{"id":"8c470640-b073-490f-9418-7014bf8b507f","arxiv_id":"2607.05448","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"iSTAR certifies saturated spins in continuous Ising dynamics, folds them into an induced field, and removes about 64% of dense interaction work on G-set without degrading same-seed baselines.","lead":"Late-stage simulated-bifurcation trajectories collapse onto a smaller active set of spins; iSTAR freezes the rest algebraically and keeps only the tail. On G-set Max-Cut-style Ising instances this cuts most of the dense O(N²) interaction work while matching the same-seed full solver.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Quality preservation under online freezing is only certified for the bSB clipping update; the FLOPs-proxy savings do not yet establish that the same-seed energy match survives realistic sparse/hardware implementations.","rationale":"The Reader correctly isolates the finite-time robust-margin certificate (Eqs. 11–14 / Prop. A.10) as the true bridge from theory to the G-set claim, and correctly notes that the global PSD recovery does not cover the tested schedules. That diagnosis is accurate and load-bearing. The additional soft spot I emphasize is the gap between the certificate (which is for dense bSB clipping) and the advertised compute reduction (a dense FLOPs proxy that ignores sparsity and overhead). Because the paper already flags both issues in the Discussion, the concern does not overturn the mechanism or the same-seed energy results; it simply keeps the practical speedup claim conditional until a sparse/wall-clock verification is supplied. Hence the verdict remains CONDITIONAL and I agree with the Reader’s weakest-assumption identification.","tokens_in":30843,"tokens_out":759,"duration_ms":8360,"concrete_test":"Re-implement the online certified rule of Table 1 on the same G1–G54 / all-Gset protocol using a sparse CSR matvec for both full bSB and the progressive reduced subsystem, counting wall-clock time (or at least true nonzero multiplies + index overhead) on a fixed CPU/GPU backend; if mean wall-clock saving falls below ~20% or any same-seed energy mismatch appears once sparse layout and induced-field updates are included, the headline 64.4% claim does not transfer.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central empirical claim (all-Gset, ηλ=0.2: 1420/1420 trigger, 0 degradations, mean 64.44% dense-work saving) rests on two pieces that are not equally secure. The algebraic frozen-set identity (Thm. 6(i)/A.9) and the one-step robust-margin certificate for the exact bSB Euler+clipping rule (Prop. A.10, Cor. A.12) are clean: once ρi(k;H)>0 and the unresolved coordinates remain inside the box under a monotone α schedule, frozen signs stay frozen and the reduced Ising objective is exact. The paper itself notes that the global hard-box PSD hypothesis fails on G-set (λ min(S)≈−48.79 for G1 vs. α≤1). Quality preservation is therefore entirely conditional on the local certificate activating and remaining valid. What is less secure is the leap from that certificate to the reported compute claim. Savings are measured only by the dense matvec proxy FLOPsred=2n^{2}Kp+2|Q|^{2}(K−Kp) (Eq. 16); the paper explicitly excludes selection overhead, submatrix extraction, memory traffic, and sparse structure. On G-set the interaction matrices are sparse (e.g. G1: n=800, m=19176), so a production sparse bSB already costs O(m) per step; progressive dense-submatrix reduction can easily lose the proxy gain once indexing and induced-field updates are counted. The same-seed energy match is therefore proven only for the idealized dense bSB dynamics that the certificate covers, not yet for the practical cost model the abstract advertises.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper argues that the O(N²) dense interaction cost of continuous Ising solvers (simulated bifurcation and relatives) is not intrinsic in late stage: trajectories collapse onto a lower-dimensional active set, and saturated coordinates can be eliminated exactly by a frozen-set identity that folds their couplings into an induced field on the unresolved subsystem. It proves large-α recovery for the external-field aSB quartic landscape, epi-convergence and hard-box vertex recovery for bSB under a PSD condition, and a robust-margin one-step/persistent freezing certificate for the clipped bSB Euler update. The algorithmic realization, iSTAR, freezes only after that certificate is verified online; on G-set with nonzero external field it matches the same-seed full bSB baseline in all reported runs and removes on average ~56–64% of dense matvec work under a FLOPs proxy.","tokens_in":31345,"tokens_out":1497,"duration_ms":19607,"significance":"If the local certificate and same-seed quality preservation hold as reported, the work supplies a clean algebraic reduction principle for continuous Ising solvers rather than a heuristic early-stop. Strengths include: (i) an exact frozen-set decomposition of the Ising energy (Thm. 6(i)/A.9); (ii) an explicit one-step and persistent freezing proof for the bSB clipping rule under a robust margin (Prop. A.10, Cor. A.12); (iii) a carefully designed online experiment that does not use the baseline to choose freeze times and reports zero same-seed degradations over large seed–instance panels; and (iv) transparent discussion that the global hard-box PSD hypothesis fails on the tested G-set schedules. The contribution is a useful bridge between variational landscape theory and adaptive active-set practice for SB-type solvers, with clear scope for follow-on sparse and wall-clock evaluation.","major_comments":[{"comment":"Eq. (16) and §3.3 measure savings only by the dense matvec proxy FLOPs = 2n²K (stagewise 2|Qk|²). G-set instances are sparse (e.g. G1: n=800, m=19176), so a production sparse bSB already costs O(m) per step; progressive submatrix extraction, induced-field updates, and indexing can erase the proxy gain. The abstract’s “removes on average 64.4% of the dense interaction work” is literally consistent with the proxy, but the computational claim as advertised needs either a sparse-aware cost model on the same runs or a sharper scoping statement that the reported saving is dense-proxy only and not yet a wall-clock or sparse-complexity result.","section":null},{"comment":"Table 2 and the zero-field contrast (§4.2, Appendix Table 11) show that certified freezing and quality preservation are strongly field-dependent: with ηλ=0.2 the online rule triggers in 540/540 (Table S2) and 1420/1420 (all-Gset) runs with 0 degradations, while μ=0 triggers in only 230/540 and the probe-500 diagnostic yields 1166/1420 losses. The abstract and main claim lead with the 64.4% figure without stating that the favorable regime is the nonzero-field setting used throughout the primary experiments. The central empirical claim should be conditioned on the presence of a symmetry-breaking field (or the paper should supply a field-free certificate that works at comparable rates).","section":null},{"comment":"§2.1–2.3 and the Discussion correctly note that the global hard-box PSD hypothesis fails on G-set (e.g. G1: λmin(S)≈−48.79 would need α≳s7 while the schedule uses α≤1), so quality preservation rests entirely on the local robust-margin certificate (Eqs. 11–14 / Prop. A.10) and on unresolved coordinates remaining in the box under a monotone α schedule. The abstract and introduction still lead with large-parameter recovery and hard-box reduction as if they underwrite the G-set result. Reorder or rephrase so that the finite-time certificate is presented as the operative theory for the experiments, with the global theorems as landscape motivation rather than as covering theorems for the reported trajectories.","section":null}],"minor_comments":[{"comment":"The diagnostic active-set selector (§3.1.2: 15% margin/velocity quantiles, |xi|<0.98, cap max{32,0.15n}) is used for probe sweeps but not ablated. A short ablation of margin quantile, saturation threshold, and velocity filter would clarify how much of the probe-sweep breadth is selector-specific.","section":null},{"comment":"Cross-variant Table 4 uses a different protocol (Goto Table S2 step counts and probe ratios) than the main all-Gset 1200-step study; the text notes this, but a single sentence in the table caption restating the protocol difference would prevent misreading the 52/54 vs 71/71 coverage gap.","section":null},{"comment":"Theorem 2’s explicit α* is intentionally non-sharp and can be large; a brief remark on typical numerical scales of R*, M*, Δ* for G-set-sized instances would help readers judge practical relevance of the large-α aSB recovery statement.","section":null},{"comment":"Figure 2 panel labels and the log10(1+ΔE) energy-gap plots are useful; ensure axis labels and the “mean first freeze” markers remain legible in grayscale print.","section":null},{"comment":"Notation: S vs J for the coupling matrix is identified in the introduction but both appear in early bSB equations; a single consistent symbol after the identification would reduce friction.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The mathematical core (frozen-set identity + robust bSB freezing certificate) is sound and the online same-seed protocol is carefully designed. The main risk for the journal is over-reading the dense FLOPs proxy as a practical complexity result on sparse G-set graphs, and the field-dependence of the empirical success. I would not reject on those grounds if the authors tighten scoping in the abstract/intro and add a sparse-cost or wall-clock caveat; the paper is a legitimate contribution to continuous Ising solver methodology. Fit for math.NA / computational optimization is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful core here is not another SB variant. It is an exact rewrite: once a set of signs is fixed, the Ising energy on the rest is the same Ising problem with an induced field μ_eff = μ_Q + S_QH v_H. They pair that with a local robust-margin certificate for the actual bSB Euler+clip step, then run an online freezes-only-when-certified procedure that, with external field, matches the same-seed full bSB energy in every reported run while cutting a large fraction of dense matvec work.\n\nWhat is new is the package: external-field large-α aSB recovery with an explicit threshold, soft-to-hard epi-convergence and hard-box reduction under a PSD condition, the frozen-set identity, and the one-step/persistent freezing corollaries for clipped bSB. The appendix proofs are explicit and coherent. Empirically they are careful: online checks every 50 steps, no baseline used to pick freeze times, 540 Table-S2 and 1420 all-Gset seed–instance runs with η_λ=0.2 all trigger and none degrade. Probe sweeps and aSB/dSB/SimCIM transfers are labeled diagnostic. Zero-field contrast is reported honestly (certificate fires less, quality preservation collapses).\n\nSoft spots, in proportion. The global hard-box PSD hypothesis does not cover the G-set schedules (they say so: G1 needs α≳7 while the schedule has α≤1). Quality preservation therefore rides entirely on the local certificate and the box bound on unresolved coordinates—not on the landscape theorems. That is fine if you read the paper as written; it is not a hidden gap. The real soft spot is the compute claim: savings are a dense 2n² proxy that excludes selection, submatrix extraction, memory traffic, and sparsity. G-set graphs are sparse, so a production sparse bSB already costs O(m) per step; the abstract’s “64.4% of the dense interaction work” is accurate for that proxy and not yet a wall-clock or sparse-runtime result. Code is promised on publication, not shipped. Free parameters (field scale, check interval, diagnostic quantiles) are standard for a methods paper and do not undercut the certified online rule.\n\nThis is for people who run continuous/quantum-inspired Ising solvers and care about late-stage cost. The citation pattern is normal for the SB/G-set literature. I would send it to peer review; the mechanism is clear, the local theory matches the algorithm, and the same-seed evidence is strong enough to deserve referees. Engage if you work on SB-type solvers or active-set ideas for continuous combinatorial dynamics; treat the FLOPs number as a dense proxy until runtime data appear.","headline":"Clean algebraic frozen-set reduction for SB solvers with strong same-seed G-set evidence; the math is solid, the savings claim is still a dense FLOPs proxy.","tokens_in":31958,"tokens_out":666,"would_cite":true,"duration_ms":6542,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65K10","90C27","82B20"],"pacs":[],"model":"grok-4.5","headline":"Late-stage continuous Ising dynamics collapse to a smaller active subsystem whose frozen spins fold exactly into an induced field, cutting most dense work without changing the discrete objective.","keywords":["Ising model","simulated bifurcation","quantum-inspired optimization","variational reduction","frozen-set method","active-set reduction","combinatorial optimization"],"falsifier":"Run the online certified iSTAR procedure on the same G-set instances and seeds; if any certified reduced trajectory returns a strictly worse Ising energy than its same-seed full ballistic baseline, or if the robust-margin check freezes a coordinate that later flips under the full dynamics, the central claim that certified freezing preserves the discrete objective fails.","tokens_in":31660,"feed_emoji":"❄️","tokens_out":1152,"duration_ms":9143,"temperature":0.7,"pith_summary":"Continuous Ising solvers such as simulated bifurcation evolve a high-dimensional continuous state and read out binary spins by sign, but each step costs a dense matrix-vector product that scales with the square of the number of variables. This paper argues that the cost is not fundamental: once many coordinates have chosen a stable sign branch, they can be removed exactly rather than merely ignored. Fixing those signs turns their couplings into an effective external field on the remaining variables, so the original discrete problem reduces to an equivalent lower-dimensional Ising problem on the active tail. The authors prove continuous-to-discrete recovery for the soft quartic landscape and for a hard-box confinement model, and they supply a finite-time robust-margin certificate that keeps saturated coordinates frozen under the clipped ballistic update. Their algorithm iSTAR detects certified frozen coordinates online, folds them into the induced field, and continues only on the unresolved subsystem. On the G-set benchmark with external fields the certified procedure matches the same-seed full baseline in every run while removing, on average, more than sixty percent of the dense interaction work. A sympathetic reader cares because the O(N^{2}) bottleneck has long been treated as the price of continuous Ising dynamics; if the collapse is real and the certificate is reliable, late-stage computation can be cut sharply without changing the combinatorial objective.","feed_headline":"Continuous Ising solvers cut 64% of late dense work by freezing spins","feed_subtitle":"A frozen-set identity turns saturated coordinates into an induced field, preserving the same-seed baseline on G-set.","key_machinery":"The frozen-set identity: for a frozen set H with fixed signs v_H, the Ising energy restricted to H is equivalent to a reduced energy on the active complement Q whose induced field is µ_eff = µ_Q + S_QH v_H. A robust freezing margin ρ_i(k;H) = D_i^H(k) − B_i^Q then certifies that a saturated coordinate with zero momentum stays fixed under the next clipped ballistic step, so the reduction can be applied online.","core_discovery":"During late-stage simulated bifurcation the trajectory collapses onto a lower-dimensional active subspace, and saturated coordinates can be eliminated exactly by a variational frozen-set identity: once a subset of signs is fixed, their couplings become an induced external field on the unresolved subsystem, reducing the original Ising instance to an equivalent lower-dimensional problem. An online certified implementation (iSTAR) that freezes only when a robust-margin condition holds preserves the same-seed full baseline on G-set and removes on average 64.4 percent of the dense interaction work.","pith_inferences":["If the induced-field identity is exact, hardware accelerators for continuous Ising solvers could hard-wire progressive variable elimination once saturation is detected, turning late-stage O(N²) work into sparse or low-rank updates.","The gap between the global PSD hard-box theorem and the local certificate used on G-set suggests a natural next theorem: a finite-time basin-stability result that guarantees entry into the certified regime without assuming the spectral condition.","Because the reduction preserves the discrete objective on the retained coordinates, it can be stacked with other classical Ising heuristics as a free late-stage accelerator rather than as a competing solver."],"forward_implications":["Once coordinates satisfy the robust-margin certificate they need not be iterated; subsequent steps cost only the active-set size squared rather than N squared.","The same induced-field reduction applies in principle to any continuous Ising solver that ends with a sign readout, not only to ballistic simulated bifurcation.","External fields create a more favorable reduction regime than the zero-field case: certificates trigger earlier and more uniformly when a symmetry-breaking field is present.","Diagnostic probe sweeps show that late-tail reducibility is broad across G-set graphs, so most instances admit a non-degrading reduced operating point well before the nominal terminal time."],"fun_headline_variants":["iSTAR freezes stable spins to cut 64% of dense Ising work","Late-stage collapse lets continuous Ising solvers drop 64% interactions","Frozen-set identity reduces Ising dynamics to active subspace only","Variational freeze on saturated coordinates removes most dense cost","iSTAR certifies spin freezes and preserves G-set same-seed baselines"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"Quality preservation on the tested G-set schedules rests on the local robust-margin certificate activating and on unresolved coordinates remaining inside the box under a monotone schedule, not on the global hard-box recovery theorems that the paper itself notes do not hold at the benchmark parameters.","fun_headline_variants_meta":{"raw":{"variants":["iSTAR freezes stable spins to cut 64% of dense Ising work","Late-stage collapse lets continuous Ising solvers drop 64% interactions","Frozen-set identity reduces Ising dynamics to active subspace only","Variational freeze on saturated coordinates removes most dense cost","iSTAR certifies spin freezes and preserves G-set same-seed baselines"]},"model":"grok-4.5","effort":"low","cost_usd":0.006636,"raw_usage":{"total_tokens":1678,"prompt_tokens":762,"num_sources_used":0,"completion_tokens":96,"cost_in_usd_ticks":66360000,"prompt_tokens_details":{"text_tokens":762,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":820,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":762,"tokens_out":96,"duration_ms":6756,"temperature":1.0,"reasoning_tokens":820,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T21:50:08.275153+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the online certified iSTAR procedure on the same G-set instances and seeds; if any certified reduced trajectory returns a strictly worse Ising energy than its same-seed full ballistic baseline, or if the robust-margin check freezes a coordinate that later flips under the full dynamics, the central claim that certified freezing preserves the discrete objective fails.","supporting_citations":[],"review_version":1}