ICML 2026 paper: Quantum AI must prioritize validity over scaling

A position paper accepted to ICML 2026 challenges the prevailing assumption that larger AI models will inevitably produce better quantum programs. Authors Junhao Song, Yu Zhou, William Knottenbelt, and Yudao Cao argue that probabilistic scaling—successful for natural language processing—fails for quantum circuit synthesis. The reason: quantum circuits operate in Hilbert space, where syntax (correct structure) and semantics (physical validity) are decoupled. Training on unverified programs teaches models to mimic surface patterns without understanding underlying quantum constraints. As the number of qubits increases, the ratio of valid to invalid circuits drops exponentially, making it mathematically impossible to filter out bad circuits after generation.

The researchers propose a paradigm shift from human-centric copilots to verifier-centric agents. Instead of learning from probabilistic language models, their approach builds validity into the generation process itself. Key techniques include hierarchical constraints that enforce circuit structure layer by layer, topological masks that respect qubit connectivity, and symbolic proxies that embed quantum rules directly into the model's architecture. This verification-aware design ensures that every generated circuit is physically realizable without relying on post-hoc checks. The authors argue this modular approach can scale to practical quantum workloads, whereas pure imitation learning cannot bridge the validity gap.