Parameter-efficient fine-tuning (PEFT) has become a popular way to adapt large pre-trained models to new tasks. Most PEFT methods update only a small subset of parameters while freezing the rest, avoiding redundant computation. As they maximize the absolute size of the updates without regard to the parameters’ original scale, the resulting changes in model behavior can be minimal. In contrast, we maximize updates relative to each parameter’s scale, yielding more meaningful downstream adaptation. We propose Gradient-to-Weight Ratio and Entropy-guided Masking (GEM), a parameter scale-aware, distribution-sensitive sparse fine-tuning framework. GEM prioritizes parameters whose updates are significant in proportion to their initial pre-trained values. It also adaptively determines how many parameters to tune at each layer based on the entropy of parameter values, thereby making the most effective use of the computational budget in PEFT. Our empirical study demonstrates the efficacy of GEM on both general-domain tasks (GLUE and SuperGLUE) and domain-specific tasks (GSM8k and MBPP), achieving up to a 1.6% improvement in fine-tuning accuracy over full fine-tuning while updating only 0.1% of model parameters.

Yavuz Faruk Bakman, Duygu Nur Yaldiz, Sungmin Kang, Tuo Zhang, Baturalp Buyukates, Salman Avestimehr, Sai Praneeth Karimireddy. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Expensive communication cost is a common performance bottleneck in Federated Learning (FL), which makes it less appealing in real-world applications. Many communication-efficient FL methods focus on discarding a part of model updates mostly based on gradient magnitude. In this study, we find that recycling previous updates, rather than simply dropping them, more effectively reduces the communication cost while maintaining FL performance. We propose *FedLUAR*, a Layer-wise Update Aggregation with Recycling scheme for communication-efficient FL. We first define a useful metric that quantifies the extent to which the aggregated gradients influences the model parameter values in each layer. *FedLUAR* selects a few layers based on the metric and recycles their previous updates on the server side. Our extensive empirical study demonstrates that the update recycling scheme significantly reduces the communication cost while maintaining model accuracy. For example, our method achieves nearly the same AG News accuracy as FedAvg, while reducing the communication cost to just 17%.

Many works have recently proposed the use of Large Language Model (LLM) based agents for performing ‘repository level’ tasks, loosely defined as a set of tasks whose scopes are greater than a single file. This has led to speculation that the orchestration of these repository-level tasks could lead to software engineering agents capable of performing almost independently of human intervention. However, of the suite of tasks that would need to be performed by this autonomous software engineering agent, we argue that one important task is missing, which is to fulfil project level dependency by installing other repositories. To investigate the feasibility of this repository level installation task, we introduce a benchmark of of repository installation tasks curated from 40 open source Python projects, which includes a ground truth installation process for each target repository. Further, we propose Installamatic, an agent which aims to perform and verify the installation of a given repository by searching for relevant instructions from documentation in the repository. Empirical experiments reveal that that 55% of the studied repositories can be automatically installed by our agent at least one out of ten times. Through further analysis, we identify the common causes for our agent's inability to install a repository, discuss the challenges faced in the design and implementation of such an agent and consider the implications that such an agent could have for developers.

Abstract Automated debugging techniques have the potential to reduce developer effort in debugging. However, while developers want rationales for the provided automatic debugging results, existing techniques are ill-suited to provide them, as their deduction process differs significantly froof human developers. Inspired by the way developers interact with code when debugging, we propose Automated Scientific Debugging ( AutoSD ), a technique that prompts large language models to automatically generate hypotheses, uses debuggers to interact with buggy code, and thus automatically reach conclusions prior to patch generation. In doing so, we aim to produce explanations of how a specific patch has been generated, with the hope that these explanations will lead to enhanced developer decision-making. Our empirical analysis on three program repair benchmarks shows that AutoSD performs competitively with other program repair baselines, and that it can indicate when it is confident in its results. Furthermore, we perform a human study with 20 participants to evaluate AutoSD -generated explanations. Participants with access to explanations judged patch correctness more accurately in five out of six real-world bugs studied. Furthermore, 70% of participants answered that they wanted explanations when using repair tools, and 55% answered that they were satisfied with the Scientific Debugging presentation.

While Large Language Models (LLMs) have demon-strated strong natural language and code processing capabilities, concern has been raised as to whether existing bug benchmarks are included in their training data. We examine the training data of the open-source LLM StarCoder, and find it likely that data from the widely used Defects4J benchmark was included, raising the possibility of its inclusion in the training data of the GPT model as well. This makes it difficult to tell how well LLM-based results on Defects4J would generalize, as for any results it would be unclear whether a technique's performance is due to LLM generalization or memorization. To remedy this issue and facilitate continued research on LLM-based SE, we present the GitHub Recent Bugs (GHRB) framework, which continuously gathers real-world Java bugs for use in evaluation of LLM-based techniques. To date, we have gathered 89 bugs reported after the GPT-3.5 training data cutoff point of September 2021.

Due to the rapid adoption of Deep Neural Networks (DNNs) into larger software systems, testing of DNN-based systems has received much attention recently. While many different test adequacy criteria have been suggested, we lack effective test input generation techniques. Inputs such as images of real-world objects and scenes are not only expensive to collect but also difficult to randomly sample. Consequently, current testing techniques for DNNs tend to apply small local perturbations to existing inputs to generate new inputs. We propose SINVAD (Search-based Input space Navigation using Variational AutoencoDers), a way to sample from, and navigate over, a space of realistic inputs that resembles the true distribution in the training data. Our input space is constructed using Variational Autoencoders (VAEs), and navigated through their latent vector space. Our analysis shows that the VAE-based input space is well-aligned with human perception of what constitutes realistic inputs. Further, we show that this space can be effectively searched to achieve various testing scenarios, such as boundary testing of two different DNNs or analyzing class labels that are difficult for the given DNN to distinguish. Guidelines on how to design VAE architectures are presented as well. Our results have the potential to open the field to meaningful exploration through the space of highly structured images.

Bug reproduction is a critical developer activity that is also challenging to automate, as bug reports are often in natural language and thus can be difficult to transform to test cases consistently. As a result, existing techniques mostly focused on crash bugs, which are easier to automatically detect and verify. In this work, we overcome this limitation by using large language models (LLMs), which have been demonstrated to be adept at natural language processing and code generation. By prompting LLMs to generate bug-reproducing tests, and via a post-processing pipeline to automatically identify promising generated tests, our proposed techniqueLibrocould successfully reproduce about one-third of all bugs in the widely used Defects4J benchmark. Furthermore, our extensive evaluation on 15 LLMs, including 11 open-source LLMs, suggests that open-source LLMs also demonstrate substantial potential, with the StarCoder LLM achieving 70% of the reproduction performance of the closed-source OpenAI LLM code-davinci-002 on the large Defects4J benchmark, and 90% of performance on a held-out bug dataset likely not part of any LLM's training data. In addition, our experiments on LLMs of different sizes show that bug reproduction usingLibroimproves as LLM size increases, providing information as to which LLMs can be used with theLibropipeline.

Many automated test generation techniques have been developed to aid developers with writing tests. To facilitate full automation, most existing techniques aim to either increase coverage, or generate exploratory inputs. However, existing test generation techniques largely fall short of achieving more semantic objectives, such as generating tests to reproduce a given bug report. Reproducing bugs is nonetheless important, as our empirical study shows that the number of tests added in open source repositories due to issues was about 28% of the corresponding project test suite size. Meanwhile, due to the difficulties of transforming the expected program semantics in bug reports into test oracles, existing failure reproduction techniques tend to deal exclusively with program crashes, a small subset of all bug reports. To automate test generation from general bug reports, we propose Libro, a framework that uses Large Language Models (LLMs), which have been shown to be capable of performing code-related tasks. Since LLMs themselves cannot execute the target buggy code, we focus on post-processing steps that help us discern when LLMs are effective, and rank the produced tests according to their validity. Our evaluation of Libro shows that, on the widely studied Defects4J benchmark, Libro can generate failure reproducing test cases for 33% of all studied cases (251 out of 750), while suggesting a bug reproducing test in first place for 149 bugs. To mitigate data contamination (i.e., the possibility of the LLM simply remembering the test code either partially or in whole), we also evaluate Libro against 31 bug reports submitted after the collection of the LLM training data terminated: Libro produces bug reproducing tests for 32% of the studied bug reports. Overall, our results show Libro has the potential to significantly enhance developer efficiency by automatically generating tests from bug reports.

Debugging takes up a significant portion of developer time. As a result, automated debugging techniques including Fault Localization (FL) and Automated Program Repair (APR) have garnered significant attention due to their potential to aid developers in debugging tasks. With the recent advance in techniques that treat the two tasks as closely coupled, such as Unified Debugging, a framework to formally express these two tasks together would heighten our understanding of automated debugging and provide a way to formally analyze techniques and approaches. To this end, we propose a Bayesian framework of understanding automated debugging. We find that the Bayesian framework, along with a concrete statement of the objective of automated debugging, can recover maximal fault localization formulae from prior work, as well as analyze existing APR techniques and their underlying assumptions. As a means of empirically demonstrating our framework, we further propose BAPP, a Bayesian Patch Prioritization technique that incorporates intermediate program values to analyze likely patch locations and repair actions, with its core equations being derived by our Bayesian framework. We find that incorporating program values allows BAPP to identify correct patches more precisely: the rankings produced by BAPP reduced the number of required patch evaluations by 68% and consequently reduced the repair time by 34 minutes on average. Further, our Bayesian framework suggests a number of changes to the way fault localization information is used in program repair, which we validate is useful for BAPP. These results highlight the potential of value-cognizant automated debugging techniques, and further verifies our theoretical framework.

Software testing is an important part of the development cycle, yet it requires specialized expertise and substantial developer effort to adequately test software. Recent discoveries of the capabilities of large language models (LLMs) suggest that they can be used as automated testing assistants, and thus provide helpful information and even drive the testing process. To highlight the potential of this technology, we present a taxonomy of LLM-based testing agents based on their level of autonomy, and describe how a greater level of autonomy can benefit developers in practice. An example use of LLMs as a testing assistant is provided to demonstrate how a conversational framework for testing can help developers. This also highlights how the often criticized “hallucination” of LLMs can be beneficial for testing. We identify other tangible benefits that LLM-driven testing agents can bestow, and also discuss potential limitations.

The rapid and widespread adoption of Deep Neural Networks (DNNs) has called for ways to test their behaviour, and many testing approaches have successfully revealed misbehaviour of DNNs. However, it is relatively unclear what one can do to correct such behaviour after revelation, as retraining involves costly data collection and does not guarantee to fix the underlying issue. This article introduces Arachne, a novel program repair technique for DNNs, which directly repairs DNNs using their input-output pairs as a specification. Arachne localises neural weights on which it can generate effective patches and uses differential evolution to optimise the localised weights and correct the misbehaviour. An empirical study using different benchmarks shows that Arachne can fix specific misclassifications of a DNN without reducing general accuracy significantly. On average, patches generated by Arachne generalise to 61.3% of unseen misbehaviour, whereas those by a state-of-the-art DNN repair technique generalise only to 10.2% and sometimes to none while taking tens of times more than Arachne. We also show that Arachne can address fairness issues by debiasing a gender classification model. Finally, we successfully apply Arachne to a text sentiment model to show that it generalises beyond convolutional neural networks.