paper contains biased and offensive examples.

The surge in Large Language Models (LLMs) development has led to improved performance on cognitive tasks as well as an urgent need to align these models with human values in order to safely exploit their power. Despite the effectiveness of preference learning algorithms like Reinforcement Learning From Human Feedback (RLHF) in aligning human preferences, their assumed improvements on model trustworthiness haven't been thoroughly testified. Toward this end, this study investigates how models that have been aligned with general-purpose preference data on helpfulness and harmlessness perform across five trustworthiness verticals: toxicity, stereotypical bias, machine ethics, truthfulness, and privacy. For model alignment, we focus on three widely used RLHF variants: Supervised Finetuning (SFT), Proximal Policy Optimization (PPO), and Direct Preference Optimization (DPO). Through extensive empirical investigations, we discover that the improvement in trustworthiness by RLHF is far from guaranteed, and there exists a complex interplay between preference data, alignment algorithms, and specific trustworthiness aspects. Together, our results underscore the need for more nuanced approaches for model alignment. By shedding light on the intricate dynamics of these components within model alignment, we hope this research will guide the community towards developing language models that are both capable and trustworthy.

Many evaluation measures are used to evaluate social biases in masked language models (MLMs). However, we find that these previously proposed evaluation measures are lacking robustness in scenarios with limited datasets. This is because these measures are obtained by comparing the pseudo-log-likelihood (PLL) scores of the stereotypical and anti-stereotypical samples using an indicator function. The disadvantage is the limited mining of the PLL score sets without capturing its distributional information. In this paper, we represent a PLL score set as a Gaussian distribution and use Kullback Leibler (KL) divergence and Jensen Shannon (JS) divergence to construct evaluation measures for the distributions of stereotypical and anti-stereotypical PLL scores. Experimental results on the publicly available datasets StereoSet (SS) and CrowS-Pairs (CP) show that our proposed measures are significantly more robust and interpretable than those proposed previously.

Transformer-based pretrained large language models (PLM) such as BERT and GPT have achieved remarkable success in NLP tasks. However, PLMs are prone to encoding stereotypical biases. Although a burgeoning literature has emerged on stereotypical bias mitigation in PLMs, such as work on debiasing gender and racial stereotyping, how such biases manifest and behave internally within PLMs remains largely unknown. Understanding the internal stereotyping mechanisms may allow better assessment of model fairness and guide the development of effective mitigation strategies. In this work, we focus on attention heads, a major component of the Transformer architecture, and propose a bias analysis framework to explore and identify a small set of biased heads that are found to contribute to a PLM's stereotypical bias. We conduct extensive experiments to validate the existence of these biased heads and to better understand how they behave. We investigate gender and racial bias in the English language in two types of Transformer-based PLMs: the encoder-based BERT model and the decoder-based autoregressive GPT model. Overall, the results shed light on understanding the bias behavior in pretrained language models. However, since PLMs and foundation models are trained on large humanwritten corpora, they often encode undesired stereotypes towards different social groups, such as gender, race, or people with disabilities (Bender et al., 2021; Blodgett et al., 2020; Hutchinson et al., 2020).