[2N6-GS-2x] Machine learning

5:45 PM - 6:00 PM JST(8:45 AM - 9:00 AM UTC)

[2N6-GS-2x-02] large Learning Deviation from Common Sense: Creative Generation Model using Counter-Intuitive Chain of Thought

〇Daisuke Niino¹, Yoshia Abe² (1. DENTSU INC., 2. AI Center, The Univ. of Tokyo)

[[online]]

Keywords:

Fine-tuning,Deviation from Common Sense,knowledge modeling

Large Language Models (LLMs) have achieved high fluency and logical reasoning capabilities through probabilistic next-token prediction based on vast amounts of text data. However, their learning objective, Maximum Likelihood Estimation, tends to converge towards statistically frequent "average solutions." This characteristic, while advantageous for tasks with unique correct answers, becomes a fatal constraint in creative tasks such as advertising copy generation, product planning, and entertainment, where deviation from existing contexts (Surprise) and non-continuous leaps are essential. In this study, we propose a new Supervised Fine-Tuning (SFT) method called "Counter-Intuitive Chain of Thought (CI-CoT)" to model the non-obvious ideation processes of skilled creators. Specifically, we created a dataset that explicitly negates common reasoning paths and enforces rhetorical thinking such as paradox, exaggeration, and combination, and trained the model using OpenAI's Fine-tuning API. Evaluation experiments using unique Japanese advertising examples (Local Creative) showed that the proposed method generates highly novel ideas that are distinct from baseline models while maintaining logical coherence. This paper argues that creativity in LLMs can be acquired not through randomness, but through learning the intentional "process of breaking common sense."