{"id":"7203abf2-3f08-4f66-9a48-a03088974586","arxiv_id":"2509.02683","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A random-forest model that picks non-uniform error budget splits reduces estimated fault-tolerant quantum computation costs by an average of 15.6% compared to a uniform split.","lead":"This paper trains a machine learning model to choose how to split a quantum circuit's total allowed error across three error sources, instead of using the usual equal split. The chosen splits cut estimated hardware costs by an average of 15.6% in simulations for over 75% of benchmark circuits.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported savings are for a per-circuit minimum of the ML prediction and the uniform baseline, so the headline claim about ML-predicted distributions instead of uniform is not yet supported.","rationale":"The load-bearing condition for the strongest_claim is that the evaluation isolates the effect of the ML-predicted distributions. It does not: the protocol chooses the lower-cost candidate per circuit, so the reported reduction is an upper bound on the ML-only improvement and masks any samples where the prediction is worse than uniform. This is more direct than the additive error-budget concern raised by the reader; an additive allocation with sum equal to the budget is conservative under the union bound, so it does not by itself break the 10% budget. The missing fixed-heuristic baseline also matters, but the min-selection issue is prior because it affects the headline number itself. I keep the reader's CONDITIONAL verdict because the concern can be settled by a straightforward re-analysis; the condition should explicitly require reporting ML-only metrics before the abstract's 'instead of a uniform distribution' wording can be accepted.","tokens_in":9744,"tokens_out":11274,"duration_ms":110042,"concrete_test":"Re-run the Section V-C evaluation on the same 383 test circuits with the released code, but do not take the minimum with uniform: estimate resources for the uniform distribution and for the normalized ML-predicted distribution only, for each total error budget (0.1%, 1%, 10%). Report the mean space-time reduction and fraction of improved circuits for the ML-only predictions, including the full distribution of differences with negative values, alongside the current min-selected results. If the ML-only mean and fraction remain close to 15.6% and 75%, the concern is resolved; if not, the abstract and conclusion must attribute the improvement to the hybrid selection strategy or be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section V-A states: 'the error budget distribution with the lowest estimated space-time cost was chosen from the machine learning model's predictions and a uniform distribution.' Each reported reduction is therefore (uniform - min(uniform, ML))/uniform, which is never negative, and Figure 4 contains no negative-difference bars. The 15.6% average and 'over 75%' fraction describe a hybrid strategy that may discard an unfavorable ML prediction, not 'machine-learning-predicted error budget distributions instead of a uniform distribution' as claimed in the abstract. For the circuits counted as 'without improvement', the ML-only prediction may actually be worse than uniform, and that loss is masked by the selection. This directly affects the strongest_claim: the data do not show how much of the reduction is due to the ML model itself. The additive error-budget assumption is less decisive: if the three budgets are per-part event probabilities, the union bound makes their sum a conservative upper bound on the total failure probability, so the 10% budget is not silently exceeded; the only error is second-order conservatism.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses the problem of allocating a user-specified total error budget among three parts of a fault-tolerant quantum circuit—logical qubit implementation, T-state production, and rotation-gate approximation—in order to reduce estimated physical resource costs (space-time product). The authors sample 1000 random error-budget distributions per circuit for 1530 circuits from MQT Bench, use Microsoft's Azure Quantum Resource Estimator to label each circuit with the best sampled distribution, and train a random forest to predict a good distribution from logical gate counts. They evaluate on a held-out test set of 383 circuits across total error budgets of 0.1%, 1%, and 10%, reporting that the estimated space-time cost is reduced for over 75% of circuits with an average reduction of 15.6% and a maximum reduction of 77.7% compared to a uniform distribution. They also extract general guidelines on how to distribute the budget and make the implementation available on GitHub.","tokens_in":9950,"tokens_out":6155,"duration_ms":62914,"significance":"If the reported evaluation were correctly attributed, the paper would make a useful practical contribution to fault-tolerant quantum resource estimation: it shows that the error-budget distribution is a meaningful degree of freedom and provides an automated way to exploit it, backed by a reasonably large dataset, a held-out test split, three total-budget settings, and public code. The guidelines and the dataset are also potentially reusable by other researchers. However, the central quantitative claim is currently not supported as stated because the evaluation protocol selects the cheaper of the ML prediction and the uniform baseline for every test circuit; therefore the reported reductions describe a hybrid strategy, not the ML model alone. This is a fixable issue, but it affects the headline result and needs to be addressed before the paper can be accepted.","major_comments":[{"comment":"The evaluation protocol in Section V-A states 'the error budget distribution with the lowest estimated space-time cost was chosen from the machine learning model's predictions and a uniform distribution.' This means that every reported improvement is computed as (uniform - min(uniform, ML))/uniform, which is never negative and which discards the ML prediction whenever it is worse than uniform. The histograms in Figure 4 therefore contain no negative-difference bars for the predicted distributions, and the reported 'over 75%' and '15.6% average reduction' describe a hybrid strategy rather than 'machine-learning-predicted error budget distributions instead of a uniform distribution' as claimed in the abstract. Please report the distribution of (cost(uniform) - cost(ML-only))/cost(uniform) for every test circuit, including negative values, and give the mean, median, and fraction of improving circuits for the ML-only predictions. The abstract and Section V-C should be revised to describe the evaluation honestly, or the evaluation should be changed to compare ML-only predictions against the uniform baseline without per-circuit selection.","section":"V-A, V-C, and Abstract"},{"comment":"The paper does not compare the ML predictions against a simple fixed non-uniform heuristic. The guidelines derived in Section V-B (small logical budget, larger T-state and rotation budgets) can be instantiated as a constant distribution, such as the mean optimized distribution over the training set, and applied to all test circuits. Without this baseline, the observed gains may be attributable simply to shifting away from a uniform split rather than to per-circuit tailoring by the ML model. Please add a fixed-heuristic baseline and report paired comparisons (ML-only versus fixed-heuristic, and fixed-heuristic versus uniform) so that the value added by machine learning per circuit is quantified.","section":"V-B and V-C"}],"minor_comments":[{"comment":"The additive error-budget model ('The sum of these individual error budgets refers to the total error budget') is stated without justification. If the three budgets are per-part error probabilities, the sum is a conservative upper bound on the total failure probability by the union bound, so the 10% budget case does not silently exceed the user's budget; please state this interpretation explicitly so that the model is not misread as an exact equality.","section":"II-C"},{"comment":"Please report the random-forest hyperparameters (number of trees, maximum depth, minimum samples per leaf, etc.) and the random seed used for the train/test split, since these are needed to reproduce the exact numbers.","section":"V-A"},{"comment":"The blue and yellow histograms are drawn in the same panel and overlap substantially; transparent bars or separate panels would make the comparison readable.","section":"Figure 4"},{"comment":"The claim that the rotations error budget 'behave[s] similarly to the T states budget' because rotation approximation includes T-state usage is plausible but not demonstrated; a scatter plot or correlation coefficient between the two budgets across circuits would support this guideline.","section":"V-B"},{"comment":"The phrase 'including samples without improvement' is ambiguous given the min-selection protocol; clarify that these are circuits for which the uniform baseline was selected because the ML prediction was worse, not circuits where the ML-only prediction happened to match uniform.","section":"V-C"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid engineering contribution with a reproducible pipeline, but the headline result must be recalculated and re-stated for ML-only predictions. The missing fixed-heuristic baseline is also important for assessing whether the ML component earns its place. I saw no problematic citation or disclosure issues; the main risk is that the abstract overstates what the experiments actually show."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the work is a genuine, reproducible application of ML to error budget allocation for FTQC resource estimation, and the authors are transparent about their setup. But the headline reduction is measured against a hybrid that takes the better of the ML prediction and the uniform baseline, so the claim that ML alone beats uniform is not supported as stated. And the comparison lacks a trivial fixed-heuristic baseline, so we don't know how much of the gain is due to learning.\n\nWhat's actually new: to my knowledge, no prior resource estimation work trains a model to choose the split of the error budget among logical, T states, and rotations. The dataset construction (1000 random splits per circuit, pick the lowest space-time product) is reasonable, and training a random forest on logical counts is straightforward. The evaluation uses a properly held-out test set, and the code is on GitHub. The histograms and guidelines are a nice byproduct, even if they're just descriptive summaries of the training data.\n\nWhere it's soft:\n\n1. The selection issue is real and material. Section V-A explicitly says the cost of the ML prediction and the uniform distribution are compared and the lower one is chosen for every circuit. That means the plotted differences are all non-negative by construction and the 15.6% average is the gain of an oracle that can discard bad ML guesses. The abstract and Section V-C repeatedly call this 'the ML model's predictions' compared to uniform. To make the claim clean, they need to report the ML-only win rate and the distribution of losses when ML is worse. My guess is the gains would shrink, maybe substantially.\n\n2. No baseline against a fixed rule extracted from their own guidelines, e.g., allocate roughly 0.1/0.5/0.4 to logical/T/rotations. Such a heuristic might capture most of the improvement, which would shrink the value of the trained model.\n\n3. Generalization is only within MQT Bench, with a random circuit split. A leave-one-family-out or a second benchmark would tell us whether the model actually transfers to unseen circuit structures. The paper's claim about 'arbitrary circuits' is too strong.\n\n4. The additive error model: if the three budgets are per-part error probabilities, their sum is a conservative upper bound on the total failure probability via the union bound, so the 10% case is safe; the paper should just say this. Minor.\n\nBottom line: the contribution is a useful engineering step, and the paper deserves a serious referee. With the ML-only numbers, a heuristic baseline, and a scope statement, it could be a solid venue-level paper. For now it's a conditional accept that needs those numbers.","headline":"Useful, reproducible ML approach to error-budget allocation, but the headline savings are inflated by a hybrid selection rule and the lack of a simple-heuristic baseline leaves the marginal value of the model unclear.","tokens_in":10421,"tokens_out":3392,"would_cite":true,"duration_ms":31578,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a machine-learning-predicted split of a quantum circuit's error budget lowers estimated fault-tolerant hardware costs for over 75% of circuits, by an average of 15.6%.","keywords":["fault-tolerant quantum computing","error budget distribution","resource estimation","space-time cost optimization","machine learning","quantum error correction","T-state production","logical qubits"],"falsifier":"Re-run the resource estimation for the same test circuits with a multiplicative error model, where the total success probability is the product of the three parts' success probabilities, and compare the space-time costs of the predicted distributions against the uniform baseline; if the predicted gains disappear or the budget is violated at the 10% total error level, the additive assumption is doing the work.","tokens_in":9565,"feed_emoji":"⚛️","tokens_out":6195,"duration_ms":53914,"temperature":0.7,"pith_summary":"The paper sets out to show that the way a fault-tolerant quantum circuit's total error budget is split among its main parts is a practical lever for reducing hardware requirements. End users normally assume the budget is split evenly, but the paper argues that some circuit parts, such as T-state production, can absorb errors at lower resource cost than others, so the split should be lopsided. To make this concrete, it builds a labeled dataset by randomly sampling many budget splits, estimating each with a resource estimator, and keeping the split with the lowest product of physical qubits and execution time. A machine learning model is then trained to predict such a split directly from a circuit's logical gate counts. On 383 unseen circuits, the predicted splits reduce estimated space-time costs for over 75% of circuits, with an average reduction of 15.6% and a maximum of 77.7% compared with a uniform split.","feed_headline":"Uneven error budgets cut estimated quantum costs by 15.6%","feed_subtitle":"Fault-tolerant circuits can tolerate more error in cheap parts and less in costly ones, saving qubits and time.","key_machinery":"The load-bearing object is the error budget distribution, a triple of error probabilities assigned to logical-qubit implementation, T-state production, and rotation-gate approximation, normalized so the three sum to the user's total error budget. The argument is carried by a supervised machine learning model trained on a dataset built by sampling many random triples per circuit, having a resource estimator compute the resulting space-time product, and labeling each circuit with the triple that minimized that product. The model's input features are the circuit's logical gate counts, so the learned mapping, once trained, predicts a resource-efficient distribution for an arbitrary new circuit without further search.","core_discovery":"The central discovery is that the uniform error budget distribution is almost always suboptimal: for a fixed total error budget, shifting allowed error away from implementing logical qubits and toward T-state production and rotation-gate approximation lowers the estimated physical resources needed to execute the circuit fault-tolerantly. The paper states this as a quantitative claim about a resource estimator: using a machine-learning-predicted distribution instead of a uniform one reduced estimated space-time costs for more than 75% of 383 previously unseen circuits, by an average of 15.6% including circuits that did not improve, and by up to 77.7%.","pith_inferences":["Editorially: because the paper's cost labels come from an estimator that adds the three error budgets linearly, the reported savings are predictions under that additive model; a test that composes failure probabilities multiplicatively could change the optimal split, especially at the 10% total budget level.","Editorially: the same data-accumulation pipeline could be relabeled with a different objective, such as physical qubit count alone or execution time alone, so the method is not tied to the space-time product.","Editorially: a natural next experiment is to check whether distributions trained under one error-correcting code, with its specific T-state cost, still help when the code or magic-state factory changes; the learned mapping would likely need retraining if relative costs shift.","Editorially: an empirical validation on hardware or detailed simulation with a multiplicative error model would separate the genuine budgeting insight from artifacts of the additive estimator."],"forward_implications":["If the paper is right, fault-tolerant resource estimates for a given application can be reduced by simply changing how a fixed error budget is allocated, without tightening hardware error rates or changing the error correction code.","The same general pattern of a lower logical error budget and higher T-state and rotation budgets should transfer across circuits and across total budgets, since the effect appeared at 0.1%, 1%, and 10% total error.","Predicted distributions can be produced cheaply once the model is trained, because the input is just logical gate counts, avoiding the need to sample thousands of distributions for each new circuit.","Most circuits, but not all, benefit from the predicted split, so the method supports per-circuit tailoring rather than a single universal distribution rule."],"supporting_citations":[{"why":"Provides the resource estimation tool that computes all qubit counts and execution times used to label the dataset.","marker":"[25]"},{"why":"Defines the assumed computation scheme and error-correction parameters that determine how logical gate counts translate into physical resources.","marker":"[26]"},{"why":"Supplies the benchmark quantum circuits from which the training and test data were drawn.","marker":"[35]"},{"why":"Provides the random-forest implementation used to learn the mapping from circuits to optimized error budget distributions.","marker":"[36]"}],"fun_headline_variants":["Uneven error budgets cut estimated quantum costs by 15.6%","ML-optimized error budgets reduce quantum overhead by 15.6%","Error budget reallocation slashes quantum resource estimates 15.6%","Quantum cost estimates drop 15.6% with optimized error budgets","Smarter error distribution cuts quantum costs by 15.6%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's estimates assume the total error is exactly the sum of the three part budgets; if real failure probabilities combine differently, such as multiplying survival probabilities, the optimized split could silently exceed the user's intended total error.","fun_headline_variants_meta":{"raw":{"variants":["Uneven error budgets cut estimated quantum costs by 15.6%","ML-optimized error budgets reduce quantum overhead by 15.6%","Error budget reallocation slashes quantum resource estimates 15.6%","Quantum cost estimates drop 15.6% with optimized error budgets","Smarter error distribution cuts quantum costs by 15.6%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000588,"raw_usage":{"total_tokens":2770,"prompt_tokens":967,"completion_tokens":1803,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":583,"completion_tokens_details":{"reasoning_tokens":1707}},"tokens_in":583,"tokens_out":1803,"duration_ms":12622,"temperature":1.0,"reasoning_tokens":1707,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:35:14.668421+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the resource estimation for the same test circuits with a multiplicative error model, where the total success probability is the product of the three parts' success probabilities, and compare the space-time costs of the predicted distributions against the uniform baseline; if the predicted gains disappear or the budget is violated at the 10% total error level, the additive assumption is doing the work.","supporting_citations":[{"cited_title":"Using azure quantum resource estimator for assessing per- formance of fault tolerant quantum computation,","cited_arxiv_id":null,"evidence_quote":"Provides the resource estimation tool that computes all qubit counts and execution times used to label the dataset."},{"cited_title":"Assessing requirements to scale to practical quantum advantage,","cited_arxiv_id":null,"evidence_quote":"Defines the assumed computation scheme and error-correction parameters that determine how logical gate counts translate into physical resources."},{"cited_title":"Scikit-learn: Machine learning in Python,","cited_arxiv_id":null,"evidence_quote":"Provides the random-forest implementation used to learn the mapping from circuits to optimized error budget distributions."}],"review_version":2}