Researchers quantify and mitigate LLM test generation misguidance from buggy code

A new study by Junda Zhao, Shurui Zhou, and Eldan Cohen (University of Toronto) tackles a critical flaw in using LLMs for automated unit test generation. When prompted with buggy source code, LLMs often produce tests that validate the erroneous behavior rather than expose it — a phenomenon they term the 'misguidance effect.' The researchers develop a quantitative metric to measure this effect and confirm its severity: buggy code causes a twofold impact — it significantly increases the number of tests that assert incorrect behavior while simultaneously suppressing the generation of tests that actually catch bugs. By analyzing internal model preferences, they show that the buggy code skews the model's probability distribution toward tests that match the flawed logic.

To mitigate this, the team introduces a specification-based test generation paradigm. Instead of feeding the buggy code into the prompt, they replace it with an LLM-generated specification docstring that describes the intended behavior. Test results show this approach dramatically reduces misguided tests and boosts the number of effective, bug-finding tests. The method also improves multi-round, feedback-driven test generation pipelines and works equally well for bug-free code. The findings suggest that decoupling the test generation from the actual code — using specifications instead — is a promising strategy for avoiding misguidance. The paper is accepted at ISSTA 2026 and will appear in PACMSE.