Hesam Sheikh Hassani
• Mehregan Nazarmohsenifakori
• Safoura Banihashemi
abstract
Our research proposal aims to explore the potential of mitigating social biases in Large Language Models (LLMs)
by combining multi-agent frameworks and reflective reasoning tools. Specifically, we intend to leverage CAMEL’s
multi-agent framework to enable multiple agents with distinct roles to collaboratively critique and refine outputs,
along with Anthropic’s ”think” tool which invokes intermediate reasoning steps. By measuring bias reduction
compared to traditional baselines, this project aims to show that agent collaboration and reflective self-correction
can significantly boost the fairness and reliability of AI-driven systems
outcomes