Interactive Thinking and Deep Reasoning in LLMs via Knowledge-Augmented Generation

View a PDF file from the paper entitled Kag-Hinker: Interactive Thinking and Deep Thinking in LLM

PDF HTML (experimental) view

a summary:In this paper, we offer Kag-Shinker, which upgrade KAG to multi-turn interactive thinking and a deep-thinking framework supported by a large language model dedicated to parameters (LLM). Our approach is to build an organized thinking process to solve complex problems, enhance logical cohesion and contextual consistency of the process of thinking about the tasks of answers (Q&A) on the rules of knowledge of the field (KBS) within LLMS. By \ Textbf {Form Logical} The KAG retrieval and thinking technology path first analyzes complex questions to independently solution sub -problems (which are also referred to as logical forms) through \ Textbf {Displesposition}. Each such logical model is represented in two models equivalent to the natural language and the logical function, after which it was classified either as a task for recovery of knowledge or thinking analysis. The dependency and the parameter that passes between these tasks are explicitly designed through logical functions. In the solution process, the recovery function performs retrieval tasks. It recalls organized and unorganized information from the specific knowledge unit. While mathematics and conclusion are used to perform thinking analysis tasks. Second, it should be noted that in the tasks of the sub -paralysis of knowledge, LLMS and the sources of external knowledge KBS are equivalent. We use the \ Textbf unit {the boundaries of knowledge} to determine the optimal source using self -regulating mechanisms such as calibration of confidence and reflective thinking, and the use of the \ Textbf unit {depth} to enhance the comprehensiveness of knowledge acquisition …