{"id":"60329e7e-b37e-4b6d-bc49-74afb2b99228","arxiv_id":"2512.01558","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"Training networks on residual corrections after subtracting the Keplerian solution cuts validation error 28-54x versus learning full trajectories, but the paper's key claim of a sharp capacity transition at chaos onset is undermined by internal inconsistencies between the abstract and full text.","lead":"This paper proposes Neural Network Perturbation Theory (NNPT), training networks to predict only the residual difference between a full three-body simulation and the analytic Keplerian orbit, rather than the whole trajectory. On a Sun-Earth-Jupiter testbed with artificially scaled Jupiter mass, the authors report that the network capacity needed for fixed accuracy jumps sharply near the onset of chaos, but the abstract and main text disagree on the shape of this transition.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equalized-accuracy protocol breaks down in chaotic regime: at f≥20 (notably f=30) no tested architecture may actually reach E_thr, so the reported sevenfold capacity jump is a best-effort failure, not a minimal capacity achieving fixed accuracy.","rationale":"The reader's weakest-assumption identification is exactly the load-bearing concern: the equalized-accuracy protocol's threshold is not demonstrated to be reachable in the chaotic regime, and the paper's fallback rule ('if no architecture meets the threshold, report the best-performing configuration') allows Table II to report capacities that are not successes. This is not merely a minor caveat; the sevenfold jump at f_c=15.6 is the paper's headline quantitative result, and it depends on P(f) being the minimal capacity that achieves E_thr. If P(f) for f≥20 is instead the best-performing failure, the transition could be an artifact of the coarse architecture grid and the arbitrary threshold, not a fundamental capacity barrier imposed by chaos. In addition, there is a direct internal contradiction between the abstract (non-monotonic capacity, peak at f=5 with 3×32, decrease to 2×32 at f≥17) and the full text/Table II (monotonic 2×32 for f≤15, 3×64 for f≥20). This contradiction compounds the protocol issue, but the protocol breakdown alone is sufficient to undermine the central claim. The residual-learning advantage (28–54× lower MSE) is a credible, separable contribution, but it does not rescue the capacity-transition conclusion. Thus the reader's REJECT verdict stands without modification.","tokens_in":15789,"tokens_out":4639,"duration_ms":47939,"concrete_test":"Re-run the capacity-shrinking protocol at f=20, 25, and 30 for all seeds and all architectures in Table I, recording validation MSE for each architecture. Check specifically whether the picked 3×64 architecture has validation MSE ≤ E_thr = 1.06×10^-2 AU² within 3000 epochs; also check whether any larger architecture meets the threshold. If at f=30 no architecture reaches E_thr, then the reported P(f)=8,578 is a best-effort failure rather than an equalized-accuracy success, and the capacity-jump conclusion for the chaotic regime is invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of a sevenfold capacity jump at f_c=15.6 rests on the equalized-accuracy protocol (Section II.H). The protocol states: 'If no architecture meets the threshold, we report the best-performing configuration.' Figure 3 shows the reference 10×128 network itself crosses the acceptance threshold E_thr=1.06×10^-2 AU² at f≈20. For f=30, where Table II reports 3×64 (8,578 parameters) as the picked architecture, the paper never demonstrates that any architecture actually achieves MSE ≤ E_thr. Table II lists only parameter counts, not the validation MSE of the picked networks. Thus the 'minimal capacity' at f=30 (and possibly f=20) may not be a capacity that reaches the target accuracy; it is merely the best-performing failure. Under those conditions, the jump from 2×32 to 3×64 does not establish a capacity barrier at fixed accuracy—it may simply be the smallest architecture in the grid whose performance is closest to an unmet threshold. Because this protocol failure affects the chaotic regime—the very region used to define the capacity barrier—the central claim is not supported by the evidence presented.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Neural Network Perturbation Theory (NNPT), in which MLPs are trained to predict only the residual y(t)=r_full-r_Kepler after subtracting the analytic Keplerian solution, using the planar circular restricted three-body problem as a testbed. The Jovian mass factor f is varied from 0.05 to 30. A reference 10×128 network defines a global target E*=9.64×10^-3 AU^2 (median validation MSE across all f and seeds), and an equalized-accuracy protocol shrinks the architecture until E_thr=1.06×10^-2 AU^2 is met. The full text claims a sharp sevenfold minimal-capacity jump (1,186→8,578 parameters, 2×32→3×64) and a threefold training-time increase at f_c=15.6±1.0, aligned with Chirikov's resonance-overlap criterion, while autoencoder analysis claims a fourfold intrinsic-dimensionality jump (d_min: 8→32) at the same f_c. The abstract, however, reports a non-monotonic capacity profile peaking at f=5 and decreasing to 2×32 at f≥17. The paper concludes that NNPT provides a general parameter-efficient surrogate strategy and that physical chaos imposes fundamental capacity barriers on fixed-architecture networks.","tokens_in":16104,"tokens_out":7492,"duration_ms":74540,"significance":"The NNPT decomposition idea is attractive, and the direct comparison showing 28–54× lower MSE for correction learning versus full-trajectory learning is a useful sanity check. The use of a symplectic integrator with energy-drift monitoring is also methodologically sound. However, the central quantitative claims are not supported by the evidence as presented. The equalized-accuracy target is derived from the median error of the same reference network that is later shown to exceed the acceptance threshold in the chaotic regime; the capacity values reported for f=20 and f=30 may be best-effort failures rather than capacities that achieve the stated fixed accuracy. This, together with the direct contradiction between the abstract and the full text, means the claimed chaos-onset capacity barrier and its Chirikov alignment are not established. If the protocol were repaired and the inconsistencies resolved, the underlying research question would remain interesting, but the current manuscript does not deliver a reliable answer.","major_comments":[{"comment":"The headline result is internally contradictory. The abstract states that capacity 'peaks at f=5 in the late integrable regime (3×32, 2242 parameters), remains elevated through the transition region, then decreases in the fully chaotic regime (f≥17, requiring only 2×32 with 1186 parameters)—a 47% reduction from peak.' The full text (Fig. 4, Table II) instead reports a monotonic sevenfold jump at f_c=15.6, with f≥20 requiring 3×64 (8,578 parameters). The abstract also quotes f_c=16.6±2.8 while the body uses f_c=15.6±1.0. These are mutually incompatible descriptions of the paper's central observation; as written, the paper does not have a well-defined headline result.","section":"Abstract vs. §III.B and Table II"},{"comment":"The equalized-accuracy protocol fails exactly in the regime used to define the capacity barrier. In §II.H, E* is the median validation MSE of the 10×128 reference over all (f,seed) pairs, and E_thr=1.06×10^-2 AU^2. Figure 3 shows the reference network itself has MSE above E_thr for f≈20 and larger. The protocol states that if no architecture meets the threshold, the best-performing configuration is reported. Table II lists 3×64 for f=20 and f=30 but gives no validation MSE for these choices. Therefore P(f)=8,578 may be the smallest grid architecture whose performance is closest to an unmet threshold, not the minimal capacity that actually reaches E*. The sevenfold jump is thus not evidence of a capacity barrier at fixed accuracy.","section":"§II.H, Fig. 3, Table II"},{"comment":"The autoencoder dimensionality claim has the same threshold problem. The reconstruction criterion is E_AE_thr=10^-4 AU^2, but Fig. 7(a) reports that at f=30 even d=64 gives E_rec≈5×10^-4 AU^2, above the threshold. Consequently, the values d_min≈32–40 shown in Fig. 7(b) for f≥20 are not minimal dimensions that meet the stated reconstruction criterion; they are best-effort values at the largest tested latent dimension. The claimed fourfold dimensionality jump (8→32) at f_c=15.6 is therefore not supported by the data presented.","section":"§II.J.3, Fig. 7"}],"minor_comments":[{"comment":"Several numerical values are inconsistent and need reconciliation: f_c=16.6±2.8 (abstract) vs. 15.6±1.0 (body); energy drift bound 2×10^-4 (abstract) vs. 2×10^-7 (body); 10×128 parameter count 168,322 (§II.E and Fig. 3) vs. 149,122 (Table I and Eq. (8)). The abstract's ||y2||/||y1||≈0.997 also disagrees with §III.C, which reports the second-stage residual is 13–19% larger.","section":"Throughout"},{"comment":"The architecture grid is not strictly increasing in size: 2×48 has 2,546 parameters while 3×32 has 2,242. Section II.H says architectures are tested 'in order of increasing size,' but the order between these two is ambiguous. This could affect which architecture is picked and should be clarified.","section":"§II.H, Table I"},{"comment":"The caption marks 'Transition fc = 20.0', while §III.B reports f_c=15.6. Please clarify which quantity is being indicated.","section":"Fig. 3 caption"},{"comment":"The reproducibility statement says data and code 'can be obtained from the corresponding author upon request.' For a computational study whose central claims rely on specific training runs and threshold choices, a public repository would materially strengthen verifiability.","section":"Data and Code Availability"}],"recommendation":"reject","confidential_remarks":"For the editor: the reader's strongest concern is valid and, on reading the manuscript, it lands. The equalized-accuracy protocol's target is self-referential and unmet in the chaotic regime, and the abstract/body contradiction is severe. The paper also contains multiple unreconciled numerical values. The load-bearing threshold failure means the central claim is not supported, and I do not see how a revision short of re-running the core experiments with a properly defined, per-mass accuracy target could fix it. The NNPT idea itself has merit and could be resubmitted after such a redesign and after all reported quantities are made consistent."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: don't trust the central claim yet. The abstract reports a non-monotonic capacity profile peaking at f=5 and needing only 2×32 (1,186 params) at f≥17; the full text reports a monotonic sevenfold jump to 3×64 (8,578 params) at f_c=15.6. The abstract even quotes the sequential-correction ratio as ~0.997 while the body says the second-stage residual is 13–19% larger. That is not a wording slip; it's the paper's main result changing between front matter and body.\n\nCredit where due: the correction-learning comparison is solid and useful. With identical 2×32 networks, residual learning beats direct trajectory learning by 28–54× in validation MSE. That is a real, reproducible-style result and consistent with existing delta-learning literature (refs 6, 21), so the framing is not new but the controlled demonstration is clean. The symplectic integrator energy check is good. The autoencoder dimensionality jump is a nice idea, though it inherits the same f-region problems.\n\nSoft spots are serious. The equalized-accuracy protocol defines E* as the median validation error of the 10×128 reference network, then asks smaller networks to meet it. Figure 3 shows that same reference network crosses the threshold at f≈20; at f=30 it is well above. Table II nevertheless reports 3×64 as the 'picked' architecture for f=20 and 30, without showing the validation MSE of the picked networks. Since the protocol says 'if no architecture meets the threshold, we report the best-performing configuration,' those points are likely best-effort failures, not capacities that achieve fixed accuracy. The sevenfold jump from 2×32 to 3×64 may just be the grid's next-smallest architecture after the threshold stops being reachable. The coarse grid (2×32, 2×48, 3×32, 3×64...) and arbitrary 10% tolerance can manufacture a step.\n\nAlso minor: Table I counts 10×128 as 149,122 parameters while the text says 168,322; the formula gives 149,122.\n\nBottom line: The residual-learning contribution is worth keeping, but the chaos-capacity barrier claim is not supported by the evidence as written. This deserves a serious referee only if the authors can resolve the abstract/body contradiction and show actual MSE values for the picked architectures in the chaotic regime. I would engage with a revised version; I would not cite the capacity transition in its current form.","headline":"The paper's headline claim—a sevenfold capacity jump at f≈15.6—is contradicted by its own abstract, and the equalized-accuracy protocol that produces it breaks down exactly in the chaotic regime.","tokens_in":16579,"tokens_out":3112,"would_cite":false,"duration_ms":34767,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["05.45.-a","07.05.Mh"],"model":"deepseek-v4-flash","headline":"The paper claims that neural networks trained on only the residual perturbation of a three-body system hit a sharp capacity barrier at chaos onset, requiring about seven times more parameters once the Jovian mass factor exceeds roughly 15.6","keywords":["neural network perturbation theory","residual learning","three-body problem","capacity transition","chaos onset","KAM tori","autoencoder","equalized-accuracy protocol"],"falsifier":"Check whether any architecture from the grid achieves validation MSE below E_thr=1.06e-2 AU^2 at f=20 and f=30; if none does, the 8,578-parameter figure is not a capacity that meets the equalized-accuracy target. Also compare the abstract's claim that f≥17 requires only 2×32 with Table II, which reports 3×64 at f≥20; both cannot describe the same protocol.","tokens_in":15642,"feed_emoji":"🪐","tokens_out":8771,"duration_ms":90493,"temperature":0.7,"pith_summary":"The paper introduces Neural Network Perturbation Theory (NNPT): instead of training a network to reproduce an entire trajectory, subtract the exactly solvable Keplerian orbit and train the network to predict only the residual. Applying this to a Sun–Earth–Jupiter system with Jovian mass varied from 0.05 to 30 times its physical value, the authors report that the minimal network capacity needed to hold validation error fixed jumps sharply from about 1,200 to about 8,600 parameters at a mass factor of about 15.6. At the same point, the latent dimension required to autoencode the residual jumps from about 8 to about 32. The authors interpret this as a fundamental capacity barrier at the onset of chaos, aligned with Chirikov's resonance-overlap criterion and preceding visible KAM torus destruction.","feed_headline":"Chaos onset forces a sevenfold neural-network capacity jump","feed_subtitle":"At Jovian mass f≈15.6, minimal network size jumps from ~1.2k to ~8.6k parameters — a sharp, physics-driven barrier.","key_machinery":"The load-bearing object is the residual signal y(t)=r_full(t)−r_exact(t), which isolates the three-body correction from the analytically known Keplerian baseline. The load-bearing protocol is equalized accuracy: compute a reference threshold E* from a large 10×128 network, then shrink architecture until that threshold is met, recording the smallest architecture that succeeds. A piecewise-linear change-point fit locates f_c; a convolutional autoencoder measures the minimal latent dimension d_min of y(t) as an independent geometric diagnostic.","core_discovery":"On the paper's own terms, the central discovery is that a fixed-architecture neural network's minimal capacity to achieve a fixed error is not a smooth function of a physical control parameter. In the integrable regime (f≤15) every seed hits the target with a 2×32 MLP (1,186 parameters); in the chaotic regime (f≥20) all seeds need a 3×64 MLP (8,578 parameters). The breakpoint f_c=15.6±1.0 is selected by piecewise-linear regression with Bayesian information criterion; the 68% bootstrap confidence interval spans only two mass factors. Independent autoencoder analysis gives a fourfold jump in minimal latent dimension at the same f_c. Energy drift below 2×10^-7 across all f is used to argue the","pith_inferences":["The internal contradiction between the abstract's non-monotonic capacity profile (peak at f=5, fall to 2×32 at f≥17) and the full text's monotonic jump (2×32 for f≤15, 3×64 for f≥20) means the reported 'minimal capacity' values are protocol-sensitive; a reader should verify which numbers correspond to the published tables.","If the capacity barrier is truly about the dynamics, the same discontinuity should appear for other fixed-capacity approximators such as polynomials or Fourier series; running the equalized-accuracy sweep with a non-neural regressor would test the claim without relying on ML training details.","One could sharpen the Chirikov connection by computing resonance-overlap widths for the specific Sun–Earth–Jupiter parameters and comparing the predicted f range to the measured breakpoint, rather than relying on the approximate f~10–16 quoted.","For practical surrogate design, this suggests a cheap diagnostic: sweep a physical parameter, hold target error fixed, and watch for a capacity step; systems with such steps may require hybrid physics-plus-ML models."],"forward_implications":["In regimes where dynamics are chaotic, generic black-box surrogates will need sharply more parameters or training time; correction learning that subtracts exact baselines avoids the worst of this barrier.","Capacity metrics under equalized accuracy can serve as an early-warning diagnostic for chaos onset, detecting functional complexity before Poincaré sections show torus destruction.","NNPT's recipe — learn the residual after an exact solution — transfers to any system with a known analytic approximation, potentially cutting parameter counts by orders of magnitude.","Sequential two-stage corrections did not refine the result, meaning naive hierarchy does not circumvent the capacity barrier; better architectural priors are needed.","The autoencoder dimensionality transition gives an unsupervised way to locate the same critical point, independent of the supervised capacity sweep."],"fun_headline_variants":["Chaos onset multiples neural network size sevenfold","Minimal net capacity jumps 7× at chaos transition","Neural nets need 7x parameters as chaos kicks in","Fixed 1% error: chaotic systems demand 7x network width","Capacity cliff: chaos onset forces neural net expansion"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The protocol assumes that the single global error threshold E*≈9.64e-3 AU^2 is an attainable target at every mass factor, and that the coarse architecture grid can find the true minimal capacity; if no architecture actually reaches the threshold in the chaotic regime, the reported minimal capacities are best-effort misses rather than successes.","fun_headline_variants_meta":{"raw":{"variants":["Chaos onset multiples neural network size sevenfold","Minimal net capacity jumps 7× at chaos transition","Neural nets need 7x parameters as chaos kicks in","Fixed 1% error: chaotic systems demand 7x network width","Capacity cliff: chaos onset forces neural net expansion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000444,"raw_usage":{"total_tokens":2136,"prompt_tokens":847,"completion_tokens":1289,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":591,"completion_tokens_details":{"reasoning_tokens":1222}},"tokens_in":591,"tokens_out":1289,"duration_ms":11221,"temperature":1.0,"reasoning_tokens":1222,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T19:10:11.961536+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Check whether any architecture from the grid achieves validation MSE below E_thr=1.06e-2 AU^2 at f=20 and f=30; if none does, the 8,578-parameter figure is not a capacity that meets the equalized-accuracy target. Also compare the abstract's claim that f≥17 requires only 2×32 with Table II, which reports 3×64 at f≥20; both cannot describe the same protocol.","supporting_citations":[],"review_version":1}