Researchers identified the operation of the Astra model without a chain of reasoning

Analysis of the work of Astra has shown that the model is capable of performing tasks with minimal use of tokenized reasoning (CoT). During testing, it was found that at a low level of query complexity, the system significantly reduces the volume of intermediate calculations compared to previous versions.

The results of observations indicate that the model applies internal logic rules without the need to generate extended text chains. In the course of the experiment, the model demonstrated a result of 7 points when performing a series of tasks in 2 hours. This confirms the ability of the architecture to directly infer answers when the task does not require deep decomposition.

The use of no-CoT mode allows the model to save computational resources and reduce the response generation time. The absence of the need for a "chain of thought" for simple operations reduces the overall load on inference, which is critical for high-load systems. Researchers note that such behavior indicates a more efficient internal representation of knowledge.

This study provides an understanding of how modern LLMs optimize the information processing process. The transition to direct problem-solving without redundant reasoning tokens is an important step in the development of architectures focused on speed and reducing the cost of a single query.