IEEE Access accepts a paper co-authored by Lecturer Shimari

A paper co-authored by Lecturer Kazumasa Shimari was accepted by IEEE Access on September 19, 2026. It is now published in Volume 14, pp. 153014–153030.

Fault Localization Using LLMs With Line-by-Line Program Execution Traces

Authors: Takafumi Sakura, Ryo Soga, Hideyuki Kanuka, Kazumasa Shimari, Takashi Ishio

Publication / venue: IEEE Access, Vol.14, pp.153014-153030, 2026.

Abstract

Fault localization (FL) is a fundamental process for identifying the root causes of software bugs. Although existing methods demonstrate that integrating static code with dynamic runtime information improves accuracy, a limitation remains. Although omniscient debugging tools provide fine-grained, line-level execution traces, their effectiveness in automated fault analysis remains unclear because they are primarily used for manual navigation. In this study, we addressed this limitation by incorporating line-level execution traces into a large language model (LLM)-based FL. We evaluated the approach using 95 single-line, single-method Java bugs selected from Defects4J v1.2 and three LLMs, assuming that the buggy method was known, and compared three input conditions: no trace, stack trace, and line-level execution trace. The execution-trace condition achieved the highest mean reciprocal rank (MRR) for all models, and relative to the no-trace condition, the correct-line ranking improved by 26%–35% for the bugs. The analysis of the improved and degraded cases showed that trace-specific runtime facts were often referenced as evidence in the improved cases, whereas incomplete or mismatched trace information was observed in the degraded cases. In a supplementary analysis of 15 model–bug pairs (30 explanations), explanations with execution traces were more consistently grounded in the provided information and less reliant on speculative reasoning than those without execution traces. For the 10 selected bugs, the median total additional time associated with the execution traces was 8.4 s. These findings indicate that, within the evaluated setting, line-level execution traces can be a useful information source for LLM-based FL.

Source: Paper / presentation abstract