{"id":"b9ea68d5-1fb4-41f1-89db-2422397a4f09","arxiv_id":"2412.14795","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A real-space neural network wavefunction captures Landau level mixing in fractional quantum Hall systems and yields lower energies than lowest-Landau-level exact diagonalization at nu=1/3 and 2/5.","lead":"Using a neural network wavefunction, the authors compute fractional quantum Hall states on a sphere and obtain energies below lowest-Landau-level exact diagonalization by including Landau level mixing. The work suggests deep-learning variational methods can access correlation physics that standard exact diagonalization misses, but the LLM accuracy claim needs a stronger benchmark.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Energy gain below LLL ED is attributed to Landau level mixing, but with no multi-LL benchmark or symmetry diagnostic the physical accuracy of the LLM content (and the derived gaps) is unestablished.","rationale":"Good faith reading: the paper is a numerical method paper. The variational upper bound is valid, and the energy ordering below LLL ED necessitates some higher-LL weight. The small-kappa results (overlap 0.9976, energy close to ED, PCF matching Laughlin) are consistent with the method working in the weak-LLM limit. The calculation of NLLL via the 1-RDM appears correct. Credit is due for the design of the monopole-harmonic basis to control pole divergences and for using the same finite-size corrections in both NN and ED. However, the load-bearing step is the inference from 'lower energy than LLL ED' to 'physically accurate LLM'. Because the variational state is not constrained to the L=0 sector, a broken-symmetry variational minimum (e.g., a small Wigner cluster) could artificially lower energy and mimic LLM effects. The reader's implied test—compare against multi-LL DMRG or fp-DMC—is exactly the missing control. Given the small system sizes, such a benchmark is computationally accessible and would settle the matter. Until it is provided, the abstract's claim that LLM is captured 'up to a very high level' is stronger than the evidence. The reader's conditional verdict is appropriate; no change in verdict is needed.","tokens_in":12809,"tokens_out":10189,"duration_ms":74005,"concrete_test":"For N=6, nu=1/3 on a sphere at kappa=1 and kappa=3, perform ED in a truncated Hilbert space containing the lowest 3-4 Landau levels (single-particle angular momenta l=Q,...,Q+3) and compare the ground-state energy, NLLL, and overlap with the NN wavefunction. Also estimate <L^2> of the converged NN state. If the NN energy lies below the truncated-ED energy by more than the basis-truncation error, or if <L^2> is far from 0, physical LLM capture is not established; otherwise the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the NN variational energies in Figs. 2-3 being below LLL ED proves that the ansatz 'naturally capture[s] the extent of Landau level mixing up to a very high level' (Abstract). A variational argument guarantees the state is not LLL (otherwise ED would bound it), but it does not guarantee the higher-LL admixture is the physical one. The ansatz (Eqs. 3-5) is a single backflow determinant plus Jastrow; nothing constrains it to the L=0 sector of the rotationally invariant Hamiltonian. A finite-size variational state can lower energy by spontaneously breaking rotational symmetry (e.g., incipient Wigner crystallization), and the paper's only LLM diagnostic is the scalar NLLL (Fig. 3b), which cannot distinguish a correct LLM-corrected liquid from a symmetry-broken state with similar NLLL. The transport gap in Fig. 4c and the large-kappa phase-transition hints therefore rest on an unvalidated assumption. The appropriate reference—a converged multi-Landau-level ED or DMRG, or fixed-phase DMC result—is not supplied, and no code or data are released. The appendix's 1-RDM estimator is correct, but it is not benchmarked. This is the load-bearing gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a real-space neural-network variational Monte Carlo (NNVMC) approach for fractional quantum Hall systems on the sphere. The trial wavefunction (Eq. 3) is a Jastrow factor times a single Slater determinant whose multi-electron orbitals are built from LLL monopole harmonics multiplied by permutation-equivariant features from a Psiformer-type network. Parameters are optimized by minimizing the variational energy of the Hamiltonian in Eq. (1), so no exact-diagonalization data are used as labels. For nu=1/3 and 2/5 at kappa=1 with N=6 to 12, the neural-network energies are reported to be consistently lower than LLL-only exact diagonalization (ED). The paper also reports overlap with the Laughlin wavefunction, the number of electrons on the LLL, pair correlation functions, quasiparticle/quasihole densities, and a transport gap as functions of kappa. The central claim is that the real-space network naturally captures Landau level mixing and yields energies below LLL ED.","tokens_in":13077,"tokens_out":7505,"duration_ms":62370,"significance":"If the Landau-level-mixing content were validated, this would be a significant methodological advance: the method provides variational upper bounds below LLL ED, reaches N=12 where LLL ED is expensive, and produces a kappa-dependent transport gap with a plausible Wigner-crystal tendency at large kappa. The paper is honest in presenting the energy as a variational upper bound, and the overlap, NLLL, and pair-correlation diagnostics are evaluated after optimization rather than fit to ED, so there is no obvious circularity. However, the central LLM claim is currently unvalidated: there is no multi-Landau-level benchmark, no reported angular-momentum content of the ground state, and no release of code or data. The significance of the work therefore rests on a load-bearing assumption that the variational manifold selects the physical higher-Landau-level admixture.","major_comments":[{"comment":"The central claim that the real-space neural-network wavefunction 'naturally captures the extent of Landau level mixing' is not supported by any multi-Landau-level reference calculation. Energies below LLL ED prove only that the optimized state is not entirely in the LLL; they do not prove that the higher-LL component is the physical one. The only quantitative LLM diagnostic, NLLL in Fig. 3b (and Table S3), is a scalar and cannot distinguish a correct LLM-corrected liquid from a state with spurious higher-LL admixture. I request a systematic comparison with a multi-LL exact diagonalization for N=6 at the same flux (or a fixed-phase DMC / multi-LL DMRG reference), and a decomposition of the one-body density matrix into individual Landau-level occupations, not just their sum.","section":"Abstract; Eqs. (3)-(5); Fig. 3b; Table S3"},{"comment":"The trial state is not constrained to the L=0 sector, although the Hamiltonian (1) is rotationally invariant and the FQH ground states studied are angular-momentum singlets. A finite-size variational ansatz can lower energy by spontaneously breaking rotational symmetry, and the scalar NLLL and overlap diagnostics do not rule this out. The paper never reports the expectation value of L^2 for the ground-state calculations. I ask that the expectation value of L^2 (or an angular-momentum decomposition) be reported for the states underlying Figs. 2-4, and ideally that the ground-state optimization be performed with a projection to L=0, so that the energy gain over LLL ED can be attributed to physical Landau level mixing rather than symmetry breaking.","section":"Eqs. (3)-(5); Figs. 2-3"},{"comment":"The transport-gap prediction in Fig. 4c and the conclusion that the gap persists at kappa=10 rest on quasiparticle/quasihole energies that are not validated. Excitation states are selected only by adding a soft Lz penalty with strength beta (Table S1), and no convergence with respect to beta or any check of the L^2 content of the excited states is provided. At small kappa the neural-network gap happens to lie close to the LLL ED value, but LLL ED is not a multi-LL reference, so this does not validate the kappa dependence. Please benchmark the excitation energies against a multi-LL reference at N=6 (or against leading-order LLM perturbation theory), and show that the reported gap is stable with respect to beta and to the choice of target Lz.","section":"Fig. 4c; Table S3; Eq. (A.16)"},{"comment":"The handling of local-energy outliers is described only qualitatively. The manuscript states that outliers are rare and do not influence training, but no specification is given of how the final energy estimates in Tables S2-S3 are computed, for example whether any clipping or rejection is applied to the local-energy samples used for the reported means and error bars. Since the appendix itself explains that the local energy near the poles can diverge for a generic neural-network phase, this should be quantified: please report the histogram of local energies and the sensitivity of the final energies to outlier removal.","section":"Appendix: Computational details; Fig. S1"}],"minor_comments":[{"comment":"The unit notation is inconsistent: Fig. 3a uses 'E_c/N hbar omega_c kappa' while Table S3 uses 'E_c/N kappa hbar omega_c'; please unify the notation throughout.","section":"Fig. 3a vs Table S3"},{"comment":"The definition of the one-body reduced density matrix in Eq. (A.19) should state explicitly that the monopole harmonics are normalized and that the spinor measure on the sphere is included in the integrals, so that the trace of the 1-RDM equals N.","section":"Eq. (A.19)-(A.20)"},{"comment":"No code or data are released, which makes it difficult to reproduce the variational energies, overlaps, and gap values reported in Tables S2-S3 and Figs. 2-4; providing the optimized wavefunctions or the training code would strengthen the manuscript.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of cond-mat.str-el and the variational approach is interesting. My main reservation is the missing multi-Landau-level validation: without a benchmark or a clear, narrow reframing of the claims as 'variational energies below LLL ED' rather than 'accurately captures Landau level mixing', the central physics claim remains unsupported. I would not object to the authors citing the concurrent related work [59], but it should not replace a direct numerical reference in this paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read. The genuinely new piece is numerical: applying the Psiformer real-space VMC to FQH on the sphere with monopole harmonics, and reporting variational energies below LLL ED for N up to 12 at ν=1/3 and 2/5, plus LLM diagnostics (NLLL, Laughlin overlap, quasiparticle/quasihole densities, transport gap as a function of κ). That is a real step: the energies are valid variational upper bounds, and the comparison with ED is internally consistent because both use the same finite-size corrections. I believe the energy drops below ED are real, not artifacts.\n\nThe soft spot is what those energy drops mean. The abstract says the network 'naturally captures the extent of LLM up to a very high level.' What the calculation shows is that the ansatz can go below the LLL-ED energy. Since the Hamiltonian includes the kinetic term, a variational state not restricted to the LLL will always gain energy; the question is whether the higher-LL admixture is the physical one. The paper offers no multi-Landau-level benchmark — no two-LL ED, no DMRG with multiple Landau levels, no fixed-phase DMC, no analytic limit — and the only LLM diagnostic is NLLL, a single scalar that cannot distinguish a correct LLM-corrected liquid from a state that lowers energy by symmetry breaking (e.g., incipient Wigner crystallization). The transport gap in Fig. 4c and the phase-transition hints at large κ rest on that unvalidated assumption. The appendix gives a correct 1-RDM estimator for NLLL but it isn't benchmarked either. No code or data are released, so I can't check the pipeline.\n\nThat said, this is not a case of a wrong result; it's a case of the central claim outrunning the evidence. The paper is candid about system-size limitations and says the phase boundary is not definitive. The citation pattern is clean: the concurrent disk-geometry study is cited, and the self-citation to prior transport-gap work is not used to set parameters, so there's no circularity. I'd send it to review — a referee could push for a multi-LL benchmark and for code/data — but I wouldn't take the abstract's LLM claim at face value until that's done.","headline":"Real variational numbers, but the LLM claim outruns the evidence: no multi-LL benchmark to show the energy gain is physical.","tokens_in":13607,"tokens_out":2794,"would_cite":true,"duration_ms":22960,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["73.43.-f","71.10.Pm","02.70.Ss"],"model":"deepseek-v4-flash","headline":"A real-space neural network wavefunction lowers the variational energy of fractional quantum Hall states below lowest-Landau-level exact diagonalization by naturally including Landau level mixing.","keywords":["fractional quantum Hall effect","Landau level mixing","neural network wavefunction","variational Monte Carlo","spherical geometry","Laughlin state","composite fermion","transport gap"],"falsifier":"For a small system (e.g., N=4 or N=6 electrons at ν=1/3 on the sphere, same 2Q as in the paper), perform exact diagonalization in a Hilbert space that explicitly includes the lowest two or three Landau levels at κ=1 and compare the ground-state energy and the lowest-Landau-level occupation with the neural network result. If the neural network energy is not equal to or lower than the multi-Landau-level ED energy within statistical error, or if its inferred N_LLL disagrees with the ED value, the claim that the ansatz naturally captures Landau level mixing would be contradicted.","tokens_in":12580,"feed_emoji":"🧠","tokens_out":7665,"duration_ms":57945,"temperature":0.7,"pith_summary":"The paper claims that a real-space neural network wavefunction, built from a single determinant of permutation-equivariant multi-electron orbitals multiplied by a Jastrow factor, can represent fractional quantum Hall ground states with Landau level mixing included. For both the ν=1/3 and ν=2/5 fillings on a sphere, the variational energies are consistently lower than exact diagonalization restricted to the lowest Landau level, for systems up to N=12 electrons at ν=1/3 and N=8 at ν=2/5. The energy gap between the neural network result and the lowest-Landau-level exact diagonalization grows monotonically with the Landau level mixing parameter κ, and the network also yields observables such as pair correlation functions, quasiparticle/quasihole densities, and transport gaps that show the expected trends with mixing. A sympathetic reader would care because Landau level mixing is notoriously difficult to treat with traditional methods, and if the neural ansatz truly captures it, the approach opens a practical route to studying fractional quantum Hall physics in the experimentally relevant regime where the Coulomb interaction is comparable to the cyclotron energy.","feed_headline":"Neural wavefunctions beat Landau-level-limited quantum Hall energies","feed_subtitle":"On 1/3 and 2/5 fillings, neural wavefunctions beat lowest-Landau-level exact diagonalization and capture mixing.","key_machinery":"The load-bearing object is the single-determinant neural network wavefunction ψ_T = e^J det[φ_i(r_j; {r≠j})], where the Jastrow factor e^J satisfies the electron-electron Coulomb cusp and the multi-electron orbitals φ_i(r_j; {r≠j}) = Σ_{k,m} w_{ikm} f_k(r_j; {r≠j}) $u_j^{{Q+m}}$ $v_j^{{Q-m}}$ combine LLL monopole harmonics (defined through spinor coordinates on the sphere, which handle the Dirac-string phase) with permutation-equivariant features f_k produced by the Psiformer, a self-attention transformer architecture. The neural network features encode many-body correlations, and because they are not projected onto the LLL, the product with monopole harmonics yields components in higher Landau levels, giving the ansatz a natural mechanism for Landau level mixing. The determinant enforces antisymmetry while the Jastrow factor controls short-range behavior; variational Monte Carlo with the Kronecker-Factored Approximate Curvature optimizer minimizes the energy. This construction avoids the local-energy divergence at the poles that would arise from a bare neural-network wavefunction on the sphere.","core_discovery":"The central discovery is that the neural network variational ansatz of Eqs. (3)-(5) can go beyond the lowest Landau level approximation without explicit multi-level construction. The wavefunction ψ_T = e^J det[φ_i(r_j; {r≠j})] uses LLL monopole harmonics multiplied by neural-network features, so the orbitals are not restricted to the LLL and the determinant-plus-Jastrow form can represent higher Landau level admixtures. On the sphere, for ν=1/3 with N=6, 8, 10, 12 electrons and for ν=2/5 with N=8 at κ=1, the optimized energies fall below exact diagonalization results that only include the LLL, and the energy difference increases with κ. The fraction of electrons in the LLL, N_LLL/N, decreases from about 0.999 at κ=0.5 to about 0.96 at κ=10, while the overlap with the Laughlin wavefunction drops from about 0.998 to 0.873, showing that the network is capturing the progressive Landau level mixing. The same ansatz produces quasiparticle and quasihole charge densities, pair correlation functions, and a transport gap that shrinks with κ but stays open up to κ=10, consistent with a gapped FQH phase at moderate mixing.","pith_inferences":["The central claim could be tested against a multi-Landau-level exact diagonalization or a fixed-phase diffusion Monte Carlo calculation for the same small systems; if the neural network energy matches or lies below such a reference, the physical origin of the energy drop would be confirmed, whereas a systematic disagreement would indicate variational bias rather than true Landau level mixing.","Since the ansatz uses a single Slater determinant (rather than a sum of determinants), the expressivity is carried almost entirely by the neural-network orbitals; one could probe this by checking whether the energy converges monotonically as network width and depth grow, and whether the N_LLL value from the network matches perturbative estimates at small κ.","The observed N_LLL ≈ 0.96 even at κ=10 implies that only a few percent of electrons leave the LLL, yet this small fraction shifts energies and observables noticeably; a natural extension is to map N_LLL as a function of κ for larger N to see whether the higher-LL occupation saturates or grows further.","The methodology may transfer to other geometries such as the disk or torus, where the monopole-harmonic trick would need to be replaced by plane-wave or other bases; the concurrent disk-geometry study cited by the paper suggests the approach is geometry-flexible."],"forward_implications":["The neural network method can reach system sizes (e.g., N=12 at ν=1/3) that are inaccessible to LLL exact diagonalization, while still including Landau level mixing, so it can extend FQH phase-diagram studies to larger systems and stronger mixing.","Because the energy gap between the neural network result and LLL ED grows with κ, the method provides a quantitative handle on how much Landau level mixing affects ground-state energies and observables in the experimentally relevant regime.","The computed transport gap decreases with κ but remains open up to κ=10, suggesting that the ν=1/3 FQH state survives moderate Landau level mixing, in agreement with earlier studies that the paper cites.","The same ansatz works for the more complex ν=2/5 composite-fermion filling, indicating that the approach may generalize to other Jain states and potentially to non-Abelian FQH states with a modified pairing factor.","The pair correlation function develops a peak at θ=π at large κ, hinting at a tendency toward Wigner crystallization, though the paper explicitly notes the system size is too small to pin down the phase boundary."],"supporting_citations":[{"why":"Provides the Laughlin wavefunction, the reference state for the 1/3 filling and the overlap measurement.","marker":"[2]"},{"why":"Documents the limitation that exact diagonalization typically treats only the lowest Landau level, motivating the need for a method that captures higher levels.","marker":"[9]"},{"why":"Supplies the Psiformer neural network architecture whose permutation-equivariant features form the multi-electron orbitals in the ansatz.","marker":"[24]"},{"why":"Gives the Kato cusp condition that the Jastrow factor is designed to satisfy.","marker":"[38]"},{"why":"Defines the monopole harmonics used as the angular basis on the sphere, which handle the Dirac-string phases and prevent local-energy divergences.","marker":"[39]"},{"why":"Introduces the composite-fermion wavefunction that describes the 2/5 filling, used as the benchmark for that state.","marker":"[41]"},{"why":"Provides earlier transport-gap calculations with Landau level mixing that the paper compares its gap results against.","marker":"[50]"},{"why":"The exact diagonalization software used to produce the LLL ED energies that the neural network results are compared with.","marker":"[60]"}],"fun_headline_variants":["Neural wavefunctions beat Landau-limited quantum Hall energies","Deep learning tames Landau level mixing in fractional quantum Hall","AI wavefunctions capture Landau mixing in FQH states","Neural nets go beyond lowest Landau level in quantum Hall","Beyond LLL: neural wavefunctions win in fractional quantum Hall"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The variational manifold spanned by the neural network wavefunction is expressive enough that the energy lowering below the lowest-Landau-level exact diagonalization comes from physical Landau level mixing, rather than from generic variational freedom or a systematic bias in the ansatz.","fun_headline_variants_meta":{"raw":{"variants":["Neural wavefunctions beat Landau-limited quantum Hall energies","Deep learning tames Landau level mixing in fractional quantum Hall","AI wavefunctions capture Landau mixing in FQH states","Neural nets go beyond lowest Landau level in quantum Hall","Beyond LLL: neural wavefunctions win in fractional quantum Hall"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001079,"raw_usage":{"total_tokens":4550,"prompt_tokens":1018,"completion_tokens":3532,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":634,"completion_tokens_details":{"reasoning_tokens":3447}},"tokens_in":634,"tokens_out":3532,"duration_ms":23915,"temperature":1.0,"reasoning_tokens":3447,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:54:36.126816+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a small system (e.g., N=4 or N=6 electrons at ν=1/3 on the sphere, same 2Q as in the paper), perform exact diagonalization in a Hilbert space that explicitly includes the lowest two or three Landau levels at κ=1 and compare the ground-state energy and the lowest-Landau-level occupation with the neural network result. If the neural network energy is not equal to or lower than the multi-Landau-level ED energy within statistical error, or if its inferred N_LLL disagrees with the ED value, the claim that the ansatz naturally captures Landau level mixing would be contradicted.","supporting_citations":[{"cited_title":"Regnault, J","cited_arxiv_id":null,"evidence_quote":"Documents the limitation that exact diagonalization typically treats only the lowest Landau level, motivating the need for a method that captures higher levels."},{"cited_title":"von Glehn, J","cited_arxiv_id":null,"evidence_quote":"Supplies the Psiformer neural network architecture whose permutation-equivariant features form the multi-electron orbitals in the ansatz."},{"cited_title":"Kato, Communications on Pure and Applied Mathe- matics 10, 151 (1957)","cited_arxiv_id":null,"evidence_quote":"Gives the Kato cusp condition that the Jastrow factor is designed to satisfy."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the monopole harmonics used as the angular basis on the sphere, which handle the Dirac-string phases and prevent local-energy divergences."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides earlier transport-gap calculations with Landau level mixing that the paper compares its gap results against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The exact diagonalization software used to produce the LLL ED energies that the neural network results are compared with."}],"review_version":1}