Modular programming, which aims to construct the final program by integrating smaller, independent building blocks, has been regarded as a desirable practice in software development. However, with the rise of recent code generation agents built upon large language models (LLMs), a question emerges: is this traditional practice equally effective for these new tools? In this work, we assess the impact of modularity in code generation by introducing a novel metric for its quantitative measurement. Surprisingly, unlike conventional wisdom on the topic, we find that modularity is not a core factor for improving the performance of code generation models. We also explore potential explanations for why LLMs do not exhibit a preference for modular code compared to non-modular code.

A second possibility is that contamination happens indirectly through the use Code generation has emerged as an important skill of synthetic data--a widespread paradigm used in for large language models to master. Measuring recent particular to increase code capabilities by generating progress in code generation has relied on few, additional code training tokens. Finally, we critical benchmarks to judge performance between argue that final model selection might have been model families and checkpoints. While many recent overly influenced by their performance on these sophisticated evaluation datasets have been datasets, overfitting to performance on these metrics proposed (Jain et al., 2024; Jimenez et al., 2024), over general-purpose code-oriented skills.

Recent developments show that Large Language Models (LLMs) produce state-of-the-art performance on natural language (NL) to code generation for resource-rich general-purpose languages like C++, Java, and Python. However, their practical usage for structured domain-specific languages (DSLs) such as YAML, JSON is limited due to domain-specific schema, grammar, and customizations generally unseen by LLMs during pre-training. Efforts have been made to mitigate this challenge via in-context learning through relevant examples or by fine-tuning. However, it suffers from problems, such as limited DSL samples and prompt sensitivity but enterprises maintain good documentation of the DSLs. Therefore, we propose DocCGen, a framework that can leverage such rich knowledge by breaking the NL-to-Code generation task for structured code languages into a two-step process. First, it detects the correct libraries using the library documentation that best matches the NL query. Then, it utilizes schema rules extracted from the documentation of these libraries to constrain the decoding. We evaluate our framework for two complex structured languages, Ansible YAML and Bash command, consisting of two settings: Out-of-domain (OOD) and In-domain (ID). Our extensive experiments show that DocCGen consistently improves different-sized language models across all six evaluation metrics, reducing syntactic and semantic errors in structured code. We plan to open-source the datasets and code to motivate research in constrained code generation.

Large Language Models have revolutionized code generation ability by converting natural language descriptions into executable code. However, generating complex code within real-world scenarios remains challenging due to intricate structures, subtle bugs, understanding of advanced data types, and lack of supplementary contents. To address these challenges, we introduce the CoCoST framework, which enhances complex code generation by online searching for more information with planned queries and correctness testing for code refinement. Moreover, CoCoST serializes the complex inputs and outputs to improve comprehension and generates test cases to ensure the adaptability for real-world applications. CoCoST is validated through rigorous experiments on the DS-1000 and ClassEval datasets. Experimental results show that CoCoST substantially improves the quality of complex code generation, highlighting its potential to enhance the practicality of LLMs in generating complex code.

Data curation is commonly considered a "secret-sauce" for LLM training, with higher quality data usually leading to better LLM performance. Given the scale of internet-scraped corpora, data pruning has become a larger and larger focus. Specifically, many have shown that de-duplicating data, or sub-selecting higher quality data, can lead to efficiency or performance improvements. Generally, three types of methods are used to filter internetscale corpora: embedding-based, heuristic-based, and classifier-based. In this work, we contrast the former two in the domain of finetuning LLMs for code generation. We find that embedding-based methods are often confounded by length, and that a simple heuristic-- pruning long files--outperforms other methods in compute-limited regimes. Our method can yield up to a 2x efficiency benefit in training (while matching performance) or a 3.5% absolute performance improvement on HumanEval (while matching compute). However, we find that perplexity on held-out long files can increase, begging the question of whether optimizing data mixtures for common coding benchmarks (HumanEval, MBPP) actually best serves downstream use cases. Overall, we hope our work builds useful intuitions about code data (specifically, the low quality of extremely long code files) provides a compelling heuristic-based method for data pruning, and brings to light questions in how we evaluate code generation models.

While language models (LMs) have proven remarkably adept at generating code, many programs are challenging for LMs to generate using their parametric knowledge alone. Providing external contexts such as library documentation can facilitate generating accurate and functional code. Despite the success of retrieval-augmented generation (RAG) in various text-oriented tasks, its potential for improving code generation remains under-explored. In this work, we conduct a systematic, large-scale analysis by asking: in what scenarios can retrieval benefit code generation models? and what challenges remain? We first curate a comprehensive evaluation benchmark, CodeRAG-Bench, encompassing three categories of code generation tasks, including basic programming, open-domain, and repository-level problems. We aggregate documents from five sources for models to retrieve contexts: competition solutions, online tutorials, library documentation, StackOverflow posts, and GitHub repositories. We examine top-performing models on CodeRAG-Bench by providing contexts retrieved from one or multiple sources. While notable gains are made in final code generation by retrieving high-quality contexts across various settings, our analysis reveals room for improvement -- current retrievers still struggle to fetch useful contexts especially with limited lexical overlap, and generators fail to improve with limited context lengths or abilities to integrate additional contexts. We hope CodeRAG-Bench serves as an effective testbed to encourage further development of advanced code-oriented RAG methods.

In the scenario-based evaluation of machine learning models, a key problem is how to construct test datasets that represent various scenarios. The methodology proposed in this paper is to construct a benchmark and attach metadata to each test case. Then a test system can be constructed with test morphisms that filter the test cases based on metadata to form a dataset. The paper demonstrates this methodology with large language models for code generation. A benchmark called ScenEval is constructed from problems in textbooks, an online tutorial website and Stack Overflow. Filtering by scenario is demonstrated and the test sets are used to evaluate ChatGPT for Java code generation. Our experiments found that the performance of ChatGPT decreases with the complexity of the coding task. It is weakest for advanced topics like multi-threading, data structure algorithms and recursive methods. The Java code generated by ChatGPT tends to be much shorter than reference solution in terms of number of lines, while it is more likely to be more complex in both cyclomatic and cognitive complexity metrics, if the generated code is correct. However, the generated code is more likely to be less complex than the reference solution if the code is incorrect.

Recently, large code generation models trained in a self-supervised manner on extensive unlabeled programming language data have achieved remarkable success. While these models acquire vast amounts of code knowledge, they perform poorly on code understanding tasks, such as code search and clone detection, as they are specifically trained for generation. Pre-training a larger encoder-only architecture model from scratch on massive code data can improve understanding performance. However, this approach is costly and time-consuming, making it suboptimal. In this paper, we pioneer the transfer of knowledge from pre-trained code generation models to code understanding tasks, significantly reducing training costs. We examine effective strategies for enabling decoder-only models to acquire robust code representations. Furthermore, we introduce CL4D, a contrastive learning method designed to enhance the representation capabilities of decoder-only models. Comprehensive experiments demonstrate that our approach achieves state-of-the-art performance in understanding tasks such as code search and clone detection. Our analysis shows that our method effectively reduces the distance between semantically identical samples in the representation space. These findings suggest the potential for unifying code understanding and generation tasks using a decoder-only structured model.

We introduce a new benchmark, ChartMimic, aimed at assessing the visually-grounded code generation capabilities of large multimodal models (LMMs). ChartMimic utilizes information-intensive visual charts and textual instructions as inputs, requiring LMMs to generate the corresponding code for chart rendering. ChartMimic includes 1,000 human-curated (figure, instruction, code) triplets, which represent the authentic chart use cases found in scientific papers across various domains(e.g., Physics, Computer Science, Economics, etc). These charts span 18 regular types and 4 advanced types, diversifying into 191 subcategories. Furthermore, we propose multi-level evaluation metrics to provide an automatic and thorough assessment of the output code and the rendered charts. Unlike existing code generation benchmarks, ChartMimic places emphasis on evaluating LMMs' capacity to harmonize a blend of cognitive capabilities, encompassing visual understanding, code generation, and cross-modal reasoning. The evaluation of 3 proprietary models and 11 open-weight models highlights the substantial challenges posed by ChartMimic. Even the advanced GPT-4V, Claude-3-opus only achieve an average score of 73.2 and 53.7, respectively, indicating significant room for improvement. We anticipate that ChartMimic will inspire the development of LMMs, advancing the pursuit of artificial general intelligence.

Large language models (LLMs) have demonstrated remarkable capabilities in code generation tasks. However, repository-level code generation presents unique challenges, particularly due to the need to utilize information spread across multiple files within a repository. Existing retrieval-based approaches sometimes fall short as they are limited in obtaining a broader and deeper repository context. In this paper, we present CatCoder, a novel code generation framework designed for statically typed programming languages. CatCoder enhances repository-level code generation by integrating relevant code and type context. Specifically, it leverages static analyzers to extract type dependencies and merges this information with retrieved code to create comprehensive prompts for LLMs. To evaluate the effectiveness of CatCoder, we adapt and construct benchmarks that include 199 Java tasks and 90 Rust tasks. The results show that CatCoder outperforms the RepoCoder baseline by up to 17.35%, in terms of pass@k score. Furthermore, the generalizability of CatCoder is assessed using various LLMs, including both code-specialized models and general-purpose models. Our findings indicate consistent performance improvements across all models, which underlines the practicality of CatCoder.