Diverse Thinking Schemata Enhance LLM Reasoning and Robustness
This work introduces Diverse Schemata Policy Optimization (DiScO), a framework designed to improve the reasoning abilities of Large Language Models (LLMs) by increasing the diversity of their "thinking schemata," which encompass reasoning transitions and answer candidates. Experiments on mathematical reasoning benchmarks show DiScO surpasses standard methods in both accuracy and the model's resilience to erroneous solution paths, suggesting that scaling along the diversity dimension is a promising research direction for LLMs.