Can I understand what I create? Self-Knowledge Evaluation of Large Language Models
Zhiquan Tan, Lai Wei, Jindong Wang, Xing Xie, Weiran Huang
TL;DR
The paper addresses the difficulty of evaluating whether large language models truly understand their own outputs. It proposes a self-knowledge framework based on a two-step “first generate, then evaluate” process, yielding a self-knowledge score that captures consistency between generated answers and their verification. Extensive experiments across seven LLMs and nine tasks, plus two LMMs, reveal persistent gaps in self-knowledge, with misalignment to human attention and varying sensitivity to prompts and protocols. The findings suggest that fine-tuning on self-generated data can improve certain tasks (e.g., GSM-8k), and that expert prompting can boost self-knowledge, offering a scalable, practical path toward more reflective and reliable models.
Abstract
Large language models (LLMs) have achieved remarkable progress in linguistic tasks, necessitating robust evaluation frameworks to understand their capabilities and limitations. Inspired by Feynman's principle of understanding through creation, we introduce a self-knowledge evaluation framework that is easy to implement, evaluating models on their ability to comprehend and respond to self-generated questions. Our findings, based on testing multiple models across diverse tasks, reveal significant gaps in the model's self-knowledge ability. Further analysis indicates these gaps may be due to misalignment with human attention mechanisms. Additionally, fine-tuning on self-generated math task may enhance the model's math performance, highlighting the potential of the framework for efficient and insightful model evaluation and may also contribute to the improvement of LLMs.