LLM-Generated Code Suffers from "Patchwork Problem" of Incoherence.

The "patchwork problem" describes a critical flaw in code generated by Large Language Models (LLMs): while individual code segments might appear correct and pass local tests, they often lack global structural coherence, leading to failures when deployed. Examples include undeclared configuration keys, missing package imports, or absent authentication guards on new routes. These issues are particularly insidious because conventional Continuous Integration (CI) tools frequently fail to detect them. The research formalizes structural coherence as consistency invariants across various graph representations of repository artifacts, such as import, call, dependency, and configuration graphs. It also proposes an eight-category taxonomy for these failures. To combat this, a hybrid verification framework is introduced, combining existing static analysis tools with purpose-built detectors for cross-cutting invariants that current toolchains overlook. Empirical evaluations confirm that most structural failures bypass type checking, testing, and static application security testing (SAST), highlighting a significant blind spot in current software quality assurance for LLM-generated code.