Noor Nashid

Noor Nashid, Daniel Ding, Keheliya Gallaba et al.

Automated program repair has traditionally focused on single-hunk defects, overlooking multi-hunk bugs that are prevalent in real-world systems. Repairing these bugs requires coordinated edits across multiple, disjoint code regions, posing substantially greater challenges. We present the first systematic study of LLM-driven coding agents (Claude Code, Codex, Gemini-cli, and Qwen Code) on this task. We evaluate these four state-of-the-art agents on 404 multi-hunk bugs from the PolyHunk dataset, yielding 1,616 repair trajectories for large-scale behavioral analysis. We employ fine-grained metrics to assess localization, repair accuracy, regression behavior, and operational dynamics across agents. We find that localization capability varies substantially, with Codex achieving the highest success rate (75.3%) and Qwen Code the lowest (40.4%). Repair accuracy also differs widely, ranging from 26.98% (Qwen Code) to 92.82% (Claude Code), and consistently declines with increasing bug dispersion and complexity (hunk divergence and spatial proximity). High-performing agents (Claude Code and Codex) demonstrate superior semantic consistency, achieving positive average regression reduction, whereas lower-performing agents often introduce new test failures. Notably, agents do not fail fast; failed repairs consume substantially more resources (33%-440% more input tokens) and require longer execution time (35%-330%). Additionally, we developed Maple to provide agents with repository-level context. Empirical results show that Maple improves repair accuracy of Gemini-cli by ~21% through enhanced localization. By analyzing fine-grained metrics and trajectory-level analysis, this study moves beyond accuracy to explain how coding agents localize, reason, and act during multi-hunk repair. Our findings underscore the impact of bug divergence and spatial proximity on multi-hunk repair success for coding agents.

Marjane Namavar, Noor Nashid, Ali Mesbah

Training a deep learning model on source code has gained significant traction recently. Since such models reason about vectors of numbers, source code needs to be converted to a code representation before vectorization. Numerous approaches have been proposed to represent source code, from sequences of tokens to abstract syntax trees. However, there is no systematic study to understand the effect of code representation on learning performance. Through a controlled experiment, we examine the impact of various code representations on model accuracy and usefulness in deep learning-based program repair. We train 21 different generative models that suggest fixes for name-based bugs, including 14 different homogeneous code representations, four mixed representations for the buggy and fixed code, and three different embeddings. We assess if fix suggestions produced by the model in various code representations are automatically patchable, meaning they can be transformed to a valid code that is ready to be applied to the buggy code to fix it. We also conduct a developer study to qualitatively evaluate the usefulness of inferred fixes in different code representations. Our results highlight the importance of code representation and its impact on learning and usefulness. Our findings indicate that (1) while code abstractions help the learning process, they can adversely impact the usefulness of inferred fixes from a developer's point of view; this emphasizes the need to look at the patches generated from the practitioner's perspective, which is often neglected in the literature, (2) mixed representations can outperform homogeneous code representations, (3) bug type can affect the effectiveness of different code representations; although current techniques use a single code representation for all bug types, there is no single best code representation applicable to all bug types.