LimICE's LLM-powered framework infers loop invariants 63% faster
Solves 349 linear benchmarks in 15.2 seconds on average
Loop invariant synthesis is a core challenge in program verification, but undecidability makes it notoriously hard. Machine learning approaches often fail because they treat invariant generation as a monolithic problem. Kai Fan and colleagues from NUDT propose Incremental ICE, a framework that borrows IC3's incremental philosophy and layers it onto the ICE (Implication, Counterexample, Equivalence) learning paradigm. Their tool, LimICE, uses LLMs to generate an ordered sequence of lemmas rather than a single formula, with ICE-DT as a fallback to complement incomplete sequences.
On 367 linear and 50 nonlinear benchmarks, LimICE solves 349 linear problems (avg 15.2s) and 47 nonlinear problems (avg 8.8s). Compared to the state-of-the-art LLM-based baseline, it solves 12-24% more instances while running 36-63% faster. It also outperforms strong non-LLM baselines, solving at least 86 additional linear and 27 additional nonlinear benchmarks. These results demonstrate that combining LLM-generated lemma ordering with sound incremental learning can push the boundaries of automated verification.
- LimICE integrates LLMs into the ICE framework with IC3-style incremental learning, generating ordered lemma sequences for loop invariants.
- Solves 349/367 linear and 47/50 nonlinear benchmarks, beating LLM baselines by 12-24% more solved instances while running 36-63% faster.
- Outperforms non-LLM baselines by at least 86 additional linear and 27 additional nonlinear solved benchmarks.
Why It Matters
Faster and more complete loop invariant inference means stronger automated verification for critical software, reducing manual proof effort.