A natural language programming system capable of dynamically changing the level of abstraction
This system proposes a system to address the challenge of "Abstraction Matching"—the difficulty of mapping abstract natural language commands to concrete code—in LLM-based programming. While existing methods typically provide only a single level of concretization, this system allows users to iteratively select and adjust the level of abstraction to match their specific intent and expertise.
The system decomposes natural language input into "blocks" based on semantic units, which users can then click to progressively expand into more detailed, concrete descriptions. Furthermore, by linking specific parameters to global variables and providing GUI elements like sliders for direct manipulation, the system enhances the user's sense of control over the resulting program.