Recent advances in Code Large Language Models (CodeLLMs) have primarily focused on open-ended code generation, often overlooking the crucial aspect of code understanding & reasoning. To bridge this gap, we introduce CodeMMLU, a comprehensive multiple-choice benchmark designed to evaluate the depth of software and code comprehension in LLMs. CodeMMLU includes nearly 20,000 questions spanning diverse domains, including code analysis, defect detection, and software engineering principles across multiple programming languages. Unlike traditional benchmarks that emphasize code generation, CodeMMLU assesses a model’s ability to reason about programs across a wide-range of tasks such as code repair, execution reasoning, and fill-in-the-blank challenges. Our extensive evaluation reveals that even state-of-the-art models struggle with CodeMMLU, highlighting significant gaps in comprehension beyond generation. By emphasizing the essential connection between code understanding and effective AI-assisted development, CodeMMLU provides a critical resource for advancing more reliable and capable coding assistants.

Large language models (LLMs) pretrained on vast source code have achieved prominent progress in code intelligence. However, existing code LLMs have two main limitations. First, they often adopt a specific architecture (encoder-only or decoder-only) or rely on a unified encoder-decoder network for different downstream tasks, lacking the flexibility to operate in the optimal architecture for a specific task. Secondly, they often employ a limited set of pretraining objectives which might not be relevant to some tasks and hence result in substantial performance degrade. To address these limitations, we propose “CodeT5+”, a family of encoder-decoder LLMs for code in which component modules can be flexibly combined to suit a wide range of code tasks. Such flexibility is enabled by our proposed mixture of pretraining objectives, which cover span denoising, contrastive learning, text-code matching, and causal LM pretraining tasks, on both unimodal and bimodal multilingual code corpora. Furthermore, we propose to initialize CodeT5+ with frozen off-the-shelf LLMs without training from scratch to efficiently scale up our models, and explore instruction-tuning to align with natural language instructions. We extensively evaluate CodeT5+ on over 20 code-related benchmarks in different settings, including zero-shot, finetuning, and instruction-tuning. We observe state-of-the-art (SoTA) performance on various code-related tasks, and our instruction-tuned CodeT5+ 16B achieves new SoTA results of 35.0% pass@1 and 54.5% pass@10 on the HumanEval code generation task against other open code LLMs, even surpassing the OpenAI code-cushman-001 model.

Existing approaches to improving the robustness of source code models concentrate on recognizing adversarial samples rather than valid samples that fall outside of a given distribution, which we refer to as out-of-distribution (OOD) samples. To this end, we propose to use an auxiliary dataset (out-of-distribution) such that, when trained together with the main dataset, they will enhance the model’s robustness. We adapt energy-bounded learning objective function to assign a higher score to in-distribution samples and a lower score to out-of-distribution samples in order to incorporate such out-of-distribution samples into the training process of source code models. In terms of OOD detection and adversarial samples detection, our evaluation results demonstrate a greater robustness for existing source code models to become more accurate at recognizing OOD data while being more resistant to adversarial attacks at the same time.

Despite the recent trend of developing and applying neural source code models to software engineering tasks, the quality of such models is insufficient for real-world use. This is because there could be noise in the source code corpora used to train such models. We adapt data-influence methods to detect such noises in this paper. Data-influence methods are used in machine learning to evaluate the similarity of a target sample to the correct samples in order to determine whether or not the target sample is noisy. Our evaluation results show that data-influence methods can identify noisy samples from neural code models in classification-based tasks. This approach will contribute to the larger vision of developing better neural source code models from a data-centric perspective, which is a key driver for developing useful source code models in practice.

Recently program learning techniques have been proposed to process source code based on syntactical structures (e.g., abstract syntax trees) and/or semantic information (e.g., dependency graphs). While graphs may be better than trees at capturing code semantics, constructing the graphs from code inputs through the semantic analysis of multiple viewpoints can lead to inaccurate noises for a specific software engineering task. Compared to graphs, syntax trees are more precisely defined on the grammar and easier to parse; unfortunately, previous tree-based learning techniques have not been able to learn semantic information from trees to achieve better accuracy than graph-based techniques. We have proposed a new learning technique, named TreeCaps, by fusing together capsule networks with tree-based convolutional neural networks to achieve a learning accuracy higher than some existing graph-based techniques while it is based only on trees. TreeCaps introduces novel variable-to-static routing algorithms into the capsule networks to compensate for the loss of previous routing algorithms. Aside from accuracy, we also find that TreeCaps is the most robust to withstand those semantic-preserving program transformations that change code syntax without modifying the semantics. Evaluated on a large number of Java and C/C++ programs, TreeCaps models outperform prior deep learning models of program source code, in terms of both accuracy and robustness for program comprehension tasks such as code functionality classification and function name prediction. Our implementation is publicly available at: https://github.com/bdqnghi/treecaps.

Learning code representations has found many uses in software engineering, such as code classification, code search, comment generation, and bug prediction, etc. Although representations of code in tokens, syntax trees, dependency graphs, paths in trees, or the combinations of their variants have been proposed, existing learning techniques have a major limitation that these models are often trained on datasets labeled for specific downstream tasks, and as such the code representations may not be suitable for other tasks. Even though some techniques generate representations from unlabeled code, they are far from being satisfactory when applied to the downstream tasks. To overcome the limitation, this paper proposes InferCode, which adapts the self-supervised learning idea from natural language processing to the abstract syntax trees (ASTs) of code. The novelty lies in the training of code representations by predicting subtrees automatically identified from the contexts of ASTs. With InferCode, subtrees in ASTs are treated as the labels for training the code representations without any human labelling effort or the overhead of expensive graph construction, and the trained representations are no longer tied to any specific downstream tasks or code units. We have trained an instance of InferCode model using Tree-Based Convolutional Neural Network (TBCNN) as the encoder of a large set of Java code. This pre-trained model can then be applied to downstream unsupervised tasks such as code clustering, code clone detection, cross-language code search, or be reused under a transfer learning scheme to continue training the model weights for supervised tasks such as code classification and method name prediction. Compared to prior techniques applied to the same downstream tasks, such as code2vec, code2seq, ASTNN, using our pre-trained InferCode model higher performance is achieved with a significant margin for most of the tasks, including those involving different programming languages. The implementation of InferCode and the trained embeddings are available at the link: https://github.com/bdqnghi/infercode.

We propose Corder, a self-supervised contrastive learning framework for source code model. Corder is designed to alleviate the need of labeled data for code retrieval and code summarization tasks. The pre-trained model of Corder can be used in two ways: (1) it can produce vector representation of code which can be applied to code retrieval tasks that do not have labeled data; (2) it can be used in a fine-tuning process for tasks that might still require label data such as code summarization. The key innovation is that we train the source code model by asking it to recognize similar and dissimilar code snippets through acontrastive learning objective. To do so, we use a set of semantic-preserving transformation operators to generate code snippets that are syntactically diverse but semantically equivalent. Through extensive experiments, we have shown that the code models pretrained by Corder substantially outperform the other baselines for code-to-code retrieval, text-to-code retrieval, and code-to-text summarization tasks.

Despite being adopted in software engineering tasks, deep neural networks are treated mostly as a black box due to the difficulty in interpreting how the networks infer the outputs from the inputs. To address this problem, we propose AutoFocus, an automated approach for rating and visualizing the importance of input elements based on their effects on the outputs of the networks. The approach is built on our hypotheses that (1) attention mechanisms incorporated into neural networks can generate discriminative scores for various input elements and (2) the discriminative scores reflect the effects of input elements on the outputs of the networks. This paper verifies the hypotheses by applying AutoFocus on the task of algorithm classification (i.e., given a program source code as input, determine the algorithm implemented by the program). AutoFocus identifies and perturbs code elements in a program systematically, and quantifies the effects of the perturbed elements on the network's classification results. Based on evaluation on more than 1000 programs for 10 different sorting algorithms, we observe that the attention scores are highly correlated to the effects of the perturbed code elements. Such a correlation provides a strong basis for the uses of attention scores to interpret the relations between code elements and the algorithm classification results of a neural network, and we believe that visualizing code elements in an input program ranked according to their attention scores can facilitate faster program comprehension with reduced code.

To save effort, developers often translate programs from one programming language to another, instead of implementing it from scratch. Translating application program interfaces (APIs) used in one language to functionally equivalent ones available in another language is an important aspect of program translation. Existing approaches facilitate the translation by automatically identifying the API mappings across programming languages. However, these approaches still require large amount of parallel corpora, ranging from pairs of APIs or code fragments that are functionally equivalent, to similar code comments.

Algorithm classification is to automatically identify the classes of a program based on the algorithm(s) and/or data structure(s) implemented in the program. It can be useful for various tasks, such as code reuse, code theft detection, and malware detection. Code similarity metrics, on the basis of features extracted from syntax and semantics, have been used to classify programs. Such features, however, often need manual selection effort and are specific to individual programming languages, limiting the classifiers to programs in the same language.To recognize the similarities and differences among algorithms implemented in different languages, this paper describes a framework of Bilateral Neural Networks (Bi-NN) that builds a neural network on top of two underlying sub-networks, each of which encodes syntax and semantics of code in one language. A whole Bi-NN can be trained with bilateral programs that implement the same algorithms and/or data structures in different languages and then be applied to recognize algorithm classes across languages.We have instantiated the framework with several kinds of token-, tree- and graph-based neural networks that encode and learn various kinds of information in code. We have applied the instances of the framework to a code corpus collected from GitHub containing thousands of Java and C++ programs implementing 50 different algorithms and data structures. Our evaluation results show that the use of Bi-NN indeed produces promising algorithm classification results both within one language and across languages, and the encoding of dependencies from code into the underlying neural networks helps improve algorithm classification accuracy further. In particular, our custom-built dependency trees with tree-based convolutional neural networks achieve the highest classification accuracy among the different instances of the framework that we have evaluated. Our study points to a possible future research direction to tailor bilateral and multilateral neural networks that encode more relevant semantics for code learning, mining and analysis tasks.