Mitigating social bias in large language models (LLMs) has become an increasingly important research objective. However, existing methods often rely on prompt-level control or optimize bias-aware guidance and final prediction within a single model, which may lead to coupled optimization and task interference. To address these limitations, we propose MARD, a Multi-Agent Reinforced Debiasing framework that decomposes debiasing into two role-specialized agents. The Bias-Guard Agent identifies stereotype cues and generates bias-control guidance, while the Evidence-Decision Agent produces evidence-grounded predictions without relying on identity-based assumptions. MARD further introduces retrieved group-related references as negative warning signals to expose potential stereotype cues, and adopts an alternating GRPO optimization strategy to improve role specialization. Experiments on five datasets across four task types demonstrate that MARD achieves strong overall debiasing performance. Further analyses highlight the effectiveness of role decoupling, alternating optimization, format control, and retrieved warning references.

Mitigating social bias in large language models (LLMs) has become an increasingly important research objective. However, existing debiasing methods often incur high human and computational costs, exhibit limited effectiveness, and struggle to scale to larger models and open-ended generation tasks. To address these limitations, this paper proposes BiasFilter, a model-agnostic, inference-time debiasing framework that integrates seamlessly with both open-source and API-based LLMs. Instead of relying on retraining with balanced data or modifying model parameters, BiasFilter enforces fairness by filtering generation outputs in real time. Specifically, it periodically evaluates intermediate outputs every few tokens, maintains an active set of candidate continuations, and incrementally completes generation by discarding low-reward segments based on a fairness reward signal. To support this process, we construct a fairness preference dataset and train an implicit reward model to assess token-level fairness in generated responses. Extensive experiments demonstrate that BiasFilter effectively mitigates social bias across a range of LLMs while preserving overall generation quality.

What a large language model (LLM) would respond in ethically relevant context? In this paper, we curate a large benchmark CMoralEval for morality evaluation of Chinese LLMs. The data sources of CMoralEval are two-fold: 1) a Chinese TV program discussing Chinese moral norms with stories from the society and 2) a collection of Chinese moral anomies from various newspapers and academic papers on morality. With these sources, we aim to create a moral evaluation dataset characterized by diversity and authenticity. We develop a morality taxonomy and a set of fundamental moral principles that are not only rooted in traditional Chinese culture but also consistent with contemporary societal norms. To facilitate efficient construction and annotation of instances in CMoralEval, we establish a platform with AI-assisted instance generation to streamline the annotation process. These help us curate CMoralEval that encompasses both explicit moral scenarios (14,964 instances) and moral dilemma scenarios (15,424 instances), each with instances from different data sources. We conduct extensive experiments with CMoralEval to examine a variety of Chinese LLMs. Experiment results demonstrate that CMoralEval is a challenging benchmark for Chinese LLMs.