The field of large language models (LLMs) is rapidly advancing, with a growing focus on mitigating bias and ensuring fairness. Recent research has highlighted the importance of auditing LLMs for political bias, with studies demonstrating that these models can exhibit significant ideological biases. To address this issue, researchers are exploring new methods for decentralizing LLM alignment, including the use of context, pluralism, and participation. Additionally, there is a growing recognition of the need to assess and mitigate bias in LLMs, particularly in regards to demographic variables such as gender, age, and background. Several studies have proposed novel methods for bias assessment and mitigation, including the use of cross-lingual analysis and prompt-instructed mitigation strategies. Furthermore, researchers are also exploring the application of LLMs in new domains, such as urban policy-intelligence and historical structural oppression measurement. Noteworthy papers in this area include the proposal of ButterflyQuant, a novel method for ultra-low-bit LLM quantization that achieves state-of-the-art results while minimizing performance loss, and Fair-GPTQ, a bias-aware quantization method that reduces unfairness in large language models. Overall, the field of LLMs is moving towards a greater emphasis on fairness, transparency, and accountability, with significant implications for the responsible deployment of AI systems.
Mitigating Bias and Advancing Fairness in Large Language Models
Sources
POW: Political Overton Windows of Large Language Models
Decentralising LLM Alignment: A Case for Context, Pluralism, and Participation
ButterflyQuant: Ultra-low-bit LLM Quantization through Learnable Orthogonal Butterfly Transforms
Discrimination by LLMs: Cross-lingual Bias Assessment and Mitigation in Decision-Making and Summarisation
Phi: Preference Hijacking in Multi-modal Large Language Models at Inference Time
Don't Change My View: Ideological Bias Auditing in Large Language Models
Advancing Conversational AI with Shona Slang: A Dataset and Hybrid Model for Digital Inclusion
Simulating a Bias Mitigation Scenario in Large Language Models
From Pixels to Urban Policy-Intelligence: Recovering Legacy Effects of Redlining with a Multimodal LLM
Fair-GPTQ: Bias-Aware Quantization for Large Language Models
Assessing Historical Structural Oppression Worldwide via Rule-Guided Prompting of Large Language Models