arXiv 2502.21074

CODI: Compressing Chain-of-Thought into Continuous Space via Self-Distillation

By Zhenyi Shen, Hanqi Yan, et al.

Published 2025-02-28

Mindmap

Browse the paper's core ideas, clusters, and relationships in a structured outline.

Chain-of-Thought (CoT) reasoning enhances Large Language Models (LLMs) by encouraging step-by-step reasoning in natural language. However, leveraging a latent continuous space for reasoning may offer benefits in terms of both efficiency and robustness. Prior implicit CoT methods attempt to bypass language completely by reasoning in continuous space but have consistently underperformed compared to the standard explic…

View the original paper on arXiv