Fairness Mediator: Neutralize Stereotype Associations to Mitigate Bias in Large Language Models
Large Language Models (LLMs) have demonstrated remarkable performance across diverse applications, yet they inadvertently absorb spurious correlations from training data, leading to stereotype associations between biased concepts and specific social groups. These associations perpetuate and even amplify harmful social biases, raising significant concerns about fairness. To mitigate such biases, prior studies have attempted to project model embeddings into unbiased spaces during inference. However, these approaches have shown limited effectiveness due to their weak alignment with downstream social biases. Inspired by the observation that concept cognition in LLMs is primarily represented through a linear associative memory mechanism, where key-value mapping occurs in the MLP layers, we posited that biased concepts and social groups are similarly encoded as entity (key) and information (value) pairs, which can be manipulated to promote fairer associations. To this end, we propose Fairness Mediator (FairMed), an effective and efficient bias mitigation framework that neutralizes stereotype associations. Our framework comprises two main components: a stereotype association prober and an adversarial debiasing neutralizer. The prober captures stereotype associations encoded within MLP layer activations by employing prompts centered around biased concepts (keys) to detect the emission probabilities for social groups (values). Subsequently, the adversarial debiasing neutralizer intervenes in MLP activations during inference to equalize the association probabilities among different social groups. Extensive experiments across nine protected attributes demonstrate that our FairMed significantly outperforms state-of-the-art methods in effectiveness, achieving average bias reductions of up to 84.42% and 80.36% for $s_{\text{DIS}}$ and $s_{\text{AMB}}$, respectively. Compared to the most effective baseline, FairMed presents competitive efficiency by cutting mitigation overhead by hundreds of minutes. FairMed also maintains the LLM’s language understanding capabilities without compromising overall performance. Our codes can be found on our website.
Wed 25 JunDisplayed time zone: Amsterdam, Berlin, Bern, Rome, Stockholm, Vienna change
11:00 - 12:15 | Fairness and LLM TestingResearch Papers at Cosmos 3A Chair(s): Andreas Metzger University of Duisburg-Essen | ||
11:00 25mTalk | Fairness Mediator: Neutralize Stereotype Associations to Mitigate Bias in Large Language Models Research Papers Yisong Xiao Beihang University, Aishan Liu Beihang University; Institute of Dataspace, Siyuan Liang National University of Singapore, Xianglong Liu Beihang University; Institute of Dataspace; Zhongguancun Laboratory, Dacheng Tao Nanyang Technological University DOI | ||
11:25 25mTalk | ClassEval-T: Evaluating Large Language Models in Class-Level Code Translation Research Papers Pengyu Xue Shandong University, Linhao Wu Shandong University, Zhen Yang Shandong University, Chengyi Wang Shandong University, Xiang Li Shandong University, Yuxiang Zhang Shandong University, Jia Li Tsinghua University, Ruikai Jin Shandong University, Yifei Pei Shandong University, Zhaoyan Shen Shandong University, Xiran Lyu Shandong University, Jacky Keung City University of Hong Kong DOI | ||
11:50 25mTalk | No Bias Left Behind: Fairness Testing for Deep Recommender Systems Targeting General Disadvantaged Groups Research Papers Zhuo Wu Tianjin International Engineering Institute, Tianjin University, Zan Wang Tianjin University, Chuan Luo Beihang University, Xiaoning Du Monash University, Junjie Chen Tianjin University DOI |
Cosmos 3A is the first room in the Cosmos 3 wing.
When facing the main Cosmos Hall, access to the Cosmos 3 wing is on the left, close to the stairs. The area is accessed through a large door with the number “3”, which will stay open during the event.