The logical reasoning dilemma of LLM: a mapping deficit in representation acquisition (2024)Abstract Currently, the Large Language Model (LLM) based on the Generative Pre-trained Transformer architecture of the natural language processing neural network model has made breakthroughs in multilingual language processing tasks and performed similar language processing capabilities to the human, However, its performance in dealing with language tasks that involve the logical reasoning capability aspect shows significant shortcoming. In this paper, we argue that the premise of a subject having logical reasoning capability is that it has representation acquisition capability. Based on this, in order to assess whether LLM has the potential to solve the shortcomings of its logical reasoning capability in its subsequent development, this paper gives an interpretation grounded in the linguistic representational similarity analysis of the human and LLM. This interpretation reveals that although LLMs use similar representations to humans in processing multilingual language tasks, but they do not have the same representation acquisition capability as humans, and there is a fundamental difference in the process of representation acquisition between LLMs and humans. We refer to this difference as a mapping deficit in the representation acquisition of LLM, which can successfully explain why LLMs are successful in processing multilingual language tasks even though they do not have the same representation acquisition capability as humans and why LLM performance in linguistic tasks that involve logical reasoning capability shows significant shortcomings. This paper aims to make a preliminary work for enhancing the logical reasoning capability for LLM in the future, and we believe that if the mapping deficit in representation acquisition of LLM is solved, then The logical reasoning dilemma of LLM will also be able to find an effective solution. Do large language models have the basis of conceptual Capacities What we have before we have representation (2023) Abstract Distance between GPT-4 and human-level concepts (2023) Abstract GPT-4, one of the latest large language models, has achieved exceptional success in dealing with various language tasks at the human level. It is natural to ask whether the language processing ability shown by GPT-4 indicates that it understands the text it outputs. To solve this complex question, we need to first ask whether it has concepts. This paper tries to make a preliminary discussion on the two questions, the first is whether GPT-4 has human-level concepts, and the Second is the distance between GPT-4 and human-level concepts. This paper argues that although GPT -4 does not have concepts at the human level, it might have different concepts than the human concept that could explain its performance on various language task processing. This paper is divided into three parts for discussion. The first part will clarify the definition of the concept and analyze several features of the concept. The second part will analyze how GPT -4 works and how it performs in various language task processing. The third part will discuss whether the performance of GPT-4 on various language processing tasks can indicate it has human-level concepts and what GPT-4 needs to do to achieve the human-level concept. By answering the question of whether the GPT-4 has a human-level concepts, we can better evaluate the foundations of the GPT-4, and can also help us to better understand human intelligence ingredients by using machine intelligence in the future.