D
Dalek
Text
LLMs for Implementation Generation in Open-Ended Problem Solving
A survey and discussion mapping how large language models and their extensions - prompting, Reinforcement Learning, and Retrieval-Augmented Generation - can support problem-solving activities for open-ended problems (problem framing, exploring solving approaches, feature elaboration and combination, implementation assessment, and handling unexpected situations) that traditional automated implementation-generation methods cannot. It contrasts LLM capabilities against four categories of traditional approaches (high-level specifications, evolutionary algorithms, agent-based methods, and cognitive architectures) and identifies open research requirements such as controlling elaboration, robust cross-abstraction assessment, and knowledge memorization during learning.