<ul><li>Large language models (LLMs) trained with chain-of-thought (CoT) supervision have shown remarkable reasoning capabilities.</li><li>A new method called EPiC has been introduced to condense CoT traces for resource-efficient reasoning training.</li><li>EPiC selectively retains problem understanding and solution convergence stages in the reasoning trace, reducing training time by over 34% without compromising reasoning accuracy.</li><li>This approach aims at achieving lossless reasoning supervision while enhancing efficiency in training reasoning models.</li></ul>

EPiC: Towards Lossless Speedup for Reasoning Training through Edge-Preserving CoT Condensation

Discover more