This paper explores the Leap-of-Thought (LoT) abilities within large language models (LLMs) - a non-sequential, creative paradigm involving strong associations and knowledge leaps. The authors first build a multimodal and multilingual Oogiri-GO dataset, which contains over 130,000 samples from the Oogiri game, a traditional Japanese creative game that challenges participants to provide unexpected and humorous responses to prompts in the form of images, text, or a combination of both.
Through extensive experiments, the authors find that even advanced LLMs and reasoning frameworks struggle to exhibit sufficient LoT ability for creative humor generation. To address this, they propose the Creative Leap-of-Thought (CLoT) paradigm, which consists of two stages:
Associable Instruction Tuning: This stage formulates the Oogiri-GO dataset into LoT-oriented instruction data to train the LLM to improve its LoT-based humor generation and discrimination abilities.
Explorative Self-Refinement: This stage encourages the LLM to generate more creative LoT data by exploring parallels between seemingly unrelated concepts and selects high-quality data to train itself for self-refinement.
The experimental results show that CLoT can greatly enhance the LoT ability of LLMs like Qwen and CogVLM across several types of Oogiri games. CLoT-integrated LLMs also achieve higher quantitative performance than the corresponding vanilla and CoT-integrated LLMs. Additionally, CLoT can boost creative abilities on other tasks like "cloud guessing game" and "divergent association task", demonstrating its remarkable generalization ability.
These findings advance the understanding of LoT in LLMs and offer a pathway to improve their creative capacities for innovative applications across domains.
To Another Language
from source content
arxiv.org
Principais Insights Extraídos De
by Shanshan Zho... às arxiv.org 04-23-2024
https://arxiv.org/pdf/2312.02439.pdfPerguntas Mais Profundas