CoFEE: Cognitive Reasoning Control for LLM Feature Discovery

The Cognitive Feature Engineering Engine (CoFEE) is an innovative technique that enhances reasoning control in Large Language Models (LLMs) to better extract features from intricate unstructured data. This method tackles the core reasoning challenge of recognizing predictive abstractions while preventing leakage, proxies, and signals that occur post-outcome. By imposing cognitive behaviors as structured inductive biases on the features proposed by LLMs, CoFEE seeks to yield more robust features compared to unrestricted generation. This research has been shared in the arXiv preprint 2604.21584v1.