A Survey of Deep Learning Methods for Source Code Summarization

A Survey of Deep Learning Methods for Source Code Summarization

Valeev A.I. (Innopolis University, Innopolis, Russia)
Korshuk A.O. (Innopolis University, Innopolis, Russia)
Ivanov V.V. (Innopolis University, Innopolis, Russia)

Abstract

Deep neural models have become the state of the art for code summarization. Before the era of Large Language Models, many specialized models applied various methods to improve baseline sequence-to-sequence models, leveraging code structure and similarities between code snippets and summaries, and employing multi-task learning. This survey elicits the design building blocks of these methods and builds a taxonomy of them. It also describes how different models apply the same method or how a single model applies different methods. Furthermore, we discuss the current flaws of these models, potential research directions, and the applicability of large language models in code summarization, comparing them with the surveyed methods.

Keywords

code summarization; transformer; RNN encoder-decoder; deep learning.

Edition

Proceedings of the Institute for System Programming, vol. 38, issue 4, part 1, 2026, pp. 101-134

ISSN 2220-6426 (Online), ISSN 2079-8156 (Print).

DOI: 10.15514/ISPRAS-2026-38(4)-6

For citation

Valeev A.I., Korshuk A.O., Ivanov V.V. A Survey of Deep Learning Methods for Source Code Summarization. Proceedings of the Institute for System Programming, vol. 38, issue 4, part 1, 2026, pp. 101-134 DOI: 10.15514/ISPRAS-2026-38(4)-6.

Full text of the paper in pdf Back to the contents of the volume