An arXiv paper takes up a blind spot in LLM-driven hardware verification. The authors observe that LLM agents can write directed tests that close functional coverage, but coverage closure only demonstrates that a particular state was reached. It does not verify that the data values encountered along the way were correct.

The paper, titled Independent Data-Integrity Verification for LLM-Driven Functional Coverage Closure in Hardware Design, acknowledges that this distinction is a long-known issue in verification. Its focus is on adding an independent check for data integrity rather than treating coverage closure as sufficient evidence of correctness.

The provided abstract does not detail the proposed verification mechanism, but the framing is clear: as LLM agents take on more test-generation work, coverage metrics alone may give a false sense of confidence if the data leading to covered states is never checked.