Hridesh Rajan

Effectively testing Artificial Intelligence (AI) agents remains a fundamental challenge due to their stochastic reasoning, vast and diverse input space, reliance on external tools, and operation in dynamic execution environments; factors that demand new testing methodologies explicitly tailored to the complex and interactive nature of agent-based systems. This work presents a novel methodology for testing AI agents, with a particular focus on assessing their behavioral robustness under varied operational conditions. Our approach relies on following key technical innovations: (1) a coverage-guided test input generation strategy based on agent- specific coverage objectives, (2) a capture-and-simulate mechanism that systematically emulates abnormal tool behaviors to mimic real-world execution failures, and (3) a deterministic behavioral failure detection approach that enables consistent identification of failures across different test inputs. We developed AgentInspect, a framework that automatically detects six types of behavioral failures in LangChain-based AI agents by analyzing their execution trajectories across three evaluation settings: a baseline setting using real tool responses, a simulated setting incorporating synthetic tool responses, and a hybrid setting that combines the real and simulated tool responses. To evaluate our approach, we curated a benchmark of 35 AI agents obtained from GitHub. Our results show that AgentInspect consistently identifies different behavioral failures with high precision and recall across all three execution settings. In particular, the simulated and hybrid settings expose failure modes that do not emerge during baseline execution with real tool responses, thereby enabling a more comprehensive assessment of agent robustness. Our findings highlight AgentInspect’s effectiveness in revealing critical failures and its practical utility for systematic robustness evaluation of AI agents.

Deep Learning (DL) applications have become essential in numerous domains, yet they remain plagued by subtle bugs that cause 66% of crashes in production systems. These failures primarily stem from API usage violations in complex frameworks like TensorFlow, Keras, and PyTorch, where APIs lack formal specifications and interdependencies between operations remain undocumented. Traditional static analysis tools fail to address DL-specific constraints, such as data dependency between layers. To bridge this critical gap, we propose NEURALSTATE, an approach to detect performance and program crash bugs in a DL program. NEURALSTATE follows a four-step process: (i) gather specifications for Deep Learning operations from different sources; (ii) introduce abstract states to represent these Deep Learning operations; (iii) design formal rules for transitioning between states based on the specifications; (iv) utilize a combination of standard analysis techniques (i.e., typestate and value propagation) to identify bugs in a DL program. Our evaluation on real-world benchmarks demonstrates NEURALSTATE’s effectiveness, achieving a 25% improvement in precision and 63% improvement in recall compared to state-of-the-art tools. Most importantly, NEURALSTATE successfully detects 18 subtle bugs in 45 real-world programs that existing techniques miss entirely.

Deep Learning (DL) is a class of machine learning algorithms that are used in a wide variety of applications. Like any software system, DL programs can have bugs. To support bug localization in DL programs, several tools have been proposed in the past. As most of the bugs that occur due to improper model structure known as structural bugs lead to inadequate performance during training, it is challenging for developers to identify the root cause and address these bugs. To support bug detection and localization in DL programs, in this article, we propose Theia, which detects and localizes structural bugs in DL programs. Unlike the previous works, Theia considers the training dataset characteristics to automatically detect bugs in DL programs developed using two DL libraries, Keras and PyTorch . Since training the DL models is a time-consuming process, Theia detects these bugs at the beginning of the training process and alerts the developer with informative messages containing the bug’s location and actionable fixes which will help them to improve the structure of the model. We evaluated Theia on a benchmark of 40 real-world buggy DL programs obtained from Stack Overflow . Our results show that Theia successfully localizes 57/75 structural bugs in 40 buggy programs, whereas NeuraLint, a state-of-the-art approach capable of localizing structural bugs before training localizes 17/75 bugs.

Not a day goes by without hearing about the impressive feats of large language models (LLMs), and equally, not a day passes without hearing about their challenges. LLMs are notoriously vulnerable to biases in their dataset, leading to issues such as toxicity, harmful responses, and factual inaccuracies. While domain-adaptive training has been employed to mitigate these issues, these techniques often address all model parameters indiscriminately during the repair process, resulting in poor repair quality and reduced model versatility. In this paper, drawing inspiration from fault localization via program slicing, we introduce a novel dynamic slicing-based intent-aware LLM repair strategy, IRepair. This approach selectively targets the most error-prone sections of the model for repair. Specifically, we propose dynamically slicing the model’s most sensitive layers that require immediate attention, concentrating repair efforts on those areas. This method enables more effective repairs with potentially less impact on the model’s overall versatility by altering a smaller portion of the model. Furthermore, dynamic selection allows for a more nuanced and precise model repair compared to a fixed selection strategy. We evaluated our technique on three models from the GPT2 and GPT-Neo families, with parameters ranging from 800M to 1.6B, in a toxicity mitigation setup. Our results show that IRepair repairs errors 43.6% more effectively while causing 46% less disruption to general performance compared to the closest baseline, direct preference optimization. Our empirical analysis also reveals that errors are more concentrated in a smaller section of the model, with the top 20% of layers exhibiting 773% more error density than the remaining 80%. This highlights the need for selective repair. Additionally, we demonstrate that a dynamic selection approach is essential for addressing errors dispersed throughout the model, ensuring a robust and efficient repair.

Programming languages are essential tools for developers, and their evolution plays a crucial role in supporting the activities of developers. One instance of programming language evolution is the introduction of syntactic sugars, which are additional syntax elements that provide alternative, more readable code constructs. However, the process of designing and evolving a programming language has traditionally been guided by anecdotal experiences and intuition. Recent advances in tools and methodologies for mining open-source repositories have enabled developers to make data-driven software engineering decisions. In light of this, this paper proposes an approach for motivating data-driven programming evolution by applying frequent subgraph mining techniques to a large dataset of 166,827,154 open-source Java methods. The dataset is mined by generalizing Java control-flow graphs to capture broad programming language usages and instances of duplication. Frequent subgraphs are then extracted to identify potentially impactful opportunities for new syntactic sugars. Our diverse results demonstrate the benefits of the proposed technique by identifying new syntactic sugars involving a variety of programming constructs that could be implemented in Java, thus simplifying frequent code idioms. This approach can potentially provide valuable insights for Java language designers, and serve as a proof-of-concept for data-driven programming language design and evolution.

Code intelligence tools such as GitHub Copilot have begun to bridge the gap between natural language and programming language. A frequent software development task is the management of technical debts, which are suboptimal solutions or unaddressed issues which hinder future software development. Developers have been found to "self-admit" technical debts (SATD) in software artifacts such as source code comments. Thus, is it possible that the information present in these comments can enhance code generative prompts to repay the described SATD? Or, does the inclusion of such comments instead cause code generative tools to reproduce the harmful symptoms of described technical debt? Does the modification of SATD impact this reaction? Despite the heavy maintenance costs caused by technical debt and the recent improvements of code intelligence tools, no prior works have sought to incorporate SATD towards prompt engineering. Inspired by this, this paper contributes and analyzes a dataset consisting of 36,381 TODO comments in the latest available revisions of their respective 102,424 repositories, from which we sample and manually generate 1,140 code bodies using GitHub Copilot. Our experiments show that GitHub Copilot can generate code with the symptoms of SATD, both prompted and unprompted. Moreover, we demonstrate the tool's ability to automatically repay SATD under different circumstances and qualitatively investigate the characteristics of successful and unsuccessful comments. Finally, we discuss gaps in which GitHub Copilot's successors and future researchers can improve upon code intelligence tasks to facilitate AI-assisted software maintenance.

Machine learning (ML) is increasingly being used in critical decision-making software, but incidents have raised questions about the fairness of ML predictions. To address this issue, new tools and methods are needed to mitigate bias in ML-based software. Previous studies have proposed bias mitigation algorithms that only work in specific situations and often result in a loss of accuracy. Our proposed solution is a novel approach that utilizes automated machine learning (AutoML) techniques to mitigate bias. Our approach includes two key innovations: a novel optimization function and a fairness-aware search space. By improving the default optimization function of AutoML and incorporating fairness objectives, we are able to mitigate bias with little to no loss of accuracy. Additionally, we propose a fairness-aware search space pruning method for AutoML to reduce computational cost and repair time. Our approach, built on the state-of-the-art Auto-Sklearn tool, is designed to reduce bias in real-world scenarios. In order to demonstrate the effectiveness of our approach, we evaluated our approach on four fairness problems and 16 different ML models, and our results show a significant improvement over the baseline and existing bias mitigation techniques. Our approach, Fair-AutoML, successfully repaired 60 out of 64 buggy cases, while existing bias mitigation techniques only repaired up to 44 out of 64 cases.