{"id":"6b86da57-8b85-4b88-95a5-13dff5dd5400","arxiv_id":"2506.09658","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Choosing the five strongest operators at each ADAPT-VQE step, rather than one, cuts reported quantum function calls to chemical accuracy by about 4.3 times for BeH2.","lead":"This paper proposes K-ADAPT-VQE, a version of the ADAPT-VQE quantum chemistry algorithm that adds five operators to the quantum circuit per step rather than one. Experiments on small molecules suggest this reduces the number of quantum circuit calls needed for chemical accuracy by roughly a factor of four.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 4.3x resource-reduction claim rests on unequal optimizer budgets and internally inconsistent call-count arithmetic in Section 3, so it is not yet established.","rationale":"The reader's weakest assumption—that the comparison is fair and robust despite unequal COBYLA iteration budgets and an assumed three evaluations per iteration—is exactly the load-bearing issue. My pass sharpens it by locating specific arithmetic contradictions in Section 3: the 5-ADAPT formula uses 800 iterations when the stated protocol gives 1000, and the 1-ADAPT formula uses 3200 when the stated protocol gives 25×40=1000. These are not mere typos in exposition; they are the basis for the 3300, 14100, and 4.3x numbers quoted in the abstract and conclusion. Because the empirical advantage is entirely a function of these counts, the central claim is currently unsupported. The heuristic itself may be valid, and the paper gives plausible physical reasoning for why simultaneously adding the K largest-gradient operators could help, especially when those operators overlap in qubit support. I also credit the paper for clearly stating its protocol and operator-pool choices, and for using standard quantum chemistry benchmarks. However, no code or data are provided, and the reported numbers cannot be independently audited. A conditional verdict remains appropriate: the core idea is testable and potentially useful, but the quantitative resource-reduction claim should not be accepted until the comparison is repeated with matched per-step budgets, a consistent call-counting methodology, and reproducible artifacts. My recommendation is therefore unchanged from the reader's CONDITIONAL verdict, not a move to acceptance or rejection.","tokens_in":6340,"tokens_out":4593,"duration_ms":45150,"concrete_test":"Re-run the BeH2 1.3 Å simulation with the stated operator pool and convergence threshold, instrumenting the code to count all quantum function calls, but vary the COBYLA budget symmetrically: for both 1-ADAPT and 5-ADAPT, use the same number of optimizer iterations per ADAPT step (e.g., 40, 100, 200, 400) and record the total calls needed to reach |E−E_FCI| < 10^-3 Ha. Then recompute the Section 3 call-count formulas using the stated totals (5 chunks × 200 = 1000 iterations and 25 operators × 40 = 1000 iterations) to check whether the reported 3300 and 14100 evaluation figures are reproducible; if the ratio changes materially or 1-ADAPT reaches chemical accuracy under any matched budget with comparable calls, the 4.3x claim is not robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim—roughly a 4.3-fold reduction in quantum function calls for BeH2—is computed in Section 3 from two arbitrary and asymmetric protocol choices: 5-ADAPT is allowed 200 COBYLA iterations per chunk (effectively 200 iterations per five added operators), while 1-ADAPT is allowed only 40 iterations per single added operator. Because 5-ADAPT buys five times more optimization budget per newly added operator, the comparison is not apples-to-apples. More seriously, the arithmetic used to obtain the reported 4.3x ratio is internally inconsistent. The text states that both procedures use a total of 1000 VQE iterations, yet the 5-ADAPT call-count formula reads '3×800 + 180×25/5', which assumes only 800 optimizer iterations, not 1000, and the 1-ADAPT formula reads '3×3200 + 180×25', which assumes 3200 iterations, not the 25×40=1000 stated in the same paragraph. The factor of three quantum function evaluations per COBYLA iteration is also asserted without evidence. Since the paper's headline contribution is resource reduction, these numbers are load-bearing: if the two methods are compared with matched per-step optimizer budgets and a consistent counting procedure, the claimed advantage may shrink or reverse.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes K-ADAPT-VQE, a modification of ADAPT-VQE in which the K operators with the largest gradient magnitudes are added to the ansatz per iteration, rather than a single operator. It reports classical simulations for BeH2, LiH, and N2 in the STO-3G basis, comparing K=5 (5-ADAPT) with K=1 (1-ADAPT), and claims a roughly 4.3-fold reduction in quantum function calls for BeH2 while attaining chemical accuracy.","tokens_in":6624,"tokens_out":12015,"duration_ms":127883,"significance":"If the central claim is correct, chunking operators by gradient rank is a simple, computationally cheap heuristic with clear practical value for ADAPT-VQE on noisy hardware. The paper has useful strengths: it uses external benchmarks (HF and FCI from PySCF), gives a clear algorithmic pseudocode, and makes a concrete, falsifiable resource-reduction prediction. However, the quantitative evidence for the headline claim is currently compromised by internal inconsistencies in the cost accounting, and no sensitivity analysis is provided for the key protocol choices. The significance therefore rests on a load-bearing calculation that is not yet reproducible as written.","major_comments":[{"comment":"The call-count arithmetic is internally inconsistent. The text states: 'For both the 5- and 1-ADAPT procedures, the total number of VQE iterations is 1000 (200 iterations for 5-ADAPT chunks, 40 iterations for each 1-ADAPT operator).' Immediately below, however, the 5-ADAPT formula 3×800 + 180×25/5 assumes 800 optimizer iterations, and the 1-ADAPT formula 3×3200 + 180×25 assumes 3200; neither matches the stated 1000. The terms 180×25/5 and 180×25 are never defined, and the factor 'three evaluations per COBYLA iteration' is asserted without justification or measurement. Since the 4.3x reduction is computed from these numbers, the headline claim is not reproducible as written. The inconsistency is also visible in Fig. 1(b), whose x-axis extends to 3200 iterations although the text says both runs used 1000.","section":"Section 3, BeH2 cost comparison"},{"comment":"The efficiency comparison uses a single hand-selected pair of per-step budgets (200 VQE iterations per 5-ADAPT step, 40 per 1-ADAPT step) and no sensitivity analysis. Even if the total iteration counts are meant to be equal, the optimization schedule differs between the two protocols, and the reported cost ratio depends on these choices. The paper states that K=5 was chosen empirically; without tests for other K values and for matched per-operator or per-step budgets, the general claim that K-ADAPT-VQE 'substantially reduces' calls is not established. The authors should report the cost ratio for several (K, budget) combinations or provide an argument that the advantage is insensitive to these choices.","section":"Section 3 and Fig. 1 caption"},{"comment":"The conclusion says 5-ADAPT achieves 'comparable accuracy' to 1-ADAPT, but Fig. 1(a) and the accompanying text report that 1-ADAPT errors are roughly an order of magnitude larger. If 1-ADAPT does not reach chemical accuracy at 1000 iterations, then the phrase 'to achieve chemical accuracy' in the call-count comparison is ambiguous or false; if it does reach chemical accuracy only after 3200 iterations, the statement that both procedures used 1000 iterations is wrong. The paper must clarify what the reported call counts are actually measuring: reaching chemical accuracy, reaching a fixed ansatz size, or some other stopping criterion.","section":"Section 4, Conclusion"}],"minor_comments":[{"comment":"The caption labels the N2 panel as '(b)', but (b) is the BeH2 convergence plot; the N2 panel should be (d). The same typo appears in the text where 'Fig 1(c) shows results for N2.'","section":"Fig. 1 caption"},{"comment":"The convergence threshold epsilon is never given a numerical value; the Results section should state the threshold used, since it directly affects the reported iteration counts.","section":"Section 2.4, Algorithm step 4e"},{"comment":"The paper should define what counts as a 'quantum function call' (e.g., one circuit execution, one expectation-value estimation, or one energy evaluation including shot noise) and state explicitly whether the MIMIQ simulations are noiseless. This is central to the resource metric.","section":"Section 3"},{"comment":"The maximum number of VQE iterations for 1-ADAPT is given only in the Fig. 1 caption, not in the algorithm description; it should be stated in the main text alongside the 200-iteration choice for 5-ADAPT.","section":"Section 2.4, Algorithm step 4d"}],"recommendation":"major_revision","confidential_remarks":"The load-bearing numbers in Section 3 are internally inconsistent, so I could not verify the 4.3x claim from the manuscript alone. I would ask the authors to supply the raw run logs or a reproducible script during revision; if the inconsistencies cannot be resolved, the quantitative claim should be removed or substantially qualified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: this paper proposes a simple, plausible tweak to ADAPT-VQE—add the top K operators by gradient magnitude each iteration, without requiring disjoint qubit supports. That is a legitimate variation on TETRIS-ADAPT-VQE, and the simulations across BeH2, LiH, and N2 show 5-ADAPT reaching chemical accuracy at more bond lengths than 1-ADAPT within the same nominal iteration cap. The idea is worth taking seriously.\n\nBut the central quantitative claim, the 4.3x reduction in quantum function calls, is not established. The paper states both procedures use 1000 VQE iterations total, yet the call-count formulas in Section 3 use 800 iterations for 5-ADAPT and 3200 for 1-ADAPT. That is a direct contradiction. On top of that, the comparison is asymmetric: 5-ADAPT gets 200 COBYLA iterations per chunk, 1-ADAPT gets 40 per operator. So the per-operator optimization budget differs by a factor of five. The 'three evaluations per COBYLA iteration' is asserted without any basis. If the budgets are matched and the counting made consistent, the 4.3x factor could shrink or even reverse.\n\nThe energy convergence plots are more convincing than the cost accounting. They show 5-ADAPT hitting the chemical-accuracy threshold more often across bond lengths. But there are no error bars, no sensitivity analysis for K or the per-step budget, and no code or data to check the specific runs. For an empirical claim, that is thin.\n\nWhat is genuinely new here is the selection rule—dropping the disjoint-support constraint and simply taking the top K by gradient. That is a small but real step beyond TETRIS-ADAPT-VQE, and the paper cites the relevant prior work. So the contribution is modest but not nil.\n\nWho is this for? Researchers actively working on ADAPT-VQE and near-term quantum chemistry who care about cheap heuristics for reducing circuit construction overhead. It is not a landmark result, but it is a testable idea.\n\nRecommendation: this deserves peer review, because the heuristic is simple, the failure mode is fixable, and the field would benefit from a clean comparison. But a serious referee should demand a matched-budget comparison, consistent arithmetic, error bars, and ideally code/data. Without those, the reported speedup is not credible.","headline":"The chunking idea is plausible, but the headline 4.3x savings number does not survive contact with the paper's own call-count arithmetic.","tokens_in":7150,"tokens_out":3727,"would_cite":false,"duration_ms":36855,"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":"K-ADAPT-VQE: adding five operators per step cuts quantum function calls by a factor of 4.3 in BeH2 while keeping chemical accuracy.","keywords":["ADAPT-VQE","variational quantum eigensolver","operator chunking","molecular ground state","quantum chemistry","NISQ algorithms","ansatz construction","resource reduction"],"falsifier":"Directly instrument a run of both algorithms on the same molecule, counting every call to the quantum energy function rather than assuming three per iteration, and give both procedures identical total VQE iteration budgets. If 5-ADAPT does not reach the same chemical-accuracy threshold with fewer quantum calls than 1-ADAPT under those matched conditions, the central resource-reduction claim is not established.","tokens_in":6154,"feed_emoji":"⚛️","tokens_out":7482,"duration_ms":81292,"temperature":0.7,"pith_summary":"K-ADAPT-VQE modifies ADAPT-VQE by adding the K strongest candidate operators to the trial wavefunction at once instead of one at a time, with K set to 5 in the demonstrations. The paper claims this simple scheduling change substantially cuts the number of VQE optimization iterations and quantum function calls needed to reach chemical accuracy in small molecular ground-state calculations. For linear BeH2 it reports a roughly 4.3-fold reduction in quantum function calls versus the standard one-operator procedure, with comparable or better accuracy across bond lengths, and similar qualitative gains for LiH and N2. A sympathetic reader takes the central insight to be that operator selection by gradient magnitude can be chunked without losing the physical content of the ansatz.","feed_headline":"Chunking operators cuts VQE calls 4.3x","feed_subtitle":"Adding five operators per ADAPT step, not one, reaches chemical accuracy with far fewer quantum evaluations on BeH2, LiH, and N2.","key_machinery":"The load-bearing mechanism is the selection and insertion of a chunk: a group of K=5 operators sorted by absolute gradient magnitude, added together with zero initial parameters, then optimized as part of the full accumulated ansatz. The supporting operator pool is pruned to spin-conserving double excitations that move electrons from Hartree-Fock-occupied orbitals into virtual orbitals, removing single excitations and non-spin-conserving terms before selection begins. The cost accounting that carries the central claim assumes three quantum function evaluations per VQE iteration and counts total evaluations along the convergence path, so the quantum-call reduction is a direct consequence of needing fewer optimization iterations to reach chemical accuracy.","core_discovery":"The paper's demonstrated discovery is that gradient-ranked chunking is an effective resource-reduction heuristic for ADAPT-VQE. At each iteration the algorithm computes the absolute values of the commutator expectation $\\langle \\Psi | [\\hat P_i, H] | \\Psi \\rangle$ for all pool operators, picks the top K, appends them in decreasing order of gradient with zero initial angles, and re-optimizes all parameters. In the BeH2 comparison, 5-ADAPT reaches an error below $10^{-3}$ Hartree using roughly 3300 quantum function evaluations, while 1-ADAPT needs about 14100, a cost reduction of about 4.3 times; the resulting energies stay close to full configuration interaction. The authors further claim that the ansatz does not suffer when operators within a chunk act on overlapping qubits, which is the feature that separates this strategy from approaches that require disjoint operator supports.","pith_inferences":["The 4.3x figure is a point estimate tied to the specific comparison budgets; a fairer efficiency measure would hold total optimization effort fixed or measure wall-clock time on a simulator, and the ratio would likely change.","Chunking by gradient magnitude acts like a coarse-grained update on the ansatz landscape, so the idea may transfer to other adaptive variational algorithms beyond ADAPT-VQE, not just molecular ground states.","A direct test would scan K across 1, 2, 3, ... for one molecule and plot quantum calls versus accuracy; the paper fixes K=5 and does not report where the optimum lies."],"forward_implications":["On noisy intermediate-scale hardware, where each quantum function call is expensive, a 4.3-fold reduction in calls makes adaptive ground-state calculations more feasible for small molecules.","The modification is a drop-in change to the ADAPT-VQE selection step, so it composes with other improvements such as pool pruning and better classical optimizers.","K becomes a tunable hyperparameter: larger chunks mean fewer ADAPT iterations but more parameters entering at once, so the optimal K likely depends on the molecule and the optimizer.","The observation that overlapping operators within a chunk do not degrade accuracy contradicts a design assumption of disjoint-support methods and suggests chunking can be paired with denser circuits."],"supporting_citations":[{"why":"defines the ADAPT-VQE procedure that K-ADAPT-VQE generalizes by chunking","marker":"[3]"},{"why":"introduces the VQE protocol whose optimization loop the paper counts","marker":"[8]"},{"why":"supplies the operator-pool construction method based on coupled exchange operators that the paper adapts","marker":"[10]"},{"why":"provides the Hartree-Fock and full configuration interaction reference energies used as benchmarks","marker":"[12]"},{"why":"constructs the second-quantized molecular Hamiltonian from the integral data","marker":"[6]"},{"why":"presents the disjoint-support ADAPT variant that K-ADAPT-VQE contrasts with","marker":"[1]"},{"why":"is the quantum emulator used to produce the simulated ground-state results","marker":"[4]"}],"fun_headline_variants":["Chunking operators cuts VQE calls 4.3x","K-ADAPT-VQE: 5-operator chunks slash evaluations","Add K operators per step to speed ADAPT-VQE","Chunked ADAPT-VQE needs 4.3x fewer calls","K-ADAPT-VQE: chunk size 5 cuts quantum calls"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The cost comparison depends on treating the chosen per-step optimization budgets (200 VQE iterations for each 5-ADAPT chunk versus 40 for each 1-ADAPT operator) and the fixed count of three quantum function evaluations per iteration as a fair and accurate measure; if those budgets or the counting rule were changed, the reported 4.3-fold reduction would change.","fun_headline_variants_meta":{"raw":{"variants":["Chunking operators cuts VQE calls 4.3x","K-ADAPT-VQE: 5-operator chunks slash evaluations","Add K operators per step to speed ADAPT-VQE","Chunked ADAPT-VQE needs 4.3x fewer calls","K-ADAPT-VQE: chunk size 5 cuts quantum calls"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00028,"raw_usage":{"total_tokens":1607,"prompt_tokens":837,"completion_tokens":770,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":453,"completion_tokens_details":{"reasoning_tokens":677}},"tokens_in":453,"tokens_out":770,"duration_ms":7457,"temperature":1.0,"reasoning_tokens":677,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:43:09.472176+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Directly instrument a run of both algorithms on the same molecule, counting every call to the quantum energy function rather than assuming three per iteration, and give both procedures identical total VQE iteration budgets. If 5-ADAPT does not reach the same chemical-accuracy threshold with fewer quantum calls than 1-ADAPT under those matched conditions, the central resource-reduction claim is not established.","supporting_citations":[{"cited_title":"Comparative Benchmarking of Utility-Scale Quantum Emulators","cited_arxiv_id":"2504.14027","evidence_quote":"is the quantum emulator used to produce the simulated ground-state results"}],"review_version":1}