{"id":"165e7dcd-c14c-4290-bca5-9605d51d18fa","arxiv_id":"2607.27320","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A ~4,000-parameter recurrent local network correcting the Zeldovich approximation reaches percent-level matter power-spectrum accuracy at k≲0.5 h/Mpc at z=0, matching larger U-Net emulators on Quijote N-body tests.","lead":"This paper trains a compact neural 'cellular automaton' that steps cosmic particle positions forward in time using a single small, local rule, instead of running an expensive N-body simulation. If the accuracy holds up, it gives cosmologists a cheap, differentiable way to reverse-engineer the early universe from galaxy surveys.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Exact E(3) equivariance is overclaimed: finite-difference stencils on a cubic lattice are only octahedrally equivariant, and the paper's validation does not exercise arbitrary rotations; a rotation test is needed before accepting the symmetry guarantee.","rationale":"The central claim has two components: empirical accuracy (percent-level P(k), r(k) on held-out cosmologies) and architectural guarantees (strict locality, E(3) equivariance, interpretable KAN rule, parameter efficiency). The most load-bearing vulnerability is the equivariance guarantee because the paper's novelty and intended use—coordinate-frame-independent field-level inference—depend on it, and because it is internally contradicted by the finite-difference construction. A cubic finite-difference stencil is not a discretization that commutes with arbitrary rotations; it is equivariant only under O_h. The paper's statement 'mathematically guaranteed' is therefore overclaimed. The empirical power-spectrum validation is performed in a fixed lattice orientation and cannot detect this. The same caveat applies to scalar invariants built from these stencils; they are not a complete O(3) invariant basis. This does not by itself refute the percent-level accuracy claim—a model with only octahedral symmetry could still learn near-isotropic behavior for statistically isotropic fields, and the Zeldovich backbone handles long-wavelength modes. But it undermines the stronger 'guaranteed equivariant by construction' argument and the interpretable-rule conclusion. A direct rotation test can settle whether the symmetry violation is practically negligible or large enough to bias rotated inputs. If the test shows >1% power differences, the paper should be revised to claim octahedral equivariance plus empirical isotropy, and the field-level inference application becomes less safe. If the test shows negligible violation, the overclaim is cosmetic and the conditional acceptance can be upgraded. Since the reader already marked CONDITIONAL with medium correctness risk, the verdict should remain unchanged pending this check.","tokens_in":23332,"tokens_out":8356,"duration_ms":84712,"concrete_test":"Run a numerical equivariance audit on the trained model: take one validation initial condition Ψ_IC, compute the z=0 output. Apply a rotation by 45° (and 30°) about a lattice axis to Ψ_IC, resample to the original grid with trilinear interpolation, and compute the output of the rotated input. Compare to the rotated output of the original input. Quantify the normalized L2 discrepancy in the displacement field and the induced difference in P(k) at k≤0.5 h/Mpc. If the discrepancy exceeds the claimed percent-level accuracy (e.g., >1% power transfer at k≤0.5), then the E(3) guarantee is load-bearing for the accuracy claim; if it is below ~0.1%, the overclaim is cosmetic and the verdict can be upgraded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section II ('Equivariant Formulation') constructs the scalar pool and vector menu from finite-difference Laplacians/Hessians on a fixed cubic q-lattice and concludes that the synthesized updates 'are mathematically guaranteed to be E(3) rotationally equivariant' (Eqs. 4–5). This guarantee is not true as stated. A 3×3×3 finite-difference stencil is covariant only under the discrete octahedral group O_h of the lattice, not under continuous O(3). For R∈O(3) not in O_h, rotating the input field and then applying the stencil does not equal applying the stencil and then rotating the output; the stencil couples lattice directions whose rotated images are not lattice directions. Hence the invariant pool is not rotationally complete and the learned rule is orientation-dependent. The central contribution—that a compact, exactly equivariant local rule emulates non-linear structure formation and supports coordinate-frame-independent field-level inference—rests on this property. The paper's percent-level figures are computed in the training lattice orientation and therefore cannot detect the violation. Appendix C acknowledges a related finite-receptive-field ceiling on halo morphology, but the rotational symmetry breaking is an additional, unacknowledged limitation. The accuracy claim may survive, but the 'guaranteed equivariance' claim must be weakened or verified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the Lagrangian Neural Cellular Automaton (LNCA), a recurrent, strictly local Lagrangian-frame emulator trained on Quijote Latin Hypercube simulations to predict the residual displacement correction beyond the Zeldovich approximation. The model iterates 32 substeps from z=127 to z=0, with a KAN acting on invariant scalar features and modulating an equivariant vector menu. Validation covers per-particle displacement residuals, power spectra, cross-correlation coefficients, halo mass functions, and a lightcone demonstration, with claims of percent-level fidelity for k ≲ 0.5 h/Mpc at z=0 using only ~4,000 learned parameters.","tokens_in":23752,"tokens_out":9186,"duration_ms":78829,"significance":"If the central claims withstand scrutiny, this is a significant contribution: a ~4,000-parameter recurrent local rule that produces continuous, differentiable trajectories and approaches the accuracy of a 2×10^7-parameter U-Net emulator would be a valuable building block for field-level inference and lightcone construction. The Zeldovich residual decomposition, the fixed Lagrangian connectivity, and the interpretable KAN update are genuine design strengths, and the paper is careful to compare against an established baseline and to test on held-out cosmologies. However, two load-bearing issues must be addressed before the claims as stated can be accepted: the exact E(3) rotational equivariance is not established by the finite-difference construction, and the halo validation is partly circular because the halo loss directly fits the moments used for validation.","major_comments":[{"comment":"The central symmetry claim is overclaimed. The scalar pool and vector menu are built from finite-difference Laplacians, Hessians, and vector products on a fixed cubic q-lattice. Such stencils are equivariant only under the discrete octahedral group of the lattice, not under continuous O(3). For a rotation R outside O_h, the finite-difference derivatives of the rotated input are not the rotated versions of the unrotated derivatives, so the statement that the synthesized updates are 'mathematically guaranteed to be E(3) rotationally equivariant' does not follow. The quoted percent-level validation is performed in the training lattice orientation and cannot detect this symmetry breaking. Please either weaken the claim to discrete octahedral equivariance and add a concrete rotation test (e.g., rotating initial conditions by an angle not a multiple of 90° and checking that all reported metric","section":"§II 'Equivariant Formulation', Eqs. (4)–(5); abstract"},{"comment":"The halo validation is partly circular. L_Halo is defined by fitting the center-of-mass position, bulk momentum, and position/velocity dispersions of target-identified halos; the validation in Fig. 12 therefore rewards precisely the moments used in training. Additionally, because L_Halo averages only over target halos, it contains no penalty for spurious collapsed objects, so the halo mass function comparison is not an independent test of the model's halo-collapse behavior. Please report halo statistics for an ablation trained without L_Halo, or explicitly state which halo properties are inherited from the loss and which are genuinely emergent.","section":"§II Eq. (11) and §III 'Halo Population'"},{"comment":"The phrase 'guaranteeing accuracy at large scales' is stronger than the architecture supports. The decomposition Ψ = D(t)Ψ_IC + Ψ_res is an excellent physics prior, but the learned residual is unconstrained and could in principle alter large-scale modes; large-scale fidelity is enforced by the training data, not by construction. Please replace 'guarantee' with a more precise formulation, e.g., 'enforce via the Zeldovich prior', or add an architectural constraint that suppresses long-wavelength residual power. This also affects how the low-k percent-level transfer function results should be interpreted.","section":"Abstract and §II Eq. (2)"}],"minor_comments":[{"comment":"The manuscript contains the stray text 'super secret hidden text' after Table II, and repeated 'this text left intentionally' in Appendix C. These placeholders must be removed before submission.","section":"Appendix B, Table II area"},{"comment":"The halo mass function plot has no error bars, although the text claims agreement 'within the Poisson scatter of the reference'. Please add Poisson uncertainties or otherwise quantify the comparison.","section":"Fig. 12 and §III 'Halo Population'"},{"comment":"The transfer-function and cross-correlation plots do not show error bars or sample variance. If these are single realizations, state this explicitly and, where possible, show the scatter across test realizations.","section":"Eq. (12) and Figs. 10–11"},{"comment":"The abstract says '~10^4 times fewer learned parameters', but the numbers quoted in §IV give 2×10^7/4×10^3 = 5×10^3. This is acceptable as an order-of-magnitude statement, but 'four orders of magnitude' is the more precise phrasing and avoids an apparent factor-of-two discrepancy.","section":"§IV 'Comparison to Existing Emulators'"},{"comment":"A code/data availability statement is absent. Given the reproducibility-focused claims of the paper, please provide a link to the trained model, training code, or at least pseudocode for the update rule.","section":"General"},{"comment":"The text has small typographical issues, including 'consrtained' in §II and 'using using' in §II. A careful proofread is needed.","section":"§II and §V"}],"recommendation":"major_revision","confidential_remarks":"The E(3) equivariance overclaim is the main technical gate; the accuracy claims may well survive once the claim is weakened or verified with a rotation test. The halo-loss circularity also needs explicit acknowledgment or an ablation. The stray placeholder text in the appendix is an editorial red flag that should be cleaned before any further review. I do not see a citation or novelty concern beyond the normal need to position against [31]."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know up front. The architecture is genuinely new: a fixed-connectivity Lagrangian-frame cellular automaton with a KAN polynomial update and a residual Zeldovich decomposition. The reported ~4k parameters and percent-level spectra are plausible and worth taking seriously. But the 'mathematically guaranteed E(3) equivariance' is not true as stated, and the halo validation is partially circular.\n\nWhat the paper does well: it identifies a real gap—a temporally continuous, differentiable, local forward model for field-level inference—and the LNCA combines known pieces (NCA, fixed-connectivity graph CA, residual LPT, KAN) in a way I haven't seen. The continuous trajectory output and native lightcone construction are genuine advantages over single-shot U-Nets. The interpretability analysis in the appendix is a good-faith attempt to open the black box, and the paper is unusually candid about its limitations, e.g., the finite receptive-field ceiling on internal halo morphology.\n\nWhere it gets soft. First, the equivariance claim. Section II constructs invariants from finite-difference Laplacians/Hessians on a fixed cubic lattice and then asserts the synthesized updates are 'mathematically guaranteed to be E(3) rotationally equivariant.' That guarantee is wrong. A 3x3x3 stencil is covariant only under the lattice's octahedral group, not continuous O(3). Rotating the input and applying the stencil does not commute with applying the stencil and then rotating the output unless the rotation preserves the lattice. The paper never tests arbitrary rotations, so the model's behavior under rotations is unknown. The empirical accuracy is computed in the training orientation, so it isn't invalidated, but the central methodological selling point needs to be either verified by a rotation test or downgraded to discrete cubic symmetry. That is a load-bearing flaw in the claims, not a cosmetic one.\n\nSecond, the halo validation. The L_Halo loss directly fits halo centroids, bulk momenta, and internal dispersions. The HMF is not identical to those moments, but the model is explicitly rewarded for forming compact, dense clumps at the right locations with the right velocity dispersion, so the good HMF agreement is not an independent confirmation. Third, the fiducial results are, as far as I can tell, from a cosmology in the training set; the held-out hypercube cosmologies are shown, but the quantitative residuals for those are not broken out with the same care as the fiducial case. Fourth, no code or data is released, and the manuscript contains placeholder strings ('super secret hidden text', 'this text left intentionally') that suggest it is not a final draft. These are fixable but matter for reproducibility.\n\nOverall: this is a promising direction, the paper is honestly written, and the empirical results, if reproducible, are a real step forward in parameter efficiency for differentiable emulators. But the equivariance guarantee needs a rotation test or a corrected statement, the halo claim needs an independent validation, and the fiducial-vs-held-out distinction needs to be drawn clearly. I'd send it to peer review, not desk reject, with a request for the rotation test and code release. It would be a great reading-group example of how physics-informed symmetry claims can overshoot what the discretization actually guarantees.","headline":"A genuinely new NCA-based emulator with a real equivariance overclaim; worth refereeing but expecting heavy revision.","tokens_in":24177,"tokens_out":3792,"would_cite":true,"duration_ms":33702,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A strictly local, recurrent rule of a few thousand learned coefficients can emulate non-linear cosmic structure formation at percent-level accuracy while returning full particle trajectories.","keywords":["cosmological emulation","Lagrangian neural cellular automaton","Zeldovich approximation","equivariant neural networks","differentiable forward model","field-level inference","large-scale structure","N-body surrogates"],"falsifier":"Rotate a held-out initial-condition box by 30 degrees about a lattice axis, evolve it with the trained model, and compare the power spectrum and cross-correlation to the unrotated evolution: if the equivariance is only octahedral, mode phases will disagree at some k and the declared percent-level budget will be exceeded. Alternatively, double the integration substeps or neighborhood size and see whether the low-mass halo mass function rises toward the N-body reference.","tokens_in":23225,"feed_emoji":"🌌","tokens_out":6146,"duration_ms":52123,"temperature":0.7,"pith_summary":"The paper sets out to show that cosmic structure formation can be emulated not by a large convolutional network that maps initial densities to final states in one pass, but by a compact, strictly local update rule applied over many time steps in the Lagrangian frame. The model, called a Lagrangian Neural Cellular Automaton, learns only the residual acceleration that corrects the Zeldovich approximation, so large-scale linear behavior is built in and the network's capacity is reserved for shell crossing and virialization. The authors claim percent-level accuracy in the matter power spectrum and cross-correlation out to k≈0.5 h/Mpc at z=0, with as few as about 4,000 learned parameters, roughly four orders of magnitude fewer than comparable one-shot emulators. Because the rule is recurrent and integrates predicted time derivatives, the model produces continuous trajectories from z=127 to z=0 and can construct lightcone outputs naturally, making it a candidate differentiable forward model for field-level inference of initial conditions.","feed_headline":"A 4,000-parameter rule emulates cosmic structure to 1%","feed_subtitle":"A recurrent local rule tracks particle trajectories continuously, enabling fast differentiable cosmic field-level inference.","key_machinery":"The central object is the Lagrangian Neural Cellular Automaton: a lattice of nodes whose connectivity is fixed in Lagrangian coordinates q while nodes move in physical space; each node carries displacement, velocity, and latent scalar and vector fields. The learned update is a pointwise Kolmogorov-Arnold Network (a polynomial neural network whose activations are learned Taylor expansions) that maps an E(3)-invariant scalar pool, built from Laplacians, Hessians, and products of gradients, into coefficients that linearly combine an equivariant vector menu, yielding residual acceleration updates that are integrated over time. This construction enforces locality and equivariance by design and ma","core_discovery":"The paper's central claim is that non-linear gravitational structure formation can be emulated by a strictly local, recurrent update rule that corrects the Zeldovich approximation, and that this rule, with only a few thousand learned coefficients, reproduces N-body matter fields at percent-level accuracy in power and cross spectra to k≈0.5 h/Mpc at z=0, while returning complete particle trajectories and native lightcones. The learned dynamics are built from E(3)-invariant scalar features and equivariant vector menus, integrated with a symplectic scheme, and are claimed to be exactly rotationally and translationally equivariant, temporally continuous, and interpretable as an explicit polynomi","pith_inferences":["If the finite-difference stencils are only exactly equivariant under the discrete octahedral group, the claimed continuous E(3) equivariance is approximate; one could test this by rotating a held-out initial condition by a non-octahedral angle and checking whether r(k) degrades.","The bounded information propagation means the Jacobian of the emulator is local, so gradient-based sampling could exploit local likelihood curvature to reduce cost; the paper does not discuss this, but it follows directly from the finite-speed lightcone.","The acknowledged ceiling on internal halo morphology suggests a specific testable route: adding multi-scale neighborhoods or hierarchical message passing should preferentially raise the low-mass halo abundance if the deficit is caused by the receptive-field horizon."],"forward_implications":["At z=0 the emulator achieves <1% power-spectrum residuals for k≲0.5 h/Mpc and <1% phase decoherence for k≲0.7 h/Mpc, with better performance at higher redshifts.","Recovered halo populations match the N-body reference at the high-mass end, while the abundance of low-mass halos is underpredicted.","Because positions and velocities exist at every integration step, lightcone mocks require no additional passes, and velocities are evolved jointly with density, supporting kSZ-type observables.","The compact polynomial update rule can be inspected: attribution tables show the corrections are driven mainly by local shear amplitude and the coupling of a latent field gradient to the tidal field.","The model generalizes across a broad range of cosmologies, including held-out extremes of the training distribution, without retraining."],"fun_headline_variants":["Cosmic structure emulated by a 4,000-parameter local rule","Lagrangian cellular automaton hits 1% accuracy on cosmic web","Recurrent local rule predicts cosmic formation to percent level","Tiny differentiable rule mimics N-body gravity at 1% precision","Neural automaton produces full cosmic trajectories with few params"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The central assumption is that the local derivative features computed on a fixed cubic Lagrangian lattice are sufficient to capture the full non-linear dynamics, so that a strictly local update with a 32-substep horizon can form collapsed halos at percent-level accuracy; if shell crossing requires information from beyond that stencil, the halo and percent-level claims fail.","fun_headline_variants_meta":{"raw":{"variants":["Cosmic structure emulated by a 4,000-parameter local rule","Lagrangian cellular automaton hits 1% accuracy on cosmic web","Recurrent local rule predicts cosmic formation to percent level","Tiny differentiable rule mimics N-body gravity at 1% precision","Neural automaton produces full cosmic trajectories with few params"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000216,"raw_usage":{"total_tokens":1321,"prompt_tokens":847,"completion_tokens":474,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":591,"completion_tokens_details":{"reasoning_tokens":386}},"tokens_in":591,"tokens_out":474,"duration_ms":7152,"temperature":1.0,"reasoning_tokens":386,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T09:32:40.283916+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rotate a held-out initial-condition box by 30 degrees about a lattice axis, evolve it with the trained model, and compare the power spectrum and cross-correlation to the unrotated evolution: if the equivariance is only octahedral, mode phases will disagree at some k and the declared percent-level budget will be exceeded. Alternatively, double the integration substeps or neighborhood size and see whether the low-mass halo mass function rises toward the N-body reference.","supporting_citations":[],"review_version":1}