LLMs can write themselves notes to get better at reasoning (arxiv.org)

Recent research has demonstrated that Large Language Models (LLMs) can enhance their reasoning abilities by autonomously generating and utilizing "experiential abstractions"—concepts similar to the strategies and reminders humans develop from experience. This study examines whether LLMs can distill their problem-solving experiences into retrievable libraries of natural-language abstractions, which can be used during inference and reinforced through learning with augmented training prompts. Notably, the findings show that LLMs leveraging these self-extracted abstractions exhibit improved performance on mathematical and logical reasoning tasks, comparable to those derived from teacher-extracted abstractions. This development is significant for the AI/ML community, as it suggests that LLMs can learn and apply knowledge in ways that mimic human cognitive processes, pointing towards a more intuitive and flexible model of learning. By integrating this abstraction framework, researchers can potentially enhance the capability of LLMs across various datasets and models, paving the way for more advanced applications in fields requiring complex reasoning and problem-solving. This research highlights the promise of equipping AI with the ability to reflect on past experiences, which may lead to more effective and adaptable AI systems in the future.