Whether Large Language Models Learn at the Inference Stage? Effects of Active Learning and Labelling with LLMs on their Reasoning

In-context learning (zero, one and few-shot learning), chain-of-thoughts and prompt engineering have been more and more widely researched and discussed recently, in connection with the discovered possibility of Large Language Models (LLMs) based on examples (and in particular demonstrations of instructions, some actions with examples, including demonstration of reasoning) to solve certain types of problems at the Inference stage, without the need for fine-tuning, and in general without any additional training of LLMs on these tasks. In our work, due to the lack of a general term, we refer to this emergent ability, observed in LLMs, as the L&R effect (Learning and Reasoning at the Inference Stage effect). This effect, along with the opportunities it creates, leaves open the question of what exactly is the reason for the emergence of such a surprising ability in LLMs. In this paper we formulate a hypothesis, consistent with both the literature and our experiments, that explains the cause of the L&R effect and answers the main question: Do LLMs learn at the Inference or not?