{"id":"53e56f2a-b6eb-4132-8f87-caa04b46077b","arxiv_id":"2506.09214","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"CACAO is a classical spin-dynamics algorithm, derived from FALQON/CD-FQA, that outperforms the quantum versions on a 2-SAT-related benchmark and handles 10,000 spins.","lead":"A new classical algorithm, CACAO, solves QUBO optimization problems by evolving classical spins with a feedback rule borrowed from quantum Lyapunov control and counterdiabatic driving. In benchmark tests on a small Ising spin glass it reaches lower energies than three quantum-inspired algorithms, and it scales to 10,000 spins.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Benchmark is degenerate: all-up is always the planted optimum, and no trivial/classical baseline is compared, so 'best convergence' only shows CACAO beats three quantum heuristics on a specially easy 2-SAT family.","rationale":"The reader's conditional verdict already captures the main defect: the benchmark is narrow and tests only one family without classical baselines. My stress-test sharpens this into a concrete degeneracy: every generated instance has the same planted all-up global optimum, so a constant assignment obtains the exact minimum energy on all instances. The paper's comparison omits this trivial baseline, and it reports only soft-spin relaxation energies rather than rounded binary QUBO values. This does not make the derivation wrong, and the paper is explicit that broader benchmarks are future work, so the conditional-accept verdict remains appropriate. However, the central claim that CACAO is a useful classical QUBO heuristic is not established by the current evidence; it is only established that CACAO approaches a known trivial optimum faster than three specific quantum heuristics on this family. The equivalence of CACAO to a coordinate-scaled gradient descent further reinforces the need for standard classical baselines before drawing any general conclusion.","tokens_in":7812,"tokens_out":13815,"duration_ms":160371,"concrete_test":"Re-run the Sec. III B comparison on a non-degenerate problem family, e.g. include all four penalty patterns (0,0),(0,1),(1,0),(1,1) or add random longitudinal fields so all-up is no longer the planted optimum. For each instance, after CACAO terminates round mZ_i to sign(mZ_i), report the discrete QUBO energy and the fraction of spins with |mZ_i| > 0.99. Add a constant all-up baseline and a standard simulated-annealing baseline. If CACAO's rounded energies are not at least comparable to simulated annealing and no better than the all-up baseline on the original family, the empirical claim should be downgraded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equations (4)-(5) are equivalent to dmZ_i/dt = -4(1-mZ_i^2) dE/dmZ_i, i.e. a coordinate-scaled continuous-time gradient descent on the soft-spin energy E(mZ) = sum J_ij mZ_i mZ_j + sum h_i mZ_i. This classical baseline is not identified in the paper. More importantly, the benchmark in Eq. (7) draws each clause from {(0,1),(1,0),(1,1)}, so every clause has at least one factor of the form (1-Z_i)/2. The all-up bitstring 0^N therefore satisfies every clause, and since each C_ij >= 0 for Z in [-1,1], all-up is always a global optimum with E = 0. A trivial assignment returning all-up solves every generated instance exactly; standard heuristics such as simulated annealing are absent. The reported EP(T) is also the soft-spin relaxation energy, with no rounding to discrete bitstrings and no success-probability or Hamming-distance metric. Thus the headline 'best convergence' is only a comparison against QA, FALQON, and CD-FQA on a specially easy family; it does not support CACAO as a competitive general QUBO heuristic. The paper explicitly leaves broader benchmarks to future work, so this is a scope limitation rather than an internal contradiction.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CACAO, a classical algorithm for QUBO/Ising optimization that is inspired by quantum Lyapunov control and local counterdiabatic driving. The algorithm evolves soft-spin variables according to Eqs. (4)-(5), with the coefficient alpha_i chosen so that the problem energy decreases monotonically. The paper demonstrates the dynamics on a two-spin example, compares CACAO with QA, FALQON, and CD-FQA on a 2-SAT-related Ising spin-glass model (Sec. III B), and reports results for systems up to 10,000 spins (Sec. III C). The main claim is that CACAO shows the best convergence to low-energy states among the tested algorithms on this benchmark.","tokens_in":8088,"tokens_out":3501,"duration_ms":37870,"significance":"The algorithm is simple, rigorously derived, and has modest classical resource requirements (O(NT) to O(N^2T)). The monotonic energy-decrease property and the absence of divergence are genuine strengths. However, the central performance claim is not established by the evidence presented: the benchmark family is degenerate, the algorithm is not compared with standard classical heuristics, and the reported energy is a continuous soft-spin quantity rather than a rounded discrete solution. If the paper is revised to include proper baselines and discrete metrics, the result could be a useful contribution; as it stands, the significance is mainly as a derivation of a classical spin-dynamics heuristic rather than a demonstrated competitive optimizer.","major_comments":[{"comment":"The benchmark construction in Eq. (7) is degenerate: each clause is drawn from {(0,1),(1,0),(1,1)}, so the all-zero bitstring (all-up spin state) satisfies every clause and, because each C_ij is nonnegative, is always a global optimum with energy zero. Thus the comparison in Fig. 3 only shows that CACAO beats QA, FALQON, and CD-FQA on a family where the optimal solution is a single fixed bitstring, with only 10 instances at N=9 and no classical baseline such as simulated annealing, simple gradient descent, or even the trivial all-up assignment. This does not support the broader claim that CACAO is a competitive general QUBO heuristic.","section":"III B, Eq. (7), Fig. 3"},{"comment":"The paper does not identify that the CACAO dynamics is equivalent to a coordinate-scaled continuous-time gradient descent on the soft-spin energy E(mZ)=sum J_ij mZ_i mZ_j + sum h_i mZ_i. Using mX_i^2 + mZ_i^2 = 1, Eq. (5) with Eq. (4) gives dmZ_i/dt = -4(1-mZ_i^2) dE/dmZ_i. Since this is a known classical optimization structure, the paper should either compare CACAO with vanilla gradient descent on the same soft-spin energy or explain precisely what the counterdiabatic-inspired coefficient choice adds over that baseline. Without such a comparison, the claimed advantage over classical methods is not demonstrated.","section":"II, Eqs. (4)-(5)"},{"comment":"The large-system results report only the continuous soft-spin energy EP(T)/N and do not provide any discrete metric, such as the Hamming distance to the planted all-up optimum, the success probability, or the energy after rounding the final mZ_i to bit values. Since all-up is exactly the ground state for every instance, a trivial algorithm returning all-up would achieve the global optimum, so the statement that CACAO 'worked even for the system with 10,000 spins' is not supported without a comparison to trivial or standard classical baselines. The residual energy per spin being roughly constant with N does not establish solution quality.","section":"III C, Fig. 4"}],"minor_comments":[{"comment":"The scaling fit T = exp(0.5237) Delta E^{-1.08} is presented as a finding, but the fitting procedure, the number of data points, and the goodness of fit are not reported; the abstract's later statement that convergence time 'might scale as O(Delta E^{-1})' should be labeled as a numerical observation, not a derived result.","section":"III A, Fig. 2"},{"comment":"The text contains typos such as 'squre lattice' and 'F ALQON' (with a space); the algorithm name should be typeset consistently as FALQON.","section":"III B"},{"comment":"The claim that 'the time for convergence does not depend on the system size' is inferred from Fig. 4, but the figure plots energy versus operation time and does not explicitly define or extract a convergence time; please specify the convergence criterion used.","section":"III C"},{"comment":"The discussion of feedback costs for FALQON and CD-FQA is useful, but it is qualitative; since the numerical comparison already ignores these costs in favor of the quantum algorithms, the conclusion that CACAO is competitive would be strengthened by a quantitative accounting or by a clearly stated caveat.","section":"IV"}],"recommendation":"major_revision","confidential_remarks":"The main issue is benchmark design: the chosen 2-SAT family is too easy and the absence of classical baselines weakens the central claim. The author explicitly defers broader benchmarks, so the deficiencies appear fixable within the manuscript's scope rather than fatal. I would encourage the editor to request a revised version with discrete metrics and standard classical heuristics before considering publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe punchline: Hatomura derives a classical spin-dynamics heuristic (CACAO) from Lyapunov control with a local Pauli-Y driver, and the math is clean—the energy decreases monotonically and the algorithm is straightforwardly O(NT) to O(N^2 T). But the performance section is much weaker than the abstract suggests. The benchmark family in Eq. (7) is a random 2-SAT-like model where the all-up bitstring is always the planted optimum: every clause is drawn from {(0,1),(1,0),(1,1)}, so the all-up assignment satisfies all clauses and each clause term is nonnegative. The paper actually says this explicitly. That means the problem is a biased ferromagnet-like landscape with a known trivial solution, not a generic spin glass. On top of that, there is no comparison to standard classical heuristics—simulated annealing, local search, or even plain gradient descent on the soft-spin energy. As the stress-test note points out, Eqs. (4)-(5) are exactly a coordinate-scaled continuous-time gradient descent on E(mZ), so the paper is implicitly comparing a particular classical descent against three quantum heuristics on an easy instance family. The metric is also the soft-spin energy EP(T), not a rounded discrete energy or success probability, so even the 'convergence' claim is about the relaxation, not the actual binary objective.\n\nGive credit where it's due. The derivation is correct and clearly presented. The energy-decrease property is guaranteed by Lyapunov control, and the discussion honestly acknowledges that no quantum-vs-classical conclusion should be drawn yet. The two-spin scaling analysis is a nice illustrative result, though it's a fit to one small system with two free constants. The paper is also refreshingly explicit about the costs of feedback in FALQON/CD-FQA that it ignores in the numerics.\n\nThe soft spots are real and load-bearing for the central claim. The 'best convergence' headline only holds against three quantum heuristics on a family where all-up is the ground state; it does not establish CACAO as a competitive general QUBO heuristic. The absence of any classical baseline is the biggest gap. A referee should ask for comparisons to simulated annealing or gradient descent, and benchmarks on problems without a planted all-up optimum (e.g., MaxCut, random Ising spin glasses, or weighted Max-2-SAT with random fields). The convergence-time scaling also needs more evidence if it's meant to be a general statement.\n\nWho is this for? People working on quantum-inspired optimization and on benchmarking quantum heuristics might find CACAO a useful reference point. It's a legitimate proposal and the paper is honest, so it deserves peer review, but the performance claims need substantial strengthening before they should be taken as evidence of a competitive algorithm.\n\nRecommendation: send it to review, with the expectation of major revision. The theory section is fine; the benchmark section needs more baselines and more diverse instances.\n\nBest.","headline":"A clean classical spin-dynamics heuristic, but the benchmark always has all-up as the planted optimum and no classical baselines are compared, so the strong performance claim is not yet supported.","tokens_in":8584,"tokens_out":5365,"would_cite":false,"duration_ms":50788,"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":"A classical spin-dynamics algorithm, CACAO, derived from Lyapunov feedback and local counterdiabatic driving, converges to lower-energy states than quantum annealing, FALQON, and CD-FQA on a 2-SAT-related spin glass, and scales to 10,000…","keywords":["combinatorial optimization","QUBO","Ising spin glass","quantum-inspired classical algorithm","counterdiabatic driving","Lyapunov control","FALQON","2-SAT"],"falsifier":"Run the Sec. III B benchmark with an additional control: the same spin ODEs but with $\\alpha_i(t)$ replaced by fixed random values of comparable magnitude; if CACAO and the random-coefficient dynamics reach the same final energies on the same instances, the Lyapunov feedback is not the operative mechanism and the paper's central explanation is falsified.","tokens_in":7603,"feed_emoji":"⚙️","tokens_out":8795,"duration_ms":80677,"temperature":0.7,"pith_summary":"This paper proposes CACAO, a classical algorithm for combinatorial optimization that takes the feedback structure of a quantum algorithm (FALQON) and combines it with local counterdiabatic driving. The paper's central claim is that this classical spin dynamics converges to lower-energy states of a 2-SAT-related Ising spin glass than quantum annealing, FALQON, and CD-FQA do in a head-to-head numerical benchmark. If correct, the claim matters because it means the useful part of these quantum heuristics is a Lyapunov feedback rule that can run on ordinary hardware with per-timestep cost between $O(NT)$ and $O(N^2T)$, sidestepping the measurement and feedback overhead of the quantum versions. The paper further claims the algorithm works at 10,000 spins with convergence time nearly independent of system size, while also showing that near-degenerate target states slow it down according to $T \\propto \\Delta E^{-1.08}$.","feed_headline":"Classical spin dynamics outperforms three quantum heuristics","feed_subtitle":"CACAO moves the feedback idea behind FALQON and CD-FQA onto ordinary hardware and scales to 10,000 spins.","key_machinery":"The central object is a classical spin dynamics on the Bloch expectation values of a product state, driven by a local Pauli-Y Hamiltonian $\\hat H_{\\rm CACAO}(\\{\\alpha_i\\})=\\sum_i \\alpha_i(t)\\hat Y_i$. The coefficient on each spin is set by quantum Lyapunov control, $\\alpha_i(t)=2(h_i+\\sum_{j:(i,j)\\in E}J_{ij}m_j^Z)m_i^X$, which is exactly the commutator-feedback rule used in FALQON reduced to a single site. Because the driver is purely local, no entanglement builds up and the Schr\\\"odinger evolution reduces to the two coupled ODEs $\\dot m_i^X=2\\alpha_i m_i^Z$ and $\\dot m_i^Z=-2\\alpha_i m_i^X$; solving these ODEs is the entire algorithm. The Lyapunov structure guarantees $\\frac{d}{dt}\\langle \\hat H_P\\rangle = -\\sum_i \\alpha_i^2 \\le 0$, so the energy descends monotonically until the feedback coefficients vanish.","core_discovery":"On its own terms, the paper establishes that one can take the feedback-based quantum optimization idea, restrict the driving to local $\\hat Y_i$ terms with amplitudes set by quantum Lyapunov control, and obtain a closed set of classical equations for the spin expectation values $m_i^X(t)$ and $m_i^Z(t)$: $\\alpha_i(t)=2(h_i+\\sum_{j:(i,j)\\in E} J_{ij} m_j^Z)m_i^X$, $\\dot m_i^X=2\\alpha_i m_i^Z$, $\\dot m_i^Z=-2\\alpha_i m_i^X$. The Lyapunov choice makes the problem energy $\\langle \\hat H_P\\rangle$ decrease monotonically. On ten random $L=3$ ($N=9$) instances of a spin glass related to 2-SAT, CACAO achieved the best final energy among QA, FALQON, and CD-FQA across operation times, and on systems up to $N=10{,}000$ the rescaled residual energy stayed at the same level, meaning the algorithm's per-spin performance does not degrade with size. The paper also reports a two-spin convergence-time law $T=\\exp(0.5237)\\,\\Delta E^{-1.08}$, indicating that CACAO, like any gap-sensitive heuristic, becomes slow when the ground state is nearly degenerate.","pith_inferences":["The CACAO equations are a continuous-time, deterministic mean-field descent on the Ising energy; a natural next test, not in the paper, is direct comparison with classical local search and simulated annealing on the same 2-SAT instances.","Because the dynamics is deterministic and cheap, adding random initial tilts $\\delta_i$ gives a restart-based heuristic; the paper fixes $\\delta_i=0$, leaving the restart-vs-run-time trade-off unexplored.","The residual energy growing extensively with $N$ suggests CACAO is a low-energy-state finder rather than an exact solver; using its converged $m_i^Z$ as a warm start for a discrete local search is a testable extension.","The same Lyapunov rule may transfer to other QUBO classes such as MaxCut or the Sherrington-Kirkpatrick model; this is a direct consequence of the algorithm's derivation, though the paper does not test it."],"forward_implications":["CACAO is a deterministic classical heuristic with per-step cost $O(NT)$ to $O(N^2T)$, so the benchmark results transfer directly to ordinary computers; the paper demonstrates this at $N=10{,}000$.","The two-spin convergence law $T=\\exp(0.5237)\\Delta E^{-1.08}$ predicts that CACAO's runtime grows as the target gap closes, so instances with near-degenerate optima will require long integration times.","Because the construction uses only local fields, the same Lyapunov-feedback scheme extends to higher-order binary optimization (HUBO) without new theoretical machinery, as the paper explicitly notes.","In the reported benchmark, CACAO's final energy is at least as good as the three quantum algorithms at every tested operation time; since the comparison excludes measurement and initialization costs for the quantum algorithms, the practical advantage in wall-clock terms would be larger, not smaller."],"supporting_citations":[{"why":"Introduces FALQON and its Lyapunov-control-inspired parameter choice, the quantum feedback framework CACAO adapts to classical spin dynamics.","marker":"[9, 10]"},{"why":"Survey of quantum Lyapunov control methods; supplies the energy-decreasing feedback principle behind the choice $\\alpha_i(t)$.","marker":"[11]"},{"why":"Defines CD-FQA, the counterdiabatic feedback-based quantum algorithm that CACAO benchmarks against and extends with a local Pauli-Y driver.","marker":"[20]"},{"why":"Shows Pauli-Y operators are the lowest-order counterdiabatic terms for the QA Hamiltonian, justifying the driver in Eq. (3).","marker":"[22]"},{"why":"Establishes the relation between the problem Hamiltonian (7) and 2-SAT, the benchmark problem family used in Sec. III B.","marker":"[23]"},{"why":"Quantum annealing baseline algorithm against which CACAO is compared in the benchmark.","marker":"[6]"},{"why":"Variational local counterdiabatic driving with sign functions adopted for the CD-FQA comparison.","marker":"[24]"},{"why":"Previous classical spin dynamics with counterdiabatic driving that can diverge at critical points, contrasted with CACAO's non-diverging behavior.","marker":"[25, 26]"}],"fun_headline_variants":["Classical algorithm beats quantum heuristics on optimization","CACAO: classical spin dynamics outdoes quantum optimizers","Classical feedback-inspired algorithm scales to 10,000 spins","Counterdiabatic-inspired classical algorithm wins on optimization","Classical spin dynamics beats QA, FALQON, and CD-FQA"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the local, Lyapunov-chosen Pauli-Y nudges guide the continuous spin variables to discrete low-energy configurations rather than stranding them in poor local minima of the relaxed dynamics; the paper tests this only on one family of randomly generated 2-SAT instances.","fun_headline_variants_meta":{"raw":{"variants":["Classical algorithm beats quantum heuristics on optimization","CACAO: classical spin dynamics outdoes quantum optimizers","Classical feedback-inspired algorithm scales to 10,000 spins","Counterdiabatic-inspired classical algorithm wins on optimization","Classical spin dynamics beats QA, FALQON, and CD-FQA"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00075,"raw_usage":{"total_tokens":3345,"prompt_tokens":953,"completion_tokens":2392,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":569,"completion_tokens_details":{"reasoning_tokens":2307}},"tokens_in":569,"tokens_out":2392,"duration_ms":24321,"temperature":1.0,"reasoning_tokens":2307,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:54:28.702985+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the Sec. III B benchmark with an additional control: the same spin ODEs but with $\\alpha_i(t)$ replaced by fixed random values of comparable magnitude; if CACAO and the random-coefficient dynamics reach the same final energies on the same instances, the Lyapunov feedback is not the operative mechanism and the paper's central explanation is falsified.","supporting_citations":[{"cited_title":"Cong and F","cited_arxiv_id":null,"evidence_quote":"Survey of quantum Lyapunov control methods; supplies the energy-decreasing feedback principle behind the choice $\\alpha_i(t)$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines CD-FQA, the counterdiabatic feedback-based quantum algorithm that CACAO benchmarks against and extends with a local Pauli-Y driver."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows Pauli-Y operators are the lowest-order counterdiabatic terms for the QA Hamiltonian, justifying the driver in Eq. (3)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the relation between the problem Hamiltonian (7) and 2-SAT, the benchmark problem family used in Sec. III B."},{"cited_title":"Kadowaki and H","cited_arxiv_id":null,"evidence_quote":"Quantum annealing baseline algorithm against which CACAO is compared in the benchmark."},{"cited_title":"Sels and A","cited_arxiv_id":null,"evidence_quote":"Variational local counterdiabatic driving with sign functions adopted for the CD-FQA comparison."}],"review_version":1}