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

JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 1 Human-Inspired Learning for Large Language Models via Obvious Record and Maximum-Entropy Method Discovery Hong Su Abstract—Large Language Models (LLMs) excel at extracting common patterns from large-scale corpora, yet they struggle with rare, low-resource, or previously unseen scenarios—such as niche hardware deployment issues or irregular IoT device behaviors—because such cases are sparsely represented in train- ing data. Moreover, LLMs rely primarily on implicit parametric memory, which limits their ability to explicitly acquire, recall, and refine methods, causing them to behave predominantly as intuition-driven predictors rather than deliberate, method- oriented learners. Inspired by how humans learn from rare experiences, this paper proposes a human-inspired learning framework that in- tegrates two complementary mechanisms. The first, Obvious Record, explicitly stores cause–result (or question–solution) rela- tionships as symbolic memory, enabling persistent learning even from single or infrequent encounters. The second, Maximum- Entropy Method Discovery, prioritizes and preserves methods with high semantic dissimilarity, allowing the system to capture diverse and underrepresented strategies that are typically overlooked by next-token prediction. Verification on a benchmark of 60 semantically diverse question–solution pairs demonstrates that the proposed entropy- guided approach achieves stronger coverage of unseen questions and significantly greater internal diversity than a random base- line, confirming its effectiveness in discovering more generalizable and human-inspired methods. Index Terms—Large Language Models; Human-Inspired Learning; Maximum-Entropy Method Discovery; Explicit Mem- ory (Obvious Record) I. INTRODUCTION Large Language Models (LLMs) have achieved sub- stantial progress across a wide range of reasoning, gen- eration, and problem-solving tasks [1]. Their training paradigm—predicting the next token over massive cor- pora—enables them to capture broad statistical regularities and to perform well on problems that are commonly represented in training data [2]. Despite these strengths, LLMs exhibit sig- nificant limitations when confronted with rare, low-resource, or previously unseen scenarios. Typical examples include niche hardware deployment issues (e.g., uncommon GPU models), atypical IoT device failures, or real-world system problems that lack sufficient textual docu- mentation online. For instance, mainstream software frame- works such as TensorFlow and PyTorch primarily provide H. Su is with the School of Computer Science, Chengdu University of Information Technology, Chengdu, China. E-mail: suguest@126.com. default support for widely used GPU hardware and standard operating systems, whereas newly released or less common GPUs—especially when combined with highly customized or non-mainstream operating systems—often require device- specific configurations and undocumented adaptations. Be- cause LLMs predominantly reflect commonly learned patterns, they are often ineffective when addressing such specialized GPU–OS combinations, forcing users to rely instead on tar- geted technical forums or vendor-specific documentation to obtain reliable solutions. As these cases are sparsely repre- sented in training corpora, LLM-generated responses tend to be incomplete, inaccurate, or overly generic. A key reason for this limitation lies in the nature of paramet- ric learning. In LLMs, knowledge is stored implicitly within weight matrices, and retrieval occurs through an intuition-like process in which the model selects high-probability continua- tions based on previously learned patterns. In contrast, human learners employ a dual mechanism: they rely on intuition for familiar situations while also maintaining explicit memory of specific cause–result relationships, which enables them to recall rare experiences and refine methods over time. Such explicit memory is essential for handling infrequent events that intuition alone cannot resolve, as illustrated in Fig. 1. Motivated by this gap, we propose a human-like learning framework that augments LLMs with two complementary capabilities: • Obvious Record - Explicit Memory — an explicit, symbolic, non-parametric memory for storing mappings of the form featurecause →featureresult. This mechanism enables the system to learn from single or rare encounters, preserve interpretable knowledge, and update methods when better solutions appear. • Maximum-Entropy Method Discovery — a mechanism for identifying and retaining methods that are most se- mantically different from existing knowledge. These high- entropy methods capture diverse perspectives and novel strategies that LLMs tend to overlook because they are not reinforced by next-token prediction. Together, these mechanisms form a dual-process learning model in which the LLM acts as an intuition engine while the Obvious Record and entropy-guided

This content is AI-processed based on open access ArXiv data.