Agentic Root Cause Reasoning and Localization of Unsoundness in Static Call Graph Analysis. Mohammad Rafieian, Michael Hicks, and Shiyi Wei. In Proceedings of the International Conference on Software Engineering (ICSE), April 2027.

Static call graph construction tools cannot soundly model every dynamic language feature, yet they rarely reveal when, where, or why their results are unsound. Users cannot judge whether a call graph is complete for their program, and developers lack the precise root cause and code location needed to tell an implementation bug from an intended soundness/precision tradeoff. We present CGrizz, an agentic framework addressing both needs in two phases. In an analysis phase, CGrizz compares programs’ static and dynamic call graphs to find boundary edges—calls observed at runtime but missing from the static graph—and, for each, an LLM agent suggests a root cause: the program pattern the analysis failed to model and the responsible location in the tool. Because such a suggestion may be wrong, CGrizz accepts it only when three independent pieces of evidence agree: runtime instrumentation that detects the pattern, a minimized test case that reproduces the missed edge, and a warning inserted at the implicated tool location. The validated reports, test cases, and code locations serve developers; the warnings accumulate into a warning-instrumented build that, in a deployment phase, flags unsoundness on new programs for users, with no dynamic analysis required. We evaluate CGrizz on three complementary benchmark suites with three state-of-the-art tools: WALA, OPAL, and Soot. CGrizz reasons about numerous unsoundness boundaries; the generated warnings generalize, firing on previously unseen programs during deployment; and the resulting artifacts help developers diagnose implementation issues and help users assess an analysis’s reliability and, in some cases, refactor programs to be more amenable to static analysis.

@inproceedings{rafieian27cgrizz,
  author = {Mohammad Rafieian and Michael Hicks and Shiyi Wei},
  title = {Agentic Root Cause Reasoning and Localization of Unsoundness in Static Call Graph Analysis},
  booktitle = {Proceedings of the International Conference on Software Engineering (ICSE)},
  year = 2027,
  month = apr
}

This file was generated by bibtex2html 1.99.