In-context learning (ICL) has proven to be adept at adapting large language models (LLMs) to downstream tasks without parameter updates, based on a few demonstration examples. Prior work has found that the ICL performance is susceptible to the selection of examples in prompt and made efforts to stabilize it. However, existing example selection studies ignore the ethical risks behind the examples selected, such as gender and race bias. In this work, we conduct extensive experiments and discover that (1) example selection with high accuracy does not mean low bias; (2) example selection for ICL may amplify the biases of LLMs; (3) example selection contributes to spurious correlations of LLMs. Based on the above observations, we propose the Remind with Bias-aware Embedding (ReBE), which removes the spurious correlations through contrastive learning and obtains bias-aware embedding for LLMs based on prompt tuning. Finally, we demonstrate that ReBE effectively mitigates biases of LLMs without significantly compromising accuracy and is highly compatible with existing example selection methods.

With the rapid deployment of Chinese large language models (LLMs), culturally-grounded bias evaluation remains understudied due to the dominance of English benchmarks and simplistic Chinese scenarios. To address this, we propose GeWu, a comprehensive benchmark featuring a culturally-aware dataset of 60,192 questions spanning 14 social groups with fine-grained Chinese contexts, significantly exceeding existing resources in breadth and depth. Our two-stage evaluation first quantifies bias via multiple-choice questions using a novel probability-based scoring mechanism to sensitively capture bias tendencies, distilling high-bias scenarios into GeWu-1K. This refined subset then enables multi-turn dialogue evaluations for in-depth analysis under realistic conditions. Experiments reveal that GeWu effectively exposes social biases in state-of-the-art Chinese LLMs, with 13.93% of scenarios eliciting universal bias across all models. This highlights persistent challenges and provides actionable insights for bias mitigation in Chinese contexts.

Yuxuan Li, Xinwei Guo, Jiashi Gao, Guanhua Chen, Xiangyu Zhao, Jiaxin Zhang, Quanying Liu, Haiyan Wu, Xin Yao, Xuetao Wei. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Text-to-Image (T2I) diffusion models exhibit concerning tendencies to generate harmful imagery that perpetuates social biases and stereotypes, posing significant ethical risks in real-world applications. While existing mitigation approaches predominantly employ black-box methodologies through dataset augmentation or constrained fine-tuning, they face critical limitations, including high data acquisition costs and potential exacerbation of stereotypes during model retraining. Inspired by neuroscience principles where neurological dysfunction often stems from aberrant neural activation patterns, we propose a novel framework, StereoClinic, targeting the root cause of stereotype generation through direct neural intervention. Our solution introduces two synergistic components: Diffusion Deep Taylor Decomposition (DDTD) for precisely localizing stereotype-related neurons via Layer-wise Relevance Propagation (LRP) attribution analysis, and Stereotype Neuron Suppression (SNS) implementing targeted activation damping to neutralize bias propagation. Through extensive empirical evaluations across multiple bias dimensions, we demonstrate that our method achieves significant stereotype mitigation without compromising image quality or requiring additional training data. This neuro-inspired approach establishes a new paradigm for model interpretability and ethical alignment in generative AI systems.

The online data markets have emerged as a valuable source of diverse datasets for training machine learning (ML) models. However, datasets from different data providers may exhibit varying levels of bias with respect to certain sensitive attributes in the population (such as race, sex, age, and marital status). Recent dataset acquisition research has focused on maximizing accuracy improvements for downstream model training, ignoring the negative impact of biases in the acquired datasets, which can lead to an unfair model. Can a consumer obtain an unbiased dataset from datasets with diverse biases? In this work, we propose a fairness-aware data acquisition framework (FAIRDA) to acquire high-quality datasets that maximize both accuracy and fairness for consumer local classifier training while remaining within a limited budget. Given the biases of data commodities remain opaque to consumers, the data acquisition in FAIRDA employs explore-exploit strategies. Based on whether exploration and exploitation are conducted sequentially or alternately, we introduce two algorithms: the knowledge-based offline data acquisition (KDA) and the reward-based online data acquisition algorithms (RDA). Each algorithm is tailored to specific customer needs, giving the former an advantage in computational efficiency and the latter an advantage in robustness. We conduct experiments to demonstrate the effectiveness of the proposed data acquisition framework in steering users toward fairer model training compared to existing baselines under varying market settings.

Federated learning (FL) stands as a paradigmatic approach that facilitates model training across heterogeneous and diverse datasets originating from various data providers. However, conventional FLs fall short of achieving consistent performance, potentially leading to performance degradation for clients who are disadvantaged in data resources. Influenced by the Matthew effect, deploying a performance-imbalanced global model in applications further impedes the generation of high-quality data from disadvantaged clients, exacerbating the disparities in data resources among clients. In this work, we propose anti-Matthew fairness for the global model at the client level, requiring equal accuracy and equal decision bias across clients. To balance the trade-off between achieving anti-Matthew fairness and performance optimality, we formalize the anti-Matthew effect federated learning (anti-Matthew FL) as a multi-constrained multi-objectives optimization (MCMOO) problem and propose a three-stage multi-gradient descent algorithm to obtain the Pareto optimality. We theoretically analyze the convergence and time complexity of our proposed algorithms. Additionally, through extensive experimentation, we demonstrate that our proposed anti-Matthew FL outperforms other state-of-the-art FL algorithms in achieving a high-performance global model while effectively bridging performance gaps among clients. We hope this work provides valuable insights into the manifestation of the Matthew effect in FL and other decentralized learning scenarios and can contribute to designing fairer learning mechanisms, ultimately fostering societal welfare.

Federated learning (FL) offers a machine learning paradigm that protects privacy, allowing multiple clients to collaboratively train a global model while only accessing their local data. Recent research in FL has increasingly focused on improving the uniformity of model performance across clients, a fairness principle known as egalitarian fairness. However, achieving egalitarian fairness in FL may sacrifice the model performance for data-rich clients to benefit those with less data. This trade-off raises concerns about the stability of FL, as data-rich clients may opt to leave the current coalition and join another that is more closely aligned with its expected high performance. In this context, our work rigorously addresses the critical concern: **Does egalitarian fairness lead to instability?** Drawing from game theory and social choice theory, we initially characterize fair FL systems as altruism coalition formation games (ACFGs) and reveal that the instability issues emerging from the pursuit of egalitarian fairness are significantly related to the clients’ altruism within the coalition and the configuration of the friends-relationship networks among the clients. Then, we theoretically propose the optimal egalitarian fairness bounds that an FL coalition can achieve while maintaining core stability under various types of altruistic behaviors. The theoretical contributions clarify the quantitative relationships between achievable egalitarian fairness and the disparities in the sizes of local datasets, disproving the misconception that egalitarian fairness inevitably leads to instability. Finally, we conduct experiments to evaluate the consistency of our theoretically derived egalitarian fairness bounds with the empirically achieved egalitarian fairness in fair FL settings.

Text-to-Image (T2I) has witnessed significant advancements, demonstrating superior performance for various generative tasks. However, the presence of stereotypes in T2I introduces harmful biases that require urgent attention as the T2I technology becomes more prominent. Previous work for stereotype mitigation mainly concentrated on mitigating stereotypes engendered with individual objects within images, which failed to address stereotypes engendered by the association of multiple objects, referred to as *Association-Engendered Stereotypes*. For example, mentioning ''black people'' and ''houses'' separately in prompts may not exhibit stereotypes. Nevertheless, when these two objects are associated in prompts, the association of ''black people'' with ''poorer houses'' becomes more pronounced. To tackle this issue, we propose a novel framework, MAS, to Mitigate Association-engendered Stereotypes. This framework models the stereotype problem as a probability distribution alignment problem, aiming to align the stereotype probability distribution of the generated image with the stereotype-free distribution. The MAS framework primarily consists of the *Prompt-Image-Stereotype CLIP* (*PIS CLIP*) and *Sensitive Transformer*. The *PIS CLIP* learns the association between prompts, images, and stereotypes, which can establish the mapping of prompts to stereotypes. The *Sensitive Transformer* produces the sensitive constraints, which guide the stereotyped image distribution to align with the stereotype-free probability distribution. Moreover, recognizing that existing metrics are insufficient for accurately evaluating association-engendered stereotypes, we propose a novel metric, *Stereotype-Distribution-Total-Variation*(*SDTV*), to evaluate stereotypes in T2I. Comprehensive experiments demonstrate that our framework effectively mitigates association-engendered stereotypes.

Large Language Models (LLMs) have demonstrated remarkable capabilities, surpassing human experts in various benchmark tests and playing a vital role in various industry sectors. Despite their effectiveness, a notable drawback of LLMs is their inconsistent moral behavior, which raises ethical concerns. This work delves into symmetric moral consistency in large language models and demonstrates that modern LLMs lack sufficient consistency ability in moral scenarios. Our extensive investigation of twelve popular LLMs reveals that their assessed consistency scores are influenced by position bias and selection bias rather than their intrinsic abilities. We propose a new framework tSMC, which gauges the effects of these biases and effectively mitigates the bias impact based on the Kullback–Leibler divergence to pinpoint LLMs' mitigated Symmetric Moral Consistency. We find that the ability of LLMs to maintain consistency varies across different moral scenarios. Specifically, LLMs show more consistency in scenarios with clear moral answers compared to those where no choice is morally perfect. The average consistency score of 12 LLMs ranges from $60.7\%$ in high-ambiguity moral scenarios to $84.8\%$ in low-ambiguity moral scenarios.

Most NFTs (Non-Fungible Tokens) use multi-hop URLs to address the off-chain assets due to the costly on-chain storage, but the path from NFTs to the underlying assets is fraught with instability, which may degrade its value. Hence, this paper aims to answer the question: Is the NFT-to-Asset connection fragile? This paper makes a first step towards this end by characterizing NFT-to-Asset connections of 12,353 Ethereum NFT Contracts (6,234,141 NFTs in total) from three perspectives, storage, accessibility, and duplication. In order to overcome challenges of affecting the measurement accuracy, e.g., IPFS instability and the changing availability of both IPFS and servers’ data, we propose to leverage multiple gateways to enlarge the data coverage and extend a longer measurement period with non-trivial efforts. Results of our extensive study show that such connection is very fragile in practice. The loss, unavailability, or duplication of off-chain assets could render the value of NFTs worthless. For instance, we find that assets of 25.24% of Ethereum NFT contracts are not accessible, and 21.48% of Ethereum NFT contracts include duplicated assets. Our work sheds light on the fragility along the NFT-to-Asset connection, which could help the NFT community to better enhance the trust of off-chain assets.

Traffic lights control could be regarded as a multi-agent coordinated problem. A model-free reinforcement learning (RL) approach is a powerful framework for solving such coordinated policy-making problems without prior environmental knowledge. In order to approach a global policy, communication among agents needs to be built. To enable dynamic and scalable communication, we propose a new RL model, CommNet based on Local Attention Mechanism (Attn-CommNet), which uses local selection and attention mechanism between hidden layers to facilitate cooperation. We evaluated the proposed method using synthetic and real word traffic flows under multi-scale road networks. The results demonstrate that the proposed method can get better performance in multi-scale problems, especially large-scale problems compared to the state-of-the-art methods.