The Signal
LLM-guided search produces fault-tolerant circuit designs
A preprint reports that an LLM-guided evolutionary search generated circuit designs for logical-state preparation, logical gates and routines including code switching. The authors say a circuit verifier exhaustively checked the candidates for faults.

An evolutionary circuit search
An LLM-guided search produced candidate fault-tolerant circuits for logical-state preparation, encoded gates and routines such as code switching, with each candidate checked against faults, according to a new arXiv preprint. Aritra Lahiri and Florian Marquardt submitted the paper on November 13, 2026.
The method, which the paper calls agentic FunSearch, uses a pretrained LLM to propose and revise programs in an evolutionary search. Each program is a Python function specifying a circuit; a verifier evaluates whether it meets its task and how it handles faults. The authors say the verifier exhaustively checks faults, and the approach found fault-tolerant circuits across the curated tasks they tested.
The study focuses on dynamic Clifford circuits, which can include measurements during a circuit and Pauli gates conditioned on measurement outcomes. The tasks span preparing encoded logical states, implementing logical gates, and combining operations into routines. The paper examines circuit designs rather than reporting a run on quantum hardware.
Results across codes and layouts
For logical-state preparation, the authors tested several error-correcting codes and connectivity layouts. On the rotated surface code’s native layout, one search evolved a program for distances 3, 5, 7 and 9. It produced fault-tolerant circuits with no silent fault events in its first generation; the final circuit had depth 6 for distance 3 and depth 7 for the other tested distances.
The paper also reports results for logical gates. For a logical CNOT on the [[5, 1, 3]] perfect code, its table gives a depth of 16 and 29 two-qubit gates for one generated circuit. The table’s cited comparison circuit has depth 84 and 65 two-qubit gates; these are the paper’s circuit-level figures, not measurements of a processor.
Other examples include logical Hadamard gates on a distance-2 colour code and a logical phase gate on the rotated surface code. For a multi-size search over triangular colour codes, fault-tolerant state-preparation circuits appeared in the first generation for distances up to 7, but the search did not find one program that worked across all distances.
Limits in the reported searches
Some tasks remained difficult under the search conditions described. For the logical phase gate on the rotated surface code, a simultaneous search over distances 3 and 5 did not find a circuit without undecodable events at distance 5 within a budget of 60 generations when limited to the logical patch’s own data and ancilla qubits. The authors report success for both distances after adding qubits by embedding the logical patch in a larger patch, then searching each distance separately.
The results cover a curated set of circuit-design tasks and depend on the code, connectivity and resources specified for each search. The preprint presents the circuits and verifier checks as results of its approach; it does not report their execution on a quantum processor.
Source
Related stories
- Researchers synthesize four-circuit fault-tolerant ancilla preparation
- Søren Fuglede Jørgensen Reports Proof That Clifford Gate-Budget Synthesis Is NP-Complete
- Researchers propose fault-tolerant atom arrays with global control
- ETH Zurich team teleports logical states on a 41-qubit circuit
- IBM says million-round repetition code can run in one circuit