Advances in Mitigating Bias in Large Language Models

The field of large language models (LLMs) is witnessing significant developments in addressing bias and improving their performance in simulating public opinions, analyzing climate policy, and evaluating geopolitical and cultural bias. Researchers are proposing new methods to evaluate and mitigate bias in LLMs, such as using human survey data as in-context examples, constructing manually curated datasets, and developing frameworks for evaluating LLM behavior across neutral and sensitive topics. Additionally, studies are investigating the ability of LLMs to answer complex policy questions, with promising results in analyzing climate policy documents. Noteworthy papers include: A Dual-Layered Evaluation of Geopolitical and Cultural Bias in LLMs, which offers a structured framework for evaluating LLM behavior. IMPACT: Inflectional Morphology Probes Across Complex Typologies, which introduces a synthetically generated evaluation framework to assess LLM performance in inflectional morphology. MPF: Aligning and Debiasing Language Models post Deployment via Multi Perspective Fusion, which presents a novel post-training alignment framework for LLMs.

Sources

Is DeepSeek a New Voice Among LLMs in Public Opinion Simulation?

A Dual-Layered Evaluation of Geopolitical and Cultural Bias in LLMs

Leveraging In-Context Learning for Political Bias Testing of LLMs

Computational Analysis of Climate Policy

A Detailed Factor Analysis for the Political Compass Test: Navigating Ideologies of Large Language Models

Perspective Dial: Measuring Perspective of Text and Guiding LLM Outputs

Positional Bias in Binary Question Answering: How Uncertainty Shapes Model Preferences

IMPACT: Inflectional Morphology Probes Across Complex Typologies

Designing an Adaptive Storytelling Platform to Promote Civic Education in Politically Polarized Learning Environments

Beyond Sociodemographic Prompting: Using Supervision to Align LLMs with Human Response Distributions

MPF: Aligning and Debiasing Language Models post Deployment via Multi Perspective Fusion

Built with on top of