While Code Language Models (CLMs) have demonstrated superior performance in software engineering tasks such as code generation and summarization, recent empirical studies reveal a critical privacy vulnerability: these models exhibit unintended memorization of sensitive training data, enabling verbatim reproduction of confidential information when specifically prompted. To address this issue, several approaches, including training data de-duplication and differential privacy augmentation, have been proposed. However, these methods require full-model retraining for deployed CLMs, which incurs substantial computational costs. In this paper, we aim to answer the following research question: Can sensitive information memorized by CLMs be erased effectively and efficiently? We conduct a pioneering investigation into erasing sensitive memorization in CLMs through machine unlearning - a post-hoc modification method that removes specific information from trained models without requiring full retraining. Specifically, we first quantify the memorization risks of sensitive data within CLM training datasets and curate a high-risk dataset of 50,000 sensitive memorized samples as unlearning targets. We study two widely used gradient ascent-based unlearning approaches: the vanilla and constraint-based methods, and introduce CodeEraser, an advanced variant that selectively unlearns sensitive memorized segments in code while preserving the structural integrity and functional correctness of the surrounding code. Extensive experiments on three families of CLMs, i.e., CodeParrot, CodeGen-Mono, and Qwen2.5-Coder, validate the effectiveness and efficiency of CodeEraser in erasing targeted sensitive memorization while maintaining model utility.

Code language models (so-called CodeLLMs) are now commonplace in software development. As a general rule, CodeLLMs are trained by dividing training examples into input tokens and then learn importance of those tokens in a process called machine attention. Machine attention is based solely on input token salience to output token examples during training. Human software developers are different, as humans intuitively know that some tokens are more salient than others. While intuition itself is ineffable and a subject of philosophy, clues about salience are present in human visual attention, since people tend to look at more salient words more often. In this paper, we present EyeMulator, a technique for training CodeLLMs to mimic human visual attention while training for various software development tasks. We add special weights for each token in each input example to the loss function used during LLM fine-tuning. We draw these weights from observations of human visual attention derived from a previously-collected publicly-available dataset of eye-tracking experiments in software engineering tasks. These new weights ultimately induce changes in the attention of the subject LLM during training, resulting in a model that does not need eye-tracking data during inference. Our evaluation shows that EyeMulator outperforms strong LLM baselines on several tasks such as code translation, completion and summarization. We further show an ablation study that demonstrates the improvement is due to subject models learning to mimic human attention.

Code language models have emerged as useful tools for various programming tasks, yet they often struggle when it comes to complex ones. In this paper, we explore the potential of curriculum learning in enhancing the performance of these models. While prior research has suggested that curriculum learning does not necessarily help in improving the performance of language models, our results surprisingly show that this may not be the case for code language models. We demonstrate that a well-designed curriculum learning approach significantly improves the accuracy of small decoder-only code language models on the task of code execution, while its effect on code completion is less significant. To explore the potential of curriculum learning, we train multiple GPT models with 1 million parameters each to predict the next token and evaluate them on code completion and execution tasks. Our contributions include proposing a novel code difficulty assessment metric by combining software code measures, investigating the effectiveness of Curriculum Learning for code language models, and introducing a Novel Curriculum Learning schedule that enhances the performance of small decoder-only language models in code execution tasks. The results of this paper open the door for more research on the use of curriculum learning for code language models.

In the context of the rising interest in code language models (code LMs) and vulnerability detection, we study the effectiveness of code LMs for detecting vulnerabilities. Our analysis reveals significant shortcomings in existing vulnerability datasets, including poor data quality, low label accuracy, and high duplication rates, leading to unreliable model performance in realistic vulnerability detection scenarios. Additionally, the evaluation methods used with these datasets are not representative of real-world vulnerability detection. To address these challenges, we introduce PrimeVul, a new dataset for training and evaluating code LMs for vulnerability detection. PrimeVul incorporates a novel set of data labeling techniques that achieve comparable label accuracy to human-verified benchmarks while significantly expanding the dataset. It also implements a rigorous data de-duplication and chronological data splitting strategy to mitigate data leakage issues, alongside introducing more realistic evaluation metrics and settings. This comprehensive approach aims to provide a more accurate assessment of code LMs' performance in real-world conditions. Evaluating code LMs on PrimeVul reveals that existing benchmarks significantly overestimate the performance of these models. For instance, a state-of-the-art 7B model scored 68.26% F1 on BigVul but only 3.09% F1 on PrimeVul. Attempts to improve performance through advanced training techniques and larger models like GPT-3.5 and GPT-4 were unsuccessful, with results akin to random guessing in the most stringent settings. These findings underscore the considerable gap between current capabilities and the practical requirements for deploying code LMs in security roles, highlighting the need for more innovative research in this domain.

Instruction-finetuned code language models (LMs) have shown promise in various programming tasks. They are trained, using a language modeling objective, on natural language instructions and gold code snippet pairs. Recent evidence suggests that these models, never exposed to incorrect solutions during training, often struggle to distinguish between correct and incorrect solutions. This observation raises our inquiry: Can preference learning, which trains models to prefer correct solutions over incorrect ones, help push the boundaries of code LMs even further? We propose PLUM, a novel \textbf{p}reference \textbf{l}earning framework a\textbf{u}gmented with test cases tailored for code L\textbf{M}s.PLUM aims to investigate the key success factors and potential benefits of preference learning in code LMs, which remain elusive despite its success in aligning LMs with human values. PLUM consists of three stages: (1) Generating test cases for natural language instructions, (2) sampling candidate solutions from the policy and evaluating them against the test cases to create a preference dataset, which is then used to (3) train the policy with a preference learning algorithm. Experiments demonstrate that PLUM substantially improves the performance of existing code LMs on established code generation benchmarks such as HumanEval (+) and MBPP (+), even for the state-of-the-art open-source language model CodeQwen-1.5-7B-Chat. PLUM complements the supervised fine-tuning (SFT) stage, demonstrating synergistic effects.

We introduce Stable Code, the first in our new-generation of code language models series, which serves as a general-purpose base code language model targeting code completion, reasoning, math, and other software engineering-based tasks. Additionally, we introduce an instruction variant named Stable Code Instruct that allows conversing with the model in a natural chat interface for performing question-answering and instruction-based tasks. In this technical report, we detail the data and training procedure leading to both models. Their weights are available via Hugging Face for anyone to download and use at https://huggingface.co/stabilityai/stable-code-3b and https://huggingface.co/stabilityai/stable-code-instruct-3b. This report contains thorough evaluations of the models, including multilingual programming benchmarks, and the MT benchmark focusing on multi-turn dialogues. At the time of its release, Stable Code is the state-of-the-art open model under 3B parameters and even performs comparably to larger models of sizes 7 billion and 15 billion parameters on the popular Multi-PL benchmark. Stable Code Instruct also exhibits state-of-the-art performance on the MT-Bench coding tasks and on Multi-PL completion compared to other instruction tuned models. Given its appealing small size, we also provide throughput measurements on a number of edge devices. In addition, we open source several quantized checkpoints and provide their performance metrics compared to the original model.

Retrieval-based code question answering seeks to match user queries in natural language to relevant code snippets. Previous approaches typically rely on pretraining models using crafted bi-modal and uni-modal datasets to align text and code representations. In this paper, we introduce ProCQA, a large-scale programming question answering dataset extracted from the StackOverflow community, offering naturally structured mixed-modal QA pairs. To validate its effectiveness, we propose a modality-agnostic contrastive pre-training approach to improve the alignment of text and code representations of current code language models. Compared to previous models that primarily employ bimodal and unimodal pairs extracted from CodeSearchNet for pre-training, our model exhibits significant performance improvements across a wide range of code retrieval benchmarks.

Static call graph (CG) construction often over-approximates call relations, leading to sound, but imprecise results. Recent research has explored machine learning (ML)-based CG pruning as a means to enhance precision by eliminating false edges. However, current methods suffer from a limited evaluation dataset, imbalanced training data, and reduced recall, which affects practical downstream analyses. Prior results were also not compared with advanced static CG construction techniques yet. This study tackles these issues. We introduce the NYXCorpus, a dataset of real-world Java programs with high test coverage and we collect traces from test executions and build a ground truth of dynamic CGs. We leverage these CGs to explore conservative pruning strategies during the training and inference of ML-based CG pruners. We conduct a comparative analysis of static CGs generated using zero control flow analysis (0-CFA) and those produced by a context-sensitive 1-CFA algorithm, evaluating both with and without pruning. We find that CG pruning is a difficult task for real-world Java projects and substantial improvements in the CG precision (+25%) meet reduced recall (-9%). However, our experiments show promising results: even when we favor recall over precision by using an F2 metric in our experiments, we can show that pruned CGs have comparable quality to a context-sensitive 1-CFA analysis while being computationally less demanding. Resulting CGs are much smaller (69%), and substantially faster (3.5x speed-up), with virtually unchanged results in our downstream analysis.

This work introduces self-infilling code generation, a general framework that incorporates infilling operations into auto-regressive decoding. Our approach capitalizes on the observation that recent infilling-capable code language models can self-infill: whereas infilling operations aim to fill in the middle based on a predefined prefix and suffix, self-infilling sequentially generates both such surrounding context and the infilled content. We utilize this capability to introduce novel interruption and looping mechanisms in conventional decoding, evolving it into a non-monotonic process. Interruptions allow for postponing the generation of specific code until a definitive suffix is established, enhancing control over the output. Meanwhile, the looping mechanism, which leverages the complementary nature of self-infilling and left-to-right decoding, can iteratively update and synchronize each piece of generation cyclically. Extensive experiments are conducted to demonstrate that our proposed decoding process is effective in enhancing both regularity and quality across several code generation benchmarks.

Pretrained code language models have enabled great progress towards program synthesis. However, common approaches only consider in-file local context and thus miss information and constraints imposed by other parts of the codebase and its external dependencies. Existing code completion benchmarks also lack such context. To resolve these restrictions we curate a new dataset of permissively licensed Python packages that includes full projects and their dependencies and provide tools to extract non-local information with the help of program analyzers. We then focus on the task of function call argument completion which requires predicting the arguments to function calls. We show that existing code completion models do not yield good results on our completion task. To better solve this task, we query a program analyzer for information relevant to a given function call, and consider ways to provide the analyzer results to different code completion models during inference and training. Our experiments show that providing access to the function implementation and function usages greatly improves the argument completion performance. Our ablation study provides further insights on how different types of information available from the program analyzer and different ways of incorporating the information affect the model performance.