Recent research highlights significant advancements in the reasoning capabilities of large language models (LLMs). A study by Google DeepMind indicates that LLMs can effectively reason without explicit prompting by exploring multiple paths during decoding, enhancing their natural reasoning abilities. Another study reveals that procedural knowledge from pretraining data drives LLM reasoning, suggesting that LLMs are not merely retrieving answers but employing generalizable strategies. This is supported by findings from a collaborative study involving millions of pretraining data points, which shows that LLMs rely more on procedural knowledge than on retrieval for reasoning tasks. Additionally, a two-stage retrieval-reasoning approach has been proposed to maintain accuracy as context lengths increase, emphasizing the importance of gathering facts before reasoning about them. Overall, these developments point to the potential for LLMs to serve as effective reasoning agents in various applications.