{"id":"19ed70e0-9f3a-4315-a0f5-f0bbf2cc7c34","arxiv_id":"2601.09462","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"On sparse assignment problems, ferromagnetically coupled stacked replicas of an Ising model outperform the centralized penalty-spin design, which loses solution structure when many replicas are averaged in its auxiliary layer.","lead":"This paper compares two ways of coupling multiple copies ('replicas') of a hard optimization problem in an Ising machine: a centralized penalty-spin layer versus direct neighbor-to-neighbor links. On sparse assignment problems solved with simulated annealing, the directly coupled 'stacked' model stays reliable and improves as more replicas are added, while the centralized penalty-spin model collapses at large replica counts because averaging wipes out the few '1' bits that en","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Shared schedule may cause PS collapse; topological ranking needs per-model schedule test","rationale":"The paper is a careful, honest numerical study with mechanism analysis; the reader's conditional verdict is appropriate. My stress-test converges on the same load-bearing assumption: the shared temperature schedule (§3) is the linchpin of the model comparison. The PS model's failure at large P is dramatic (Fig. 4(b)) and the §5.2 mechanism (⟨x⟩_PS→0, ⟨S⟩_1→0) is derived from runs under that schedule. Because the PS auxiliary layer's flip energy scale grows with P, a schedule fixed by the original problem's energy differences may specifically disadvantage PS; if so, 'cooperation collapse' is a schedule artifact rather than a topological necessity. The concrete test (recomputing the schedule from the full PS Hamiltonian, or scaling N_Steps with P) would settle this. I do not find a separate internal inconsistency: the metrics are defined consistently, decoding is uniform, and limitations (L≤20, noise-free) are stated. The generalization from QAP to all one-hot problems (§5.2.3) is a scope limitation, not a flaw in the QAP result.","tokens_in":22809,"tokens_out":12038,"duration_ms":119243,"concrete_test":"Rerun the PS model at P=30, μ=5, |J_P|=3 (the collapse point in Fig. 4(b)) with: (a) the original shared schedule as control; (b) a schedule recomputed from the full PS Hamiltonian by applying the same OpenJij β_init/β_final routine (Eqs. 16–20) to the PS model itself; (c) the shared schedule but N_Steps scaled linearly with P (e.g., 10× or 30× the original). Keep all other parameters and decoding identical. If P_Feasible in (b) or (c) rises from ≈0 to the C/stacked level, the cooperation collapse is a schedule artifact and the topological conclusion is not established. Also run the same (b)/(c) variants for the stacked model at J_P=−3 as a control to confirm the schedule change does not erase the stacked model's advantage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the PS model has an intrinsic large-P cooperation collapse depends on P_Feasible≈0 at P=30, |J_P|≥2 (Fig. 4(b)) and the mechanism in §5.2 (⟨x⟩_PS→0, ⟨S⟩_1→0). The fairness premise is that the same temperature schedule (Eqs. 16–20) is used for all models (§3). That schedule is derived from energy-difference samples ('2N samples') and is not recomputed from the full replica-coupled Hamiltonian that includes J_P. In the PS model, flipping the auxiliary-layer spin crosses an energy barrier that grows with J_P(P−1); as P grows, this barrier becomes large relative to the fixed schedule's temperatures, potentially freezing the PS layer in a zero-biased state. The §5.2.3 averaging mechanism is then inferred from runs that may already be in this frozen regime. Because the stacked model's per-spin flip barriers are O(J_P) independent of P, the shared schedule is less mismatched for it, so the observed topology ranking may reflect unequal schedule adequacy rather than an intrinsic topological failure. This is load-bearing for the abstract's 'decisively influences' claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper uses simulated annealing as a noise-free testbed to compare three replica-coupled Ising architectures for the quadratic assignment problem: an independent-replica baseline (C model), the penalty-spin (PS) model with a centralized auxiliary layer, and a stacked model with nearest-neighbor ferromagnetic or antiferromagnetic inter-replica couplings. With P total layers, 1000 SA runs per parameter point on QAPLIB instances tai12a-tai20a, the paper reports that the ferromagnetically coupled stacked model maintains feasibility and improves the approximation ratio as P and N_Steps grow, while the PS model loses feasibility at large P (P_Feasible ~ 0 at P=30, |J_P|>=2) and shows vanishing Jaccard-type 1-1 correlation. The proposed mechanism is that many-replica averaging in the PS layer washes out sparse one-hot information, biasing the auxiliary layer toward zero and preventing coordination. The paper also explains low-mu feasibility of the AFM-stacked model by destabilization of the all-zero state. It concludes that inter-replica coupling topology decisively influences search robustness and offers model-selection and tuning guidelines.","tokens_in":22999,"tokens_out":8836,"duration_ms":86408,"significance":"The empirical core is competently executed: the comparison uses an external QAPLIB optimum, a C-model baseline, 1000-run ensembles with standard-deviation error bars, and the infeasible-sample artifact in Fig. 4(d) is explicitly flagged rather than claimed as a performance gain. The bit-level diagnostics (<x>_PS, <S>_1) provide a concrete, falsifiable account of the proposed averaging mechanism. If the central ranking survives a fair-schedule control, the paper would provide a practically useful design principle for replica-coupled annealing: decentralized nearest-neighbor coupling is more robust than centralized auxiliary coupling for sparse one-hot constraints, and AFM coupling is a feasibility aid but not a quality aid. The main limitation is breadth: all numerical evidence comes from one problem family (QAP, L<=20) under one temperature-schedule convention, so the 'decisive topology' conclusion is not yet established at the level of generality claimed.","major_comments":[{"comment":"The load-bearing claim of an intrinsic large-P cooperation collapse in the PS model is made under a temperature schedule derived from 2N single-spin-flip samples of the un-replicated problem Hamiltonian H0, not from the replica-coupled Hamiltonian. In the PS model, flipping one auxiliary spin costs up to 2|J_P|(P-1); at P=30, |J_P|=2 this barrier is about 116 in units of J_P, whereas the stacked model's flip barriers are O(|J_P|) independent of P. The same nominal schedule therefore operates in very different parts of the two models' low-temperature kinetics, and the <x>_PS -> 0 / <S>_1 -> 0 statistics in Section 5.2 may reflect freezing of the auxiliary layer rather than information-theoretic averaging. Please add a control in which the schedule is recomputed from the full PS Hamiltonian (or otherwise matched to each model's acceptance rate) and report whether P_Feasible ~ 0 and the cor","section":"Section 3 (Eqs. 16-20) and Section 4.2 (Fig. 4(b))"},{"comment":"At fixed P, the PS model contains P-1 problem replicas plus one auxiliary layer without H0, whereas the C and stacked models contain P problem replicas. The paper treats this as an intrinsic structural cost, but it also means the PS model has fewer layers that directly optimize the objective and fewer candidates for minimum-energy decoding. To separate topology from resource allocation, compare the PS model with P+1 total layers (i.e., P problem replicas plus one auxiliary) against the stacked model with P layers. If the PS model still collapses, the topological explanation is supported; otherwise the ranking may reflect the missing problem replica rather than the centralized coupling topology.","section":"Section 2.1 (Eqs. 1-3) vs Section 4.2"},{"comment":"The conclusion is generalized to 'assignment, matching, and scheduling' problems based on QAP's one-hot sparsity. The numerical evidence, however, is confined to four QAPLIB instances (tai12a/tai15a/tai17a/tai20a), all from the same random QAP family. The mean-field averaging argument is plausible, but it is not demonstrated for any non-QAP one-hot-constrained problem. Please add at least one independent benchmark (e.g., maximum matching, N-queens, or a scheduling QUBO) or restrict the conclusions to QAP-like sparse assignment problems.","section":"Section 5.2.3 and abstract"}],"minor_comments":[{"comment":"Typo: 'We preparePlayers in total' should read 'We prepare P layers in total'.","section":"Section 2.1"},{"comment":"The axis labels and tick labels appear as garbled '/aj17/aj15...' strings in the submitted source. If this is not a rendering artifact, the figures are unreadable and must be replaced with legible axis labels.","section":"Figures 3-11"},{"comment":"The left side 2*beta_base*P_init has dimensions of inverse energy while the right side is dimensionless. Please check the formula or clarify the intended dimensionless convention.","section":"Eq. (17)"},{"comment":"The bit configurations shown are representative single samples. Please state how many seeds were inspected and confirm that the displayed patterns are typical, rather than selected outliers.","section":"Section 5.2.1, Fig. 9"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is honest and carefully analyzed, but the schedule-matching confound is significant enough that I cannot endorse the central causal claim as it stands. I expect a revision with the PS full-Hamiltonian schedule control and a qualification of scope to be publishable in JPSJ."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Net: a genuinely useful numerical study, and the authors are unusually honest about its limits. The new content is the systematic large-P comparison of PS and stacked topologies on a sparse constrained problem (QAP), plus a concrete mechanism for why the PS model degrades: the auxiliary layer, coupled to all replicas, aligns with the majority -1 bits, so as P grows the sparse 1-bit information is diluted and the PS layer stops guiding the replicas. That mechanism is simple, clearly stated, and supported by bit statistics (⟨x⟩_PS → 0, ⟨S⟩_1 → 0). The 1000-run ensembles with error bars, the honest flagging of the Fig. 4(d) infeasible-sample artifact, and the explicit limitation statements (§6) all earn credit. The ranking itself (FM-stacked over C over PS at large P on QAP) is well-supported by the data; the interpretation—that this is an intrinsic topological failure—is where I would push.\n\nThe soft spots are real. The biggest is the shared temperature schedule. It is derived from single-spin-flip energy differences of the original problem Hamiltonian, not from the full replica-coupled Hamiltonian. The PS auxiliary layer has flip barriers that grow with J_P(P−1), so at P=30 the schedule may simply be freezing the PS layer, not revealing an intrinsic topological failure. The fact that the collapse persists at |J_P|=3 even when N_steps is raised to 50,000 reduces the force of this worry, and at P=10 with |J_P|=3 the barrier is also large but the model does not fully collapse, which points toward the averaging story rather than a pure kinetic artifact. Still, a per-model schedule test (or a schedule that includes the J_P terms) is needed before \"decisively influences search robustness\" can be taken as established.\n\nSecond, the paper generalizes from QAP (L≤20) to all \"sparse feasible structures\" on the strength of an argument, not a second problem family. The mechanism is plausible for one-hot assignment problems, but matching/scheduling have different constraint graphs and the claim is broader than the evidence. Minor: μ is chosen per instance from feasibility data, which conditions the size-scaling comparison; no code/data is shipped; the relation to the authors' QCE 2025 short paper is not stated. All addressable.\n\nBottom line: worth sending to a serious referee. The comparison is reproducible in spirit, the mechanism is falsifiable, and practitioners in Ising-machine error mitigation will cite it. I would ask the authors to test schedule fairness, add one non-QAP constrained problem, and soften the abstract.","headline":"A careful SA comparison with a plausible mechanism, but the 'cooperation collapse' claim is only as strong as the untested shared-schedule premise—worth refereeing, not accepting as-is.","tokens_in":23601,"tokens_out":9065,"would_cite":true,"duration_ms":98246,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["05.10.-a"],"model":"deepseek-v4-flash","headline":"This paper claims that how replicas of an Ising model are wired together—through a central penalty-spin layer or directly to neighbors—determines whether parallel search stays robust, and that the neighbor-coupled stacked design wins on con","keywords":["Ising machines","simulated annealing","quadratic assignment problem","replica-coupled models","penalty-spin model","stacked model","constraint satisfaction","error mitigation"],"falsifier":"Re-run the PS model at P=30 and |J_P|=2–3 with inverse temperatures rescaled to the model's own coupling scale (or equivalently, normalize J_P by the constraint coefficient µ), and check whether P_Feasible stays near 1; the paper's cooperation-collapse narrative would weaken substantially if feasibility survives a schedule matched to the PS energy scale.","tokens_in":22579,"feed_emoji":"🧲","tokens_out":5377,"duration_ms":51690,"temperature":0.7,"pith_summary":"Error mitigation for Ising machines is usually discussed as protection against hardware noise; this paper argues it is better understood as a question of how replicas of the problem Hamiltonian are wired together. Using simulated annealing on quadratic assignment problems, it claims that a ferromagnetically coupled stacked model—where each replica talks only to its immediate neighbors—keeps constraints satisfied and improves solution quality over a broad range of penalty and coupling strengths. The penalty-spin model, where all replicas talk through one central auxiliary layer, instead suffers what the authors call cooperation collapse: as the number of replicas grows, averaging in the central layer dilutes the sparse 1-bits of one-hot solutions, so replicas stop coordinating and feasibility itself fails. If this is right, topology, not just coupling strength, is the decisive design variable for parallelized Ising search.","feed_headline":"Stacked replica coupling beats penalty-spin at scale","feed_subtitle":"Direct neighbor links keep assignment constraints satisfied; centralized averaging washes out the sparse solution.","key_machinery":"The comparison rests on three Hamiltonians over P layers: an independent-replica baseline, the penalty-spin (PS) model, in which P−1 problem replicas each couple to one shared auxiliary spin layer with strength J_P, and the stacked model, in which adjacent replicas couple directly in a ring with the same J_P. The diagnostic that carries the argument is the 1–1 correlation ⟨S⟩1, a Jaccard-style measure of how much adjacent replicas agree on which bits are 1; unlike ordinary spin correlations, it is not fooled by agreement on zeros. It is this measure that shows PS-layer averaging washing out sparse solution information as P grows, while local neighbor coupling preserves partial sharing of act","core_discovery":"The paper's central claim is that the topology of inter-replica couplings decisively influences search robustness in finite-time annealing, independent of hardware noise. On QAP, the stacked model with ferromagnetic inter-replica coupling stably achieves near-perfect feasibility across |J_P| ≤ 3 at fixed penalty µ=5, and its approximation ratio improves monotonically as the number of replicas P and annealing steps grow. In the PS model, the central penalty-spin layer averages configurations from many replicas; because feasible QAP solutions are extremely sparse, averaging drives the layer's mean bit value toward zero and erases the information about where the active bits are. The paper ident","pith_inferences":["The averaging-collapse mechanism, if generic, transfers beyond QAP: any centralized auxiliary layer used with assignment, matching, or scheduling constraints (all one-hot sparse) should show the same large-P degradation; testing those problems directly is a natural next step.","The shared-temperature-schedule protocol may understate PS performance, since PS energy scales include an extra J_P term the independent model lacks; re-running with per-model schedule normalization would isolate structural from energetic effects—this is an inference, not a claim in the paper.","For hardware with sparse local connectivity, the stacked model's ring topology maps more naturally onto limited graphs than the PS model's long-range couplings, strengthening the practical case for stacking beyond the noise-free results shown here.","A combined design is suggested by the two mechanisms: use AFM coupling early in annealing to avoid the all-zero trap and switch to FM coupling later to coordinate replicas; the present paper does not test this annealed-sign scheme."],"forward_implications":["If the claim is correct, the ferromagnetically stacked model is the preferred replica-coupling design for constrained problems with one-hot or sparse solutions: it converts extra replicas and longer annealing into better solutions instead of into coordination failure.","PS-model users should expect feasibility and cooperation to break down at large P; the paper's guideline is to keep the number of replicas small and pre-check that a usable (µ, J_P) window exists.","The AFM-stacked variant offers a bounded benefit: it stabilizes feasibility at low penalty coefficients by frustrating the all-zero state, but it does not improve approximation ratio, so it fits feasibility-first applications only.","Benchmark comparisons of error-mitigation methods should control for coupling topology, since the same coupling strength J_P produces opposite scaling behavior in the two architectures.","Minimum-energy decoding across all layers is a fair, decoder-agnostic baseline that treats the PS auxiliary layer as a candidate solution, making the reported differences attributable to the search dynamics rather than to readout choice."],"fun_headline_variants":["Stacked coupling outlasts penalty-spin at large parallelism","Direct replica links beat centralized penalty-spin averaging","Replica-coupling topology determines Ising search robustness","Penalty-spin collapse vs stacked scalability on QAP"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The comparison assumes that applying one shared, per-instance temperature schedule to all three models—whose energy scales differ because PS and stacked models add J_P couplings the independent model lacks—produces a fair test rather than systematically biasing one topology.","fun_headline_variants_meta":{"raw":{"variants":["Stacked coupling outlasts penalty-spin at large parallelism","Direct replica links beat centralized penalty-spin averaging","Replica-coupling topology determines Ising search robustness","Penalty-spin collapse vs stacked scalability on QAP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000561,"raw_usage":{"total_tokens":2470,"prompt_tokens":684,"completion_tokens":1786,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":428,"completion_tokens_details":{"reasoning_tokens":1721}},"tokens_in":428,"tokens_out":1786,"duration_ms":12294,"temperature":1.0,"reasoning_tokens":1721,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T10:36:59.165063+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the PS model at P=30 and |J_P|=2–3 with inverse temperatures rescaled to the model's own coupling scale (or equivalently, normalize J_P by the constraint coefficient µ), and check whether P_Feasible stays near 1; the paper's cooperation-collapse narrative would weaken substantially if feasibility survives a schedule matched to the PS energy scale.","supporting_citations":[],"review_version":1}