New Paper Unveils Sound Tests to Catch Flawed LLM Optimization Models
Zero false positives in verifying LLM-generated optimization models—first theory-backed method.
A new paper from Haifeng Li and Mo Hai tackles a critical blind spot in LLM-based code generation: optimization models that run without errors but solve the wrong problem. The authors develop a falsification-based verification framework that tests candidate models by transforming numerical slots (e.g., changing constraints) and checking solver outputs against expected patterns, without needing a reference model. The battery includes six test classes derived from duality, comparative statics, and polyhedral limits—covering directions, curvature, crush probes, prohibitive limits, annihilation, and exchange. Every test is sound: a violation means the model is definitely wrong.
Experiments on 326 ground-truth models from NL4OPT and four benchmark families validate the approach. The battery achieves a 0.0% false-positive rate (compared to 54.9% for a naive threshold perturbation tester) and detects 70% of certified conditional-class mutants. Notably, it convicts 40.4% of the mutants that execution-accuracy scoring misses entirely. The theory also characterizes what these tests can never detect and proves that no fixed-threshold tester can be simultaneously sound and nontrivial—making this a foundational step toward trustworthy AI-generated decision models.
- Sound test battery guarantees zero false positives—a violation always means the model is wrong.
- Detects 70% of certified error types and catches 40.4% of mutants invisible to execution-accuracy scoring.
- Proves no fixed-threshold perturbation tester can be both sound and nontrivial, grounding future verification methods.
Why It Matters
Reliable verification for LLM-generated optimization models is essential before deploying AI in critical decision systems.