Tag
Deep Interaction proposes a method for directly editing erroneous reasoning steps in chain-of-thought outputs from large language models, achieving over 25% improvement in correction success and 40% reduction in token usage.