New Method Improves Continuous Chain-of-Thought Training for LLMs

Continuous Chain-of-Thought (CoT) models aim to condense verbose reasoning processes into compact latent representations. Previous methods for training these models often relied on indirect supervision, which is computationally intensive and slow. A new approach, C-MTP, proposes a direct supervision method that models each latent representation as an average of the embeddings from the CoT traces being compressed. This direct supervision technique demonstrates improved efficiency and competitive performance against slower indirect methods when applied to simpler CoT tasks involving shorter reasoning sequences. However, the study also extended evaluation to more complex tasks with significantly longer reasoning traces. In these challenging scenarios, both direct and indirect supervision methods showed a substantial performance drop, highlighting a critical limitation in the current state of continuous CoT techniques for handling intricate, multi-step reasoning.