Researchers at the Max Planck Institute for the Science of Light and Friedrich-Alexander University Erlangen-Nürnberg have introduced an autonomous agentic artificial intelligence framework that designs and verifies fault-tolerant quantum error-correcting circuits. Published as an arXiv preprint on November 13, 2026, the method couples frontier large language models with an automated circuit verifier, discovering novel quantum circuits that reduce operational depth and outperform existing human-crafted and reinforcement learning benchmarks.
Autonomous Program Synthesis for Fault-Tolerant Quantum Computing
In a preprint titled Agentic AI for Discovering Fault-Tolerant Quantum Circuits (arXiv:2611.18504), physicists Aritra Lahiri and Florian Marquardt demonstrate how agentic evolutionary program synthesis can resolve a central bottleneck in quantum computing. Operating logical qubits protected by quantum error-correcting codes requires physical Clifford circuits to prepare logical states, execute gates, and perform code-switching. Historically, these circuits have relied on manual analytical design or reinforcement learning routines requiring millions of simulations.
The authors formulated circuit design as an evolutionary search over executable Python code. While earlier frameworks like DeepMind’s FunSearch (Romera-Paredes et al., Nature 2024) used language models as sampling mutation operators across thousands of generations, Lahiri and Marquardt provided frontier models with interactive tool execution. Each agent queries a circuit verifier during prompt generation, inspecting simulation feedback and ablating subroutines before submitting code to an island database.
The Agentic FunSearch Loop and Circuit Verification
The system operates in the stabilizer formalism established by Daniel Gottesman (Gottesman, 1997), tracking quantum states via binary tableaus over Galois Field GF(2). Prompts provide target stabilizer generators, physical coupling constraints, available ancillae, and permitted Clifford gates. For logical gates, the target is converted to a logical Choi state via channel-state duality.
The generated Python functions produce dynamic Clifford circuits combining unitary gates, projective measurements, and conditional Pauli corrections. The automated verifier evaluates each candidate through two stages:
- Symplectic validity: The simulator verifies that the circuit prepares the target state while disentangling data qubits from all ancillae across every combination of measurement outcomes.
- Exhaustive fault tolerance: Under circuit-level depolarizing noise, the verifier models gate failures, input errors, and measurement flips up to order t. It tracks fault propagation to determine residual Pauli errors on data qubits.
The framework utilizes flag fault tolerance (Chao and Reichardt, 2018), where deterministic ancilla measurements act as fault detectors. Any fault event leaving a heavy residual error on data qubits must fire a detector. Heavy errors that fail to trigger detectors are categorized as “silent events.” Eliminating silent events is the primary optimization objective.
Quantitative Reductions in Circuit Depth and Ancilla Overhead
The agent was benchmarked across standard codes, including the non-CSS [[5,1,3]] perfect code, rotated surface codes from distance 3 to 9, the Steane [[7,1,3]] code, and the Reed-Muller [[15,1,3]] code. The agent consistently found shallower circuits with strict fault tolerance.
| Quantum Code and Task | Hardware Coupling | Benchmark Depth | Agentic AI Depth | Silent Faults | Key Advantage |
|---|---|---|---|---|---|
| [[5,1,3]] Logical State Prep | All-to-all | 12 (RL Baseline) | 8 | 0 | 33% depth reduction using 12 two-qubit gates |
| [[9,1,3]] Surface Code d=3 | Native planar | 7 (Measurement Enc.) | 6 | 0 | Zero silent faults vs 16 in standard heuristic |
| [[9,1,3]] Surface Code d=3 | Hexagonal | Not reported | 6 | 0 | Adapted to constrained hex hardware layouts |
| [[25,1,5]] Surface Code d=5 | All-to-all | 16 (Literature) | 9 | 0 | 44% depth reduction; zero-waste option at depth 16 |
| [[7,1,3]] to [[15,1,3]] Switch | All-to-all | Custom analytical | 20 | 0 | Transversal gate switching with 40 two-qubit gates |
For rotated surface codes, the agent generated a single multi-size program scaling across distances d = 3, 5, 7, and 9. At distance 3 on planar layouts, depth reached 6 with 4 flag detectors. For distances 5 through 9, depth remained at 7. Unlike standard measurement-based encodings that produce unheralded hook errors, the agentic circuits herald all errors dynamically.
The framework also designed logical gate routines. Because the Eastin-Knill theorem prevents any single code from implementing universal gates transversally, architectures require code switching or magic state distillation. The agent synthesized an autonomous switching circuit between the Steane code (transversal Clifford gates) and the Reed-Muller code (transversal T gate), using 2 ancillae, 4 detectors, and 40 two-qubit gates.
Frontier Reasoning Tiers and Evolutionary Dynamics
The study benchmarked three model tiers: a base tier using OpenAI’s GPT-6-luna with high reasoning effort, a strong tier with maximum reasoning effort, and an elite tier using GPT-6.1-sol. Analyzing how each frontier model navigates formal verification provides key empirical insights for automated science.
The elite tier produced significantly higher record rates, establishing new best scores in fewer generations. Scratchpad analysis revealed that higher reasoning effort in GPT-6-luna expanded scratchpad notes by 25%, while switching to GPT-6.1-sol increased word count by over 125%. Richer scratchpads correlated with systematic hypothesis generation and disciplined pruning of failed circuit topologies.
Implications for Quantum Engineering and Research Administration
The research presents several practical takeaways for quantum laboratories and research administration offices:
- Hardware-tailored compilation: Device teams with custom coupling constraints can deploy agentic synthesis to generate verified circuits directly, accelerating physical processor milestones (Acharya et al., Nature 2023).
- Rigor in automated workflows: Pairing generative models with formal verifiers provides a robust model for a reproducible AI experiment, eliminating unverified code prior to hardware deployment.
- Managing research computing infrastructure: Research facilities supporting quantum initiatives must support hybrid workflows integrating frontier model APIs with numerical simulators, as highlighted in CASRAI guidance on reproducibility infrastructure.
- Institutional assessment of preprints: As autonomous agents accelerate preprint output, academic departments require consistent policies to evaluate unrefereed findings, detailed in CASRAI’s analysis of arXiv preprints.
Future Trajectories and Open Challenges
Future extensions face two main frontiers. Expanding beyond the stabilizer formalism to handle continuous-angle rotations and arbitrary magic-state distillation will require state-vector or tensor network simulators. Furthermore, verifying fault tolerance beyond distance t = 1 in large codes encounters combinatorial limits, requiring hybrid analytical verification and stochastic Monte Carlo sampling.
Sources
- Lahiri, A., and Marquardt, F. (2026). “Agentic AI for Discovering Fault-Tolerant Quantum Circuits.” arXiv preprint arXiv:2611.18504 [quant-ph]. Checked 2026-11-16. https://arxiv.org/abs/2611.18504
- Romera-Paredes, B., Barekatain, M., Novikov, A., et al. (2024). “Mathematical discoveries from program search with large language models.” Nature, 625, 468–475. Checked 2026-11-16. https://www.nature.com/articles/s41586-023-06924-6
- Gottesman, D. (1997). “An Introduction to Quantum Error Correction and Fault-Tolerant Quantum Computation.” arXiv preprint quant-ph/9705052. Checked 2026-11-16. https://arxiv.org/abs/quant-ph/9705052
- Chao, R., and Reichardt, B. W. (2018). “Quantum Error Correction with Only Two Extra Qubits.” Physical Review Letters, 121(5), 050502 / arXiv:1705.02329. Checked 2026-11-16. https://arxiv.org/abs/1705.02329
- Google Quantum AI (Acharya, R., et al.) (2023). “Suppressing quantum errors by scaling a quantum error-correcting code.” Nature, 614, 676–681. Checked 2026-11-16. https://www.nature.com/articles/s41586-022-05434-1







