acceptodds
Under review as a conference paper at ICLR 2027

Automated discovery of quantum subsystem codes with an AI scientist

Abstract

Quantum error correction (QEC) is essential for scalable fault-tolerant quantum computing, yet practical code design remains constrained by the connectivity and imperfections of the underlying hardware. This challenge is particularly acute for superconducting processors, where fixed and sparse couplings can restrict both the realizable code space and the level of logical protection that can be achieved. Here we introduce a verification-feedback-driven LLM framework for hardware-constrained subsystem-code discovery. The framework uses an LLM to generate hardware-aware code constructions, deterministic tools to verify them, and verified feedback to improve subsequent generations. Its central feature is that code generation is conditioned on the specific hardware instance, and independently verified code properties and failure evidence are fed back to guide subsequent refinement. We evaluate the framework on four calibration-informed defect layouts of the 105-qubit Willow graph and three connectivity-design tasks: two Willow augmentations and one free-connectivity search within the same qubit budget. On the four defect layouts, the selected subsystem implementations use weight-two gauge measurements and attain larger code and circuit fault distances than native surface-code templates, while reducing CZ count by \(12.5%\)–\(41.2%\) and CZ depth by \(70.0%\)–\(86.1%\) relative to code-distance-matched routed implementations. Under Willow augmentation, the search finds a distance-7 subsystem code protecting two logical qubits at maximum degree 5 and a distance-5 stabilizer code () protecting 10 logical qubits at maximum degree 6. A separate free-connectivity search finds a distance-9 subsystem construction at maximum degree 4. These results demonstrate the potential of self-evolving LLM search for QEC design under realistic hardware constraints, paving a path toward robust and resource-efficient fault-tolerant quantum computing.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI scientist, Quantum error correction, Quantum computing, AI for quantum

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_dqCEOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an LLM-based AI Scientist framework for automatically discovering hardware-constrained quantum error-correcting codes. The framework integrates LLM-generated executable code constructions with deterministic verification and iterative feedback.

The authors evaluate the framework using GPT, DeepSeek, and Claude on seven tasks based on Google's 105-qubit Willow topology, including four calibration-informed defect scenarios and three connectivity-design tasks. On defective layouts, the discovered subsystem codes achieve higher code and circuit fault distances than native surface-code baselines, while reducing CZ gate count and depth relative to distance-matched routed implementations. Additional experiments demonstrate improvements in logical-qubit capacity, code distance, and connectivity requirements under specified hardware constraints.

Critical strengths
  • Verification-grounded LLM framework for hardware-aware code discovery. For the automated discovery, one of the crucial thing is to verify and evaluate results. The paper has done a careful analysis on this aspect.
Critical weaknesses
  • Restricted search space and limited generality. Although the framework is presented as an automated approach to subsystem-code discovery, it is restricted to CSS codes with separate X- and Z-type gauge generators, without considering general non-CSS subsystem codes. Moreover, several of the strongest reported constructions are based on known Bacon–Shor code families or stabilizer codes with r=0r=0.

  • Lack of ablation studies to validate the proposed self-evolving framework. A central contribution of the paper is the iterative discovery process driven by verification feedback and accumulated search memory. However, the paper does not provide controlled ablation studies to isolate the contributions of these components. In particular, comparisons with LLM search without feedback or memory, as well as conventional evolutionary or random search under matched computational budgets, are missing. Although the reported search trajectories show improvements over successive rounds, they do not establish whether these gains arise from the proposed feedback mechanism or simply from evaluating more candidates. This makes it difficult to assess the necessity and effectiveness of the key methodological contributions.

Questions for authors

See the weakness part: could the authors elaborate more on these two points?

7 comments

W1: search-space generality and novelty of subsystem-code design

We thank the reviewer for the constructive comments on the generality of our framework. We appreciate the opportunity to clarify the scope of our search space and the interpretation of the resulting code constructions.

Generality of CSS subsystem codes. Our current framework focuses on CSS subsystem codes, where X- and Z-type gauge generators are not required to commute with each other. While this does not cover general non-CSS subsystem codes, it already provides a large and nontrivial design space. Rather than restricting candidates to predefined code families, our framework jointly explores gauge generators, physical-qubit assignments, measurement layouts, and hardware connectivity across up to 105 physical qubits. These variables are strongly coupled through algebraic validity, logical protection, and hardware feasibility. The CSS setting also offers practical advantages in syndrome extraction and decoding, making it a natural starting point for hardware-aware subsystem-code design. Extending the framework to general non-CSS subsystem codes through a full binary symplectic representation is an interesting direction for future research.

Known code families versus search generality. We agree that several strong constructions identified by our framework belong to well-established code families, including the Bacon–Shor codes in the four defective-chip experiments. This outcome is consistent with the physical characteristics of the considered hardware. On highly local superconducting architectures, Bacon–Shor codes offer geometrically structured constructions with low-weight gauge measurements, making them particularly suitable candidates. Importantly, these code families were not imposed as search restrictions, but emerged from a broader search over gauge structures and hardware realizations. The custom-layout experiments further demonstrate the versatility of our approach, producing different constructions such as ([[50,2,4,7]]) and ([[49,1,12,9]]) under distinct connectivity constraints. These results highlight the value of automated search in identifying suitable code constructions for different hardware objectives.

Inclusion of stabilizer codes. Stabilizer codes are naturally included in the subsystem formalism as the special case (r=0). We retain this possibility to allow the search to explore different gauge structures according to the specified objective. For example, Claude identifies a ([[70,10,0,5]]) stabilizer code in the Capacity task, protecting ten logical qubits at distance five using 100 physical qubits and maximum connectivity degree six. This result illustrates how the framework accommodates both subsystem and stabilizer constructions within a unified hardware-aware search process.

W2: classical search baselines and verification-feedback ablation

We thank the reviewer for this important suggestion. We agree that search trajectories alone are insufficient to establish the effectiveness of the proposed self-evolving mechanism, as improvements may also arise from evaluating more candidates. To address this concern, we conducted two complementary sets of experiments: comparisons with conventional non-LLM search methods and ablations of verification-driven feedback.

Classical search baselines. We introduce three conventional search strategies: structured random search, logical-guided simulated annealing, and mutation-population search. All three methods operate under the same hardware constraints and use the same deterministic code verifier as our framework. A short CP-SAT initialization provides generators satisfying the initial algebraic constraints. Each algorithm is evaluated with three independent random seeds, with approximately 12 CPU-hours shared across the nine runs per setting. The elapsed search time is comparable to that of our LLM searches. Across all seven settings, these classical approaches collectively generated 6,563,593 candidate proposals.

SettingGPT: fullDeepSeek: fullBest classical
2C[[49,1,36,7]][[49,1,36,7]][[24,1,0,3]]
4C[[49,1,36,7]][[49,1,36,7]][[19,1,0,3]]
9C[[25,1,16,5]][[28,1,18,4]][[49,1,0,3]]
2Q + 2C[[36,1,25,6]][[36,1,25,6]][[39,1,0,3]]
Protection (Δ≤5\Delta \leq 5)[[54,2,38,6]][[32,2,18,4]][[14,1,0,3]]
Capacity (Δ≤6\Delta \leq 6)[[75,7,52,5]][[75,7,52,5]][[13,1,6,3]]
Connectivity (d≥9d \geq 9)[[49,1,12,9]], Δ=4\Delta=4[[63,1,4,9]], Δ=6\Delta=6[[23,1,15,3]]

The classical baselines achieve a maximum code distance of only three across all four defective-chip settings, whereas our LLM-guided framework reaches distances of seven, seven, five, and six, respectively. Furthermore, the classical searches fail to meet the prescribed protection targets in the three custom-layout tasks. These results demonstrate the effectiveness of LLM-guided program search over the conventional strategies evaluated here. Unlike predefined mutation operators, LLM-generated construction programs can coordinate structural changes to gauge generators, physical-qubit roles, measurement assignments, and hardware connectivity, providing greater flexibility in exploring this coupled design space.

Verification-feedback ablation. To examine whether iterative refinement contributes beyond independent LLM-based generation, we compare the full framework against no-feedback sampling using GPT and DeepSeek under the same task settings. The full framework incorporates previously verified constructions, code properties, and failure diagnostics into subsequent generations, whereas the no-feedback variant generates candidates independently without access to previous search outcomes.

SettingGPT: fullGPT: no feedbackDeepSeek: fullDeepSeek: no feedback
2C[[49,1,36,7]][[49,1,36,7]][[49,1,36,7]][[49,1,36,7]]
4C[[49,1,36,7]][[49,1,36,7]][[49,1,36,7]][[49,1,36,7]]
9C[[25,1,16,5]][[16,1,9,4]][[28,1,18,4]][[9,1,4,3]]
2Q + 2C[[36,1,25,6]][[48,1,35,6]][[36,1,25,6]][[48,1,35,6]]
Protection (Δ≤5\Delta \leq 5)[[54,2,38,6]][[42,2,28,5]][[32,2,18,4]][[26,2,0,3]]
Capacity (Δ≤6\Delta \leq 6)[[75,7,52,5]][[75,3,48,5]][[75,7,52,5]][[50,2,32,5]]
Connectivity (d≥9d \geq 9)[[49,1,12,9]], Δ=4\Delta=4[[81,1,64,9]], Δ=6\Delta=6[[63,1,4,9]], Δ=6\Delta=6[[81,1,64,9]], Δ=8\Delta=8

The results show clear improvements on the more challenging tasks. For GPT, verification feedback increases the distance from four to five on 9C and from five to six on Protection, improves logical capacity from three to seven on Capacity at (d=5), and reduces the maximum connectivity degree from six to four on Connectivity at (d=9). DeepSeek exhibits similar improvements across these tasks. On 2Q+2C, both models achieve (d=6), while the full framework identifies constructions using fewer data qubits. For 2C and 4C, both approaches reach the native-connectivity distance bound of seven, leaving no room for improvement in the primary objective.

Role of the self-evolving mechanism. The additional experiments provide complementary evidence for the two central components of our framework. The classical baselines highlight the effectiveness of LLM-guided construction search, while the no-feedback ablation demonstrates the benefits of incorporating verified search outcomes into subsequent generations. Although the current ablation evaluates feedback and accumulated search history jointly rather than isolating each component, the results provide direct empirical support for the proposed iterative discovery mechanism. We appreciate the reviewer's constructive suggestions, which have helped us strengthen the experimental validation of our framework.

Reviewer_CEJYOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a framework based on LLMs for the automated discovery of error-correcting codes for quantum subsystems. It provides a method for exploring subsystem error-correcting codes which fit the physical connectivity characteristics and measurement limitations of quantum chips. Experimental results show this method can identify error-correcting code implementations that feature larger code distances, lower circuit overhead, or less stringent hardware connectivity requirements.

Critical strengths
  1. This research focuses on the constraints imposed by physical connectivity on specific superconducting chips and proposes an algorithm that simultaneously optimizes gauge measurements, qubit roles, and physical connectivity.

  2. The verification mechanism is quite robust. The authors use deterministic algebraic tools, enumeration, MILP, and fault distances for certain quantum circuits to verify the results. Additionally, the implementation includes caching, failure diagnosis, search memoization, and limits on program execution.

Critical weaknesses
  1. Lack of a search baselines which do not use an LLM. The authors use GPT, DeepSeek, and Claude to compare the correction code LLMs found, but all of them are LLMs, which does not fully answer the question why users cannot use genetic algorithms, simulated annealing, local search, or CP-SAT all used to accomplish the same task. The unique contributions of LLMs need stronger evidence.

  2. The results regarding the actual logical error rate are not sufficiently convincing. Some results show a higher code distance and lower CZ-gate overhead did not translate into superior logical storage performance. Also, the paper does not provide a corresponding logical error rate curve for a standard surface code for comparison.

Questions for authors

The paper only uses LLMs to generate and optimize quantum error-correcting codes, which makes it lack a comparison with traditional methods such as genetic algorithms or random search. Are there specific advantages of using LLMs over automatic non-LLM methods?

The paper includes iterative improvement of candidate codes through a verification-based feedback mechanism. However, it lacks corresponding ablation studies for feedback mechanisms. And to what extent it improves results compared to search methods that do not utilize such feedback?

8 comments

W1&Q1: classical search baselines and the necessity of LLM-guided search.

We thank the reviewer for this important suggestion. We agree that comparisons among different LLM backends alone are insufficient to establish the advantages of LLM-guided search over conventional optimization methods. To address this concern, we conducted additional experiments with three non-LLM search baselines.

Classical search baselines. We introduce three conventional search strategies: structured random search, logical-guided simulated annealing, and mutation-population search. All three methods operate under the same hardware constraints and use the same deterministic code verifier as our framework. A short CP-SAT initialization provides generators satisfying the initial algebraic constraints. Each algorithm is evaluated with three independent random seeds, with approximately 12 CPU-hours shared across the nine runs per setting. The elapsed search time is comparable to that of our LLM searches. Across all seven settings, these classical approaches collectively generated 6,563,593 candidate proposals.

SettingGPT: fullDeepSeek: fullBest classical
2C[[49,1,36,7]][[49,1,36,7]][[24,1,0,3]]
4C[[49,1,36,7]][[49,1,36,7]][[19,1,0,3]]
9C[[25,1,16,5]][[28,1,18,4]][[49,1,0,3]]
2Q + 2C[[36,1,25,6]][[36,1,25,6]][[39,1,0,3]]
Protection (Δ≤5\Delta \leq 5)[[54,2,38,6]][[32,2,18,4]][[14,1,0,3]]
Capacity (Δ≤6\Delta \leq 6)[[75,7,52,5]][[75,7,52,5]][[13,1,6,3]]
Connectivity (d≥9d \geq 9)[[49,1,12,9]], Δ=4\Delta=4[[63,1,4,9]], Δ=6\Delta=6[[23,1,15,3]]

Experimental results. The classical baselines achieve a maximum code distance of only three across all four defective-chip settings, whereas our LLM-guided framework reaches distances of seven, seven, five, and six, respectively. These results demonstrate a substantial advantage of LLM-guided search over the conventional strategies evaluated here.

Advantages of LLM-guided search. Importantly, our task involves a large, coupled design space rather than optimizing a few parameters within a predefined code family. It requires jointly determining physical-qubit roles, X/Z gauge generators, measurement-ancilla assignments, and, for custom-layout tasks, hardware connectivity. These decisions directly affect both hardware feasibility and the resulting logical protection. Unlike conventional search strategies relying on predefined proposal and mutation operators, our LLM-based approach operates at the level of executable construction programs, enabling coordinated structural modifications across these interdependent design variables. The experimental results suggest that this flexibility, together with the LLM's ability to leverage prior knowledge of code constructions, provides an effective approach to navigating the hardware-constrained subsystem-code design space.

Q2: effectiveness of verification-driven feedback

We thank the reviewer for highlighting the importance of validating our self-evolving mechanism. To isolate the contribution of verification-driven feedback, we conducted additional ablation experiments comparing the full framework against independent no-feedback sampling using GPT and DeepSeek under the same task settings.

Feedback ablation. In the full framework, previously verified constructions, code properties, and failure diagnostics are incorporated into subsequent generations, allowing the LLM to progressively refine its construction strategies. In contrast, the no-feedback variant generates candidates independently without access to previous search outcomes. This comparison evaluates whether iterative refinement provides additional benefits beyond the LLM's intrinsic code-generation capabilities.

SettingGPT: fullGPT: no feedbackDeepSeek: fullDeepSeek: no feedback
2C[[49,1,36,7]][[49,1,36,7]][[49,1,36,7]][[49,1,36,7]]
4C[[49,1,36,7]][[49,1,36,7]][[49,1,36,7]][[49,1,36,7]]
9C[[25,1,16,5]][[16,1,9,4]][[28,1,18,4]][[9,1,4,3]]
2Q + 2C[[36,1,25,6]][[48,1,35,6]][[36,1,25,6]][[48,1,35,6]]
Protection (Δ≤5\Delta \leq 5)[[54,2,38,6]][[42,2,28,5]][[32,2,18,4]][[26,2,0,3]]
Capacity (Δ≤6\Delta \leq 6)[[75,7,52,5]][[75,3,48,5]][[75,7,52,5]][[50,2,32,5]]
Connectivity (d≥9d \geq 9)[[49,1,12,9]], Δ=4\Delta=4[[81,1,64,9]], Δ=6\Delta=6[[63,1,4,9]], Δ=6\Delta=6[[81,1,64,9]], Δ=8\Delta=8

Experimental results. The full framework consistently achieves better results on the more challenging search tasks. For GPT, feedback improves the code distance from 4 to 5 on 9C and from 5 to 6 on Protection, increases the number of logical qubits from 3 to 7 on Capacity while maintaining (d=5), and reduces the maximum connectivity degree from 6 to 4 on Connectivity while maintaining (d=9). DeepSeek exhibits similar improvements, with distance increasing from 3 to 4 on both 9C and Protection, logical capacity increasing from 2 to 7, and maximum degree decreasing from 8 to 6. On 2Q+2C, both models retain (d=6) while reducing the number of data qubits from 48 to 36. For the simpler 2C and 4C tasks, both approaches reach the native-connectivity upper bound of (d=7), leaving no room for further improvement in the primary objective.

Role of iterative refinement. These results demonstrate that the improvements observed across search rounds cannot be attributed solely to generating additional independent candidates. By incorporating verified search outcomes into subsequent program generation, our framework can progressively refine code structures and hardware realizations. The ablation therefore provides direct experimental support for the effectiveness of the proposed verification-driven, self-evolving search mechanism.

W2: logical error rates and surface code comparison

We thank the reviewer for raising this important concern. We would like to clarify why our work primarily adopts distance-based evaluation rather than relying exclusively on logical error rates (LERs), and present additional circuit-level experiments that directly address the requested comparison.

Limitations of Pauli-based simulations. Conventional Stim-based LER simulations rely on stochastic Pauli noise models and corresponding detector error models (DEMs). Although computationally efficient, these models do not fully capture important hardware effects, including coherent errors, non-Pauli decoherence dynamics, leakage, and context-dependent crosstalk. Such limitations are particularly relevant to subsystem codes, whose low-weight gauge measurements may reduce circuit complexity, measurement duration, and exposure to hardware imperfections. Consequently, their practical advantages may not be fully reflected by conventional Pauli-based LER benchmarks.

Implementation dependence of LER. Circuit-level LER is not an intrinsic property of a quantum code. Its evaluation depends critically on three components: syndrome-extraction circuits, physical noise models, and decoding algorithms. Surface codes have benefited from extensive research and highly optimized implementations of these components. In contrast, general subsystem codes lack comparably mature and standardized extraction architectures and decoders. As a result, LER comparisons may reflect differences in implementation maturity rather than solely the underlying code properties.

Why we use distance-based evaluation. These considerations motivate our choice to evaluate logical protection primarily through certified code distance and circuit fault distance, complemented by hardware resource metrics such as CZ count and circuit depth. Code distance characterizes the minimum weight of nontrivial logical operators, while circuit fault distance further accounts for fault propagation through the specified extraction circuit. Both metrics can be evaluated without relying on a particular decoder or numerical physical error rate. Although they do not fully determine practical LER, they provide a principled and verifiable basis for evaluating automatically discovered quantum codes.

Additional circuit-level comparisons. To directly address the reviewer's concern, we conducted additional logical-memory experiments comparing our discovered codes against code-distance-matched routed surface codes on the same defective hardware layouts. Both implementations are evaluated under the same superconducting-inspired SI1000 noise model at (p=0.001), with both idle scaling factors set to one. We employ the same Tesseract decoder (beam size 10, queue limit 20,000) and storage duration of (T=d) measurement rounds, using 5,000 independent shots for each X- and Z-basis memory experiment. The resulting logical-memory failure probabilities are summarized below.

SettingOur code [[n,k,r,d]]Common d / TOurs: X / ZRouted surface: X / Z
2C[[49,1,36,7]]7 / 70.88% / 0.92%1.18% / 1.14%
4C[[49,1,36,7]]7 / 71.00% / 0.80%2.06% / 1.82%
9C[[25,1,16,5]]5 / 50.58% / 0.42%3.30% / 4.02%
2Q + 2C[[36,1,25,6]]6 / 61.18% / 1.14%5.04% / 4.98%

Under this common evaluation protocol, our implementations achieve lower observed logical-memory failure probabilities in both memory bases across all four defective-chip settings. These results provide direct circuit-level evidence that the advantages in distance and hardware resources can also translate into improved logical-memory performance, while broader evaluations under device-specific noise models remain an important direction for future work.

Reviewer_eceqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper builds an LLM-driven search loop for hardware-constrained subsystem codes. An orchestrator hands a structured task to a design agent, which writes executable construction programs: each program assigns physical qubits to data, measurement-ancilla and unused roles, then writes down X- and Z-type gauge generators together with their ancilla supports. Deterministic tools then derive the stabilizer center, compute the gauge and logical counts, and return certified lower bounds plus constructive upper bounds on the dressed distance before a feasibility filter checks the couplers and ranks the survivors. Verified properties and failure evidence are fed back into the next round. The evaluation covers four calibration-informed defect masks on the 105-qubit Willow graph and three connectivity-design tasks. On the defect masks the selected GPT-generated constructions are square Bacon-Shor realizations with weight-two gauges, reported at dressed distances 7, 7, 5 and 6 against 5, 4, 3 and 4 for the native surface-code templates; the circuit fault distance dc is also reported, together with a 12.5-41.2% reduction in per-cycle CZ count and a 70.0-86.1% reduction in CZ depth relative to distance-matched routed surface codes. Under Willow augmentation the search returns a [[50, 2, 4, 7]] subsystem code at maximum degree 5, [[75, 7, 52, 5]] constructions at degree 6, a [[70, 10, 0, 5]] stabilizer code, and a [[49, 1, 12, 9]] construction at maximum degree 4 for the free-connectivity task.

Critical strengths

1The architecture puts the LLM only where it belongs. The model proposes programs; every reported number - the code parameters, the stabilizer center, the distance bounds - comes from exact linear algebra over F2 and from MILP or UNSAT certificates. That removes the usual failure mode of LLM-for-science papers, where a plausible-looking claim is taken at face value. 2Hardware conditioning is real rather than rhetorical. The feasibility filter is defined on the surviving coupler set, the output is a qubit-role assignment on a named processor, and the installed edge set is distinguished from the native one. This is the difference between designing a code and designing a layout that a fab can actually run. 3The distance evidence is reported the right way: a lower bound from completed exclusion plus an upper bound from an explicit witness, with exactness claimed only when the two coincide. Section D and Appendix F.4 are unusually explicit about where the certificates stop. 4For the defect experiments, the informative comparison is against a routed surface code matched to the discovered distance, and that is the comparison the paper makes. Reporting CZ count, CZ depth and SWAP counts under a common compilation is the right currency for a hardware paper.

Critical weaknesses

1Nothing isolates the LLM. The claim is that an LLM-driven search is the right way to explore this space, but there is no control in which a classical searcher is given the same candidate-evaluation budget. Simulated annealing, a genetic algorithm, or a CP-SAT encoding over qubit-role assignments would all be plausible competitors, and the construction space is small enough that they are practical. Without that arm the paper demonstrates that search works, not that language-model search is what makes it work. Given the token and API costs in Figure 4e, a cost-matched classical baseline is the single most valuable experiment to add. 2The discovered codes are textbook constructions. All four defect-mask results are square Bacon-Shor codes - [[49, 1, 36, 7]] is the 7 x 7 Bacon-Shor, [[25, 1, 16, 5]] the 5 x 5 - and the paper says so explicitly. What the search found is an embedding of a known family on a defective graph, which is a legitimate combinatorial problem but a much narrower claim than the title suggests. The novelty should be stated as defect-aware layout of subsystem codes, not as discovery of codes. 3The protection claim rests on geometry, not on logical error rates. Everything supports distance and gate counts. The one circuit-level experiment (Appendix G.2) covers a single code, five cycles, 105 shots per point, physical error rates up to 0.6%, and no surface-code baseline; Figure 7 shows logical failure probabilities reaching 0.5. At p = 0.006 a distance-5 surface code over five rounds is many orders of magnitude below that. If those numbers are right, the discovered codes may be far worse than the baseline they claim to beat, and the distance advantage does not translate. Either add the baseline at matched p and cycle count, or scope the abstract's protection claim to distance and resources. 4dc = d for a Bacon-Shor with bare weight-two gauges needs an explanation. In a weight-two gauge measurement, a fault on the ancilla between its two CZs propagates to a correlated two-qubit data error, and the standard consequence for Bacon-Shor is a hook-induced reduction of the circuit-level distance. The certificates in Appendix F.4 report dc = 7, which means the compiled schedule must be hook-free for these particular circuits. Which ordering and ancilla scheme achieves that, and does the same scheme survive the 4- and 9-coupler masks? 5Best-of-search and best-of-model selection is not corrected for. The headline uses the four GPT constructions that attain the best dressed distance for each mask, and the custom-layout margins over the baselines are sometimes a few tenths of a percent. With three models, seven tasks and many rounds, the reported maxima are order statistics. The paper should report the spread across runs (same prompt, same model, different seed) and say whether the headline numbers are typical or best.

Questions for authors

1.What does a classical search (annealing, GA, or CP-SAT over the same role assignments) achieve with the same number of candidate evaluations and the same wall-clock? 2.Given that the discovered codes are Bacon-Shor, what exactly is the new object - the code family, the layout, or the search procedure? 3.Can you report circuit-level logical error rates for the discovered d = 7 code and for a distance-matched surface code at the same physical error rate and the same number of rounds? 4.How does dc = 7 survive hook errors from weight-two gauge extraction? Please give the schedule and, if flags are used, their cost. 5.How much does the best dressed distance vary across repeated runs of the same model and prompt?

9 comments

W1&Q1: classical search baselines and the necessity of LLM-guided search.

We thank the reviewer for this important suggestion. We agree that comparing different LLM backends alone is insufficient to establish the advantages of LLM-guided search over conventional optimization methods. To address this concern, we conducted additional experiments with three non-LLM baselines and an ablation study of our verification-feedback mechanism.

Classical search baselines. We introduce three conventional search strategies: structured random search, logical-guided simulated annealing, and mutation-population search. All three methods operate under the same hardware constraints and use the same deterministic code verifier as our framework. A short CP-SAT initialization provides generators satisfying the initial algebraic constraints. Each algorithm is evaluated with three independent random seeds, with approximately 12 CPU-hours shared across the nine runs per setting. The elapsed search time is comparable to that of our LLM searches. Across all seven settings, these classical approaches collectively generated 6,563,593 candidate proposals.

Verification-feedback ablation. To further investigate whether the self-evolving mechanism contributes beyond LLM-based generation alone, we compare our full framework against independent no-feedback sampling using GPT and DeepSeek under the same task settings. This experiment evaluates the benefit of incorporating previously verified constructions, code properties, and failure diagnostics into subsequent generations, rather than independently generating candidates without iterative feedback.

SettingGPT: fullGPT: no feedbackDeepSeek: fullDeepSeek: no feedbackClassical search
2C[[49,1,36,7]][[49,1,36,7]][[49,1,36,7]][[49,1,36,7]][[24,1,0,3]]
4C[[49,1,36,7]][[49,1,36,7]][[49,1,36,7]][[49,1,36,7]][[19,1,0,3]]
9C[[25,1,16,5]][[16,1,9,4]][[28,1,18,4]][[9,1,4,3]][[49,1,0,3]]
2Q + 2C[[36,1,25,6]][[48,1,35,6]][[36,1,25,6]][[48,1,35,6]][[39,1,0,3]]
Protection (Δ ≤ 5)[[54,2,38,6]][[42,2,28,5]][[32,2,18,4]][[26,2,0,3]][[14,1,0,3]]
Capacity (Δ ≤ 6)[[75,7,52,5]][[75,3,48,5]][[75,7,52,5]][[50,2,32,5]][[13,1,6,3]]
Connectivity (d ≥ 9)[[49,1,12,9]], Δ=4[[81,1,64,9]], Δ=6[[63,1,4,9]], Δ=6[[81,1,64,9]], Δ=8[[23,1,15,3]]

Comparison with classical search. The classical baselines achieve a maximum distance of only three across all four defective-chip settings, whereas our LLM-guided framework reaches distances of seven, seven, five, and six, respectively. This substantial gap demonstrates the effectiveness of LLM-guided program search in navigating the coupled design space of gauge structures and hardware realizations, compared with the conventional search strategies evaluated here.

Effectiveness of verification feedback. The ablation results further demonstrate the benefits of iterative refinement. For the simpler 2C and 4C settings, both the full framework and no-feedback sampling achieve the native-connectivity distance bound of seven, leaving no room for further improvement in the primary objective. In contrast, feedback produces clear gains on more challenging tasks. For GPT, the distance increases from four to five on 9C and from five to six on Protection, the logical capacity increases from three to seven on Capacity, and the maximum connectivity degree decreases from six to four on Connectivity while maintaining distance nine. DeepSeek exhibits similar improvements across these tasks. On 2Q+2C, both approaches achieve distance six, but the full framework identifies constructions with fewer data qubits. These results suggest that verification-driven feedback is particularly valuable for more challenging search objectives, where independent generation alone does not reach the strongest solutions.

W2&Q2: novelty of automated subsystem-code discovery.

We respectfully disagree with the implication that discovering known Bacon–Shor constructions reduces our contribution to embedding a predefined code family. The search space and the resulting solutions are fundamentally different concepts. Our framework does not restrict the search to Bacon–Shor codes or any predefined list of code families. Instead, it explores the broad space of CSS subsystem codes by jointly designing gauge generators, physical-qubit assignments, measurement layouts, and hardware connectivity. The fact that some optimal or competitive search outcomes belong to a known family does not imply that the search space itself is restricted. To the best of our knowledge, our work introduces the first LLM-guided framework specifically targeting hardware-constrained subsystem-code discovery, where both code constructions and their physical realizations emerge from an automated, verification-driven search.

The emergence of Bacon–Shor codes in the defective-chip experiments is also physically meaningful. On highly local, two-dimensional superconducting architectures, surface codes are particularly well suited to the native connectivity, and Bacon–Shor codes similarly exploit regular geometric structures while enabling low-weight gauge measurements. It is therefore reasonable that Bacon–Shor constructions emerge as strong candidates under these hardware constraints. Importantly, this family was not imposed in advance but emerged from the search. Moreover, our results are not limited to Bacon–Shor codes. The custom-layout experiments produce diverse constructions, including a ([[50,2,4,7]]) subsystem code, a ([[75,7,52,5]]) subsystem code, and a ([[49,1,12,9]]) code with maximum connectivity degree four. These results demonstrate that our framework can explore different code structures and adapt them to distinct hardware design objectives, rather than simply optimizing the placement of a single known family.

Nevertheless, we appreciate that the term "discovery" in the original title could be interpreted as claiming the invention of entirely new algebraic code families. To avoid this ambiguity, we will revise the title to "Automated Search for Quantum Subsystem Codes with an AI Scientist" and clarify that our primary contributions lie in the LLM-guided search methodology and the verified, hardware-adapted code constructions it produces.

W3&Q3: circuit-level LERs and surface code comparisons.

We appreciate the reviewer's concern regarding logical error rates (LERs). However, we would like to clarify why LER alone is not sufficient to characterize the practical advantages of quantum error-correcting codes, particularly subsystem codes.

Limitations of Pauli-based simulations. Conventional Stim-based LER simulations rely on stochastic Pauli noise models and the corresponding detector error models (DEMs). While computationally efficient, such models do not fully capture important hardware effects, including non-Pauli decoherence dynamics, coherent errors, leakage, and context-dependent crosstalk. These limitations are particularly relevant to subsystem codes, whose low-weight gauge measurements can potentially reduce measurement complexity, circuit duration, and exposure to hardware imperfections. Consequently, their practical benefits may not be fully reflected by conventional Pauli-noise-based LER benchmarks.

Implementation dependence of LER. Circuit-level LER is not an intrinsic property of a quantum code. Its evaluation requires three closely coupled components: syndrome-extraction circuits, a physical noise model, and an appropriate decoder. Surface codes have benefited from extensive research and highly optimized implementations of all three components. In contrast, general subsystem codes still lack comparably mature and standardized syndrome-extraction architectures and decoding algorithms. Consequently, an LER comparison may reflect differences in the maturity of these supporting components rather than the intrinsic protection capabilities of the underlying codes. This makes LER an important but implementation-dependent metric, rather than a universal criterion for code discovery.

Why we use distance-based evaluation. These considerations motivate our choice to evaluate logical protection primarily through certified code distance and circuit fault distance. Code distance characterizes the minimum weight of nontrivial logical operators, while circuit fault distance further accounts for fault propagation through a specified syndrome-extraction circuit. Unlike LER, these distance-based measures do not depend on the performance of a particular decoder or an assumed numerical physical error rate. We acknowledge that distance alone does not determine practical LERs; nevertheless, it provides a principled and verifiable basis for evaluating automatically discovered codes without confounding the comparison with potentially unequal levels of circuit and decoder optimization.

Additional circuit-level comparisons. To directly address the reviewer's concern, we conducted additional circuit-level logical-memory experiments comparing our discovered codes against code-distance-matched routed surface codes on the same defective hardware layouts. Both implementations are evaluated under the same superconducting-inspired SI1000 noise model at (p=0.001), with both idle scaling factors set to one. We employ the same Tesseract decoder (beam size 10, queue limit 20,000) and the same storage duration of (T=d) measurement rounds, using 5,000 independent shots for each X- and Z-basis memory experiment. In particular, the two distance-seven cases (2C and 4C) are compared over seven measurement rounds. The resulting logical-memory failure probabilities are summarized below.

SettingOur code [[n,k,r,d]]Common d / TOurs: X / ZRouted surface: X / Z
2C[[49,1,36,7]]7 / 70.88% / 0.92%1.18% / 1.14%
4C[[49,1,36,7]]7 / 71.00% / 0.80%2.06% / 1.82%
9C[[25,1,16,5]]5 / 50.58% / 0.42%3.30% / 4.02%
2Q + 2C[[36,1,25,6]]6 / 61.18% / 1.14%5.04% / 4.98%

Under this common evaluation protocol, our discovered implementations achieve lower observed logical-memory failure probabilities than the corresponding routed surface-code baselines in both X and Z bases across all four defective-chip settings. These additional results demonstrate that the advantages identified through distance and circuit-resource analysis can also translate into improved logical-memory performance under a matched noise and decoding protocol, while recognizing that broader hardware-specific assessments remain an important direction for future work.

Comments on the paper

1 comment

Dear Area Chair,

We sincerely thank you and the reviewers for your time, careful evaluation, and constructive feedback. We also appreciate the reviewers’ recognition of the strengths of our work. Reviewer eceq noted that “Hardware conditioning is real rather than rhetorical,” Reviewer CEJY described our verification mechanism as “quite robust,” and Reviewer dqCE acknowledged our “careful analysis” of verification and evaluation.

In our rebuttal, we have provided detailed, point-by-point responses to all ten main concerns raised across the three reviews. To address these concerns, we conducted four sets of additional experiments: comparisons with three non-LLM search baselines, verification-feedback ablations, circuit-level logical-memory comparisons against distance-matched routed surface codes, and independent-run reproducibility studies. We also clarified the scope and novelty of our framework and provided further details on syndrome-extraction schedules and circuit fault-distance certification. We have revised the manuscript accordingly and uploaded the updated PDF.

With approximately 2 days remaining in the discussion period, we would be grateful for your help in encouraging the reviewers to let us know whether our responses and additional evidence have satisfactorily addressed their concerns. If any questions remain, we are happy and ready to provide further answers and clarifications during the remaining discussion time.

Thank you again for your time and consideration.

Best regards,
The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.