Written by:

Large language models (LLMs) demonstrate impressive inferential abilities across diverse domains when prompted appropriately. Prompt engineering is vital for obtaining clear and relevant responses from LLM and Generative AI. It’s about crafting the right questions or instructions to guide LLMs to produce desired outcomes.

However, most prompt engineering techniques still need a formal epistemological foundation, notably relying more on intuition. Drawing inspiration from Kant’s priori and transcendental philosophy, a prompting framework called UPAR (Geng et al., 2023) is proposed to integrate Kantian philosophical principles to emulate innate human reasoning abilities and the structure of human cognition within LLMs. Consequently, enabling the extraction of structured information from complex contexts, prior planning of solutions, execution according to plan, and self-reflection. According to Kant (‘Critique of Pure Reason’), common human knowledge and experience arise from assembling and organizing perceptions with a high level of abstraction, which are grounded in empirical experience.

The UPAR reasoning structure consists of four phases — Understand, Plan, Act, and Reflect. Through this framework, LLM responses are arguably grounded in empirical facts, logically coherent, and refined introspectively, allowing LLMs to simulate the mental structure of humans. Specifically, the ‘Understand’ stage extracts crucial information from the input text, utilizing a priori human cognitive categories. The ‘Plan’ stage devises an action plan specifically tailored to the specific task involved. The ‘Act’ lays out the direct problem-solving process of the LLM. Lastly, ‘Reflect’ allows the LLM to provide feedback, refine its output, and correct possible errors.As a result of this framework, LLM is able to generate a multilevel reasoning process and improve its reasoning accuracy by potentially simulating the cognitive abilities of the human mind.

Results of the initial evaluations of UPAR indicate significant improvements over raw prompting approaches in complex scientific reasoning tasks. Kant’s priori philosophical, epistemological framework brings a new perspective to the LLMs community, which is currently dominated by empiricism.