[CAIS’26] Vista: Verifier-in-the-Loop Agentic Reinforcement Learning for Quantum Program Synthesis
Published in In the Proc. of the ACM Conference on AI and Agentic Systems (CAIS), 2026
We present Vista, a verifier-in-the-loop agentic reinforcement learning system for quantum program synthesis, instantiated for OpenQASM 3.0 circuit generation. Because verifier stages differ sharply in cost, latency, and informativeness, running the full verifier on every candidate is wasteful, while collapsing all outcomes into a single reward destabilizes learning. Vista introduces (i) hierarchical verified reward optimization, which turns staged verifier outcomes into stable learning signals spanning feasibility, behavior, objective quality, and utility, and (ii) budget-aware gated evaluation, which schedules expensive verification stages strategically. Presented at ACM CAIS 2026.