AI can now tune hardware verification tools without breaking correctness — and then disappears from the final code

Researchers built a system where an AI proposes small tweaks to a hardware verification algorithm, but every proposed change must pass a mathematical proof check before it's accepted. This means AI can optimize the tool without risk of introducing hidden bugs that slip through — the AI's work gets independently verified before deployment, then removed entirely from the final version.

Hardware verification is a grinding computational problem: you need to check that a chip design is safe before it goes into production, and the standard algorithm (called IC3) has dozens of tuning knobs that dramatically affect speed. Currently, engineers manually adjust these knobs, which is slow and fragile. This system lets an AI propose improvements, but enforces mathematical correctness at every step — no guesswork allowed. The result is a faster verification tool that contains no AI code at runtime, so there's no inference cost and no dependency on the model being available later.