A-CEGIS Framework Uses Counterexamples for Multi-Turn Agent Refinement

The A-CEGIS framework uses deterministic counterexamples to guide agents in natural-language-to-regex synthesis. In NL-RX-Turk tasks, providing diagnostic feedback increased success rates to 90% within four turns, significantly outperforming generic self-correction at 27%.