The field of AI research is moving towards a greater emphasis on mitigating social bias in large language models (LLMs) and vision-language models (VLMs). This is driven by the need to develop more inclusive and fair AI systems that can accommodate diverse global populations. Recent studies have highlighted the English-centric nature of the field and the significant language gap in LLM safety research. To address these challenges, researchers are exploring new representational formats, such as thick description, and developing model-agnostic debiasing frameworks. These frameworks aim to filter generation outputs in real-time and enforce fairness by discarding low-reward segments based on a fairness reward signal. Furthermore, researchers are investigating the vulnerability of distilled models to adversarial injection of biased content and proposing practical design principles for building effective adversarial bias mitigation strategies. Noteworthy papers in this regard include: BiasFilter, which proposes a model-agnostic, inference-time debiasing framework that integrates seamlessly with both open-source and API-based LLMs. Dissecting Bias in LLMs, which adopts a mechanistic interpretability approach to analyze how social biases are structurally represented within models and demonstrates that bias-related computations are highly localized, often concentrated in a small subset of layers.