Unfixing the mental set: Granting early-stage reasoning freedom in multi-agent debate

Jing Wu

Abstract: Large language models (LLMs) have demonstrated remarkable performance across a wide range of tasks in recent years. While prior work has explored leveraging LLMs to generate synthetic data for self-improvement, repeated iterations often suffer from diminishing returns due to the reliance on homogeneous reasoning patterns and limited exploration of alternative perspectives. In this paper, we introduce a novel framework that enriches the reasoning process by encouraging critical thinking among multiple agents. Rather than deploying an ensemble of models with identical prompts, we propose a strategy generator that produces customized instructions tailored to each individual LLM. Acting as a critical thinking agent, the generator is iteratively fine-tuned using carefully selected strategies that are both diverse and effective. This approach fosters specialization within each model while promoting diversity across reasoning paths, enabling the system to maintain varied solution trajectories and achieve sustained performance gains through iterative refinement. We demonstrate the effectiveness of our method across a variety of agentic frameworks and complex reasoning tasks.