IC3-Evolve: Offline LLM Heuristic Evolution for IC3 Model Checking
LLMs evolve IC3 heuristics under formal proof/witness gates, yielding standalone checkers that generalize beyond HWMCC.
An offline LLM framework evolves slot-restricted patches to IC3 implementations, admitted exclusively after independent certificate validation for inductive invariants on SAFE runs and replayable counterexample traces on UNSAFE runs (arXiv:2604.03232).
IC3 performance depends on interacting heuristics whose manual tuning is costly and brittle; IC3-Evolve automates discovery on HWMCC benchmarks and generalizes to unseen public and industrial suites while preserving soundness, extending Bradley's 2011 PDR algorithm (arXiv:1103.2146; arXiv:2604.03232). Related NeuroSAT work applied ML to guide SAT search at runtime but did not evolve standalone checker source code (arXiv:1706.08561).
Initial arXiv abstract omits explicit linkage to EDA toolchain integration patterns seen in recent AI-hardware co-design verification bottlenecks; the gated offline evolution supplies auditable, zero-inference artifacts distinct from online LLM assistants.
AXIOM: LLMs are moving from inference to offline evolution of formally gated heuristics, shortening verification cycles for AI-designed chips.
Sources (3)
- [1]Primary Source(https://arxiv.org/abs/2604.03232)
- [2]IC3 Original(https://arxiv.org/abs/1103.2146)
- [3]NeuroSAT(https://arxiv.org/abs/1706.08561)