Code comprehension remains fundamental to software development. Although recent advances in deep learning and Code Large Language Models (CodeLLMs) have significantly improved method-level code summarization, class-level understanding remains challenging due to: (1) limited dedicated research focus, (2) excessive code sequence of a class, and (3) overreliance on source code and Syntax Tree (AST) while neglecting other valuable multimodal information. To address these challenges, we present a framework (called CSMS) with two key innovations. First, it pioneers method-level summary fusion, leveraging CodeT5 to extract semantically rich features from member methods without additional input. Second, its optimized Transformer architecture combines a parameter-efficient TTSEncoder with a multiperspective TAJDecoder, enabling comprehensive feature utilization while handling long sequences. Experimental results on ClassSum and HRCE datasets demonstrate CSMS’s superiority, achieving state-of-the-art BLEU scores while effectively handling long class sequences. The framework’s ability to leverage method-level context while maintaining computational efficiency represents a significant advance in class comprehension.

Code representation plays a fundamental role in enabling a wide range of code intelligence tasks, such as code translation, bug fixing, and code completion. Existing approaches typically adopt task-specific modeling paradigms, training separate models for individual downstream tasks. However, such methods often fail to capture semantic and structural commonalities across tasks, resulting in limited generalization and transferability. Therefore, we propose MTL-CR (Multitask Learning for Code Representation), a unified multitask learning framework that jointly optimizes multiple code-related tasks to learn more generalizable and robust code representations. To balance optimization dynamics among tasks, we further introduce a dynamic task weighting strategy based on taskspecific learning speeds. Experimental results demonstrate that MTL-CR consistently outperforms single task baselines and static multitask methods in various downstream tasks and evaluation metrics. For example, in the code translation task, it improves Accuracy and BLEU by 6.29% and 2.33%.