Self-reflecting Large Language Models: A Hegelian Dialectical Approach

Iterative self-reflection (Shinn et al., 2023; Madaan et al., 2023) is another approach that has recently gained significant attention within the NLP community. This method involves models mimicking human behavior by reviewing and critiquing their own outputs, actions, or decision-making processes to improve their performance over time. While an LLM does not “reflect” in the way humans do, it can be programmed to evaluate the quality or appropriateness of its responses in certain contexts.

Self-reflection can be viewed as a form of “Self-dialectic.” Broadly speaking, “Dialectic” refers to any logical debate that involves considering opposing views to presented propositions and using contradictions to uncover the truth and validity of statements made in the debate (Cambridge University Press, n.). Typically used in philosophy, the meaning of “Dialectic” beyond this general description varies depending on the philosophical tradition employing it (Bobzien & Duncombe, 2023; Maybee, 2020). The Hegelian Dialectic, for example, refers to the method proposed by Hegel in the 19th century, which iteratively synthesizes new theses and antitheses, driving the progression of ideas in the discussion (Hegel, 1807; 1951; Maybee, 2020).

Examining NLP from a philosophical perspective has recently fascinated researchers, as it connects computational methods with traditional philosophical methodologies (Milliere & Buckner, 2023; Milliere & Buckner, 2024). In this work, we aim to propose a philosophical approach inspired by Hegelian dialectics to foster LLMs’ self-reflection and explore whether a model is capable of generating new ideas using this method. Specifically, we introduce a self-debating methodology to evoke novel answers from an LLM through a “self-reflection” process, adhering to an organic dialectical discussion in the Hegelian style. Furthermore, we explore the effect of LLMs’ temperature