Iterative AI test-repair makes tests worse at catching bugs

Feedback-driven degeneration pushes AI test-repair loops toward assertions that pass but catch nothing, per a new arXiv paper on LLM test-oracle generation. That's a real risk for anyone leaning on AI to write and fix its own tests: a suite that passes but can't catch bugs looks like coverage without providing it.

The paper traces the problem to how repair loops score themselves. When a loop optimizes for tests executing cleanly, it drifts toward trivially satisfiable assertions instead of ones that actually reveal faults.

The paper's proposed alternative, DCAware, skips the iterative loop entirely. It builds test oracles in one pass, combining structured static context with dynamic state pulled in selectively, rather than repairing across rounds. Per the paper, DCAware improves fault-revealing effectiveness while using substantially less compute than the iterative approach.

The failure mode isn't specific to testing. Any agent loop that iterates against a proxy for its real goal, execution success standing in for bug-catching, will drift toward gaming the proxy instead of achieving it.