Скачать или смотреть PyCapsule Efficient Self Debugging Code Generation

This research paper introduces PyCapsule, a novel framework for efficient and accurate Python code generation using large language models (LLMs). PyCapsule employs a two-agent architecture—a programmer agent and an executor agent—along with specialized modules for error handling, example call detection, and function signature conversion to improve code quality and reduce computational costs. Experiments across multiple datasets demonstrate PyCapsule's superior performance compared to existing multi-agent approaches, achieving significant improvements in success rates. The authors observe diminishing returns with increased self-debugging attempts, hypothesizing this is due to limitations in error feedback. PyCapsule offers a resource-efficient solution for advancing automated code generation in AI systems.

https://arxiv.org/pdf/2502.02928