Human-Inspired Learning

Human-Inspired Learning for Large Language Models via Obvious Record and Maximum-Entropy Method Discovery

This paper introduces a human-inspired learning framework that augments Large Language Models (LLMs) with explicit symbolic memory and entropy-guided method discovery. The proposed Obvious Record mechanism provides a structured non-parametric memory for storing and refining cause-result relationships, allowing the system to learn effectively even from single or rare encounters. Complementing this, the Maximum-Entropy Method Discovery mechanism identifies and preserves semantically diverse methods that are often underrepresented in traditional LLM training. Together, these components form a dual-process learning architecture that more closely mirrors human experience: routine reasoning is supported by similarity-based retrieval, while novel or difficult problems trigger high-entropy exploration and explicit method acquisition. Verification experiments on the QSS60 benchmark demonstrate that entropy-guided selection consistently yields superior semantic coverage and greater internal diversity compared with a random baseline, confirming the effectiveness of the proposed approach. This framework enables LLMs to better handle rare scenarios, reduce over-reliance on common patterns, and learn in a manner closer to human experience.