COMPASS: Where Does Mental Execution in Language Model?

Large language models (LLM) are significantly increasing the performance of using open logic steps to solve complex problems. But most of these methods require predefined mental characterization. A new research shows that a much simple signal is enough for mathematical problems: ** whether the models are directly correct to the answers they provide**.

Basic findings of this approach called COMPASS are:

• Using the accuracy signal, it can be found in a secret direction (latent direction) that mental abilities live in the model
• Unlike previous methods, Chain-of-Thought prompt design or complex property engineering is not required
• This signal, learned from the model’s own right/burning response trials, leads to the solution of mathematical problems
• Sparse Autoencoders (SAE) offers an alternative, more accessible method to complex reasoning properties derived from

The research makes it simpler to understand that artificial intelligence systems provide accurate or incorrect results. This changes the point of view to interior mechanisms of models and offers a practical way map for future development.

This finding arises as an open question that can be corrected in mathematical reasoning, and can not be warned and adapted to wider problem areas.