{"id":"172f819e-fe4b-4b2c-910d-51bdf9dd2f41","arxiv_id":"2412.00618","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A self-attention fermionic neural network variationally solves the disk-geometry fractional quantum Hall problem including Landau level mixing, outperforming LLL-projected exact diagonalization and revealing short-distance wavefunction structure.","lead":"This paper uses a neural network wavefunction to solve the fractional quantum Hall problem in a magnetic field without restricting electrons to the lowest Landau level. It reports lower energies than standard lowest-Landau-level exact diagonalization and finds a transition to a crystal-like state at strong Landau level mixing.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Short-distance 'beyond Laughlin' features are pre-encoded in the Jastrow factor (Eq. 3, β=1/4), so the claim that the unbiased NN reveals them is unsupported.","rationale":"The paper's central claim has two main pillars: (1) an unbiased, physics-agnostic NN-VMC solver that reaches beyond LLL-projected ED, and (2) the discovery of short-distance wavefunction features beyond the Laughlin ansatz, plus a possible LL-mixing-driven phase transition. The energy comparison (Fig. 3) is a genuine methodological achievement, and the paper is appropriately cautious about the phase-transition interpretation in Section VII. However, the 'beyond Laughlin' discovery is not securely supported because the variational ansatz itself encodes a Coulomb cusp. Eq. (3) fixes a Jastrow factor whose form is chosen exactly to impose the Kato cusp, and β is initialized at 1/4, a value explicitly selected to interpolate between the l=1 and l=3 cusp parameters. The SM confirms this prior knowledge is built in. Thus, the observed zero-splitting and cusp are consequences of the ansatz, not emergent network behavior. This does not invalidate the method's usefulness, but it undercuts the paper's strongest novelty claim. The reader's weakest_assumption identified exactly this issue, along with the phase-transition finite-size concern; I focused on the Jastrow issue because it directly affects the headline claim of unbiased discovery, while the phase-transition concern is already acknowledged by the authors as inconclusive. A concrete ablation test could settle the matter. Since the reader already recommended a conditional acceptance with revisions (including clarifying what is imposed versus learned), my analysis does not change that verdict; it reinforces the need for those revisions.","tokens_in":18488,"tokens_out":4774,"duration_ms":42649,"concrete_test":"Ablate the Jastrow factor: retrain the identical Psiformer at λ=1/3, N=12 with Jθ removed (or β fixed to 0), and also with β initialized to 1/7 instead of 1/4, then re-measure ∂log|Ψ|/∂log r and the zero pattern of Fig. 4. If the Coulomb cusp and trimer zero-splitting disappear or change qualitatively, the features are imposed by the Jastrow ansatz, not learned by the network.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV and Eq. (3) introduce a Jastrow factor Jθ(r)=Σ_{i<j} -β α²/(α+|ri-rj|) with β deliberately initialized to 1/4, between the l=1 (1/3) and l=3 (1/7) cusp values. This functional form is the standard two-dimensional Coulomb cusp; the cusp condition (Eq. 4/S14) fixes the linear term, and the SM explicitly states that the cusp information learned from FermiNet is \"put into\" the Psiformer ansatz. Consequently, the reported short-distance structure—Ψ~r as r→0, the splitting of each third-order zero into a trimer of first-order zeros (Fig. 4), and the phase winding l=1 at small radius—is directly produced by this Jastrow term combined with the determinant's Laughlin-like polynomial, as in Eq. (9). The network does not need to discover this; it only needs to optimize a few variational parameters. The paper's statement that \"no information about FQH physics is put in by hand during NN initialization and training\" is contradicted by the construction. The energy benchmarks are credible, but the \"beyond Laughlin\" discovery claim is not established: it is an artifact of the chosen ansatz. The phase-transition claim is separately hedged by the authors (finite-size disk, circular symmetry) and would also require scaling analysis, but the Jastrow pre-encoding is the more fundamental threat to the central claim of an unbiased solver.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper adapts the Psiformer self-attention fermionic neural network to the N-electron two-dimensional Coulomb gas in a magnetic field, working in the full Hilbert space of a disk without Landau-level (LL) truncation. The variational wavefunction is a sum of determinants of complex, permutation-equivariant generalized orbitals with isotropic Gaussian envelopes, multiplied by the real Jastrow factor exp(−β α²/(α+|ri−rj|)). The authors benchmark against LLL-projected exact diagonalization (ED) at fixed angular momentum M for N=6–10 at LL-mixing parameter λ=1/3, find slightly lower FNN energies at every N, and monitor convergence through the measured total angular momentum (converging to 197.97 for N=12, close to the Laughlin value M=198). Fixing N−1 electrons, they visualize the phase and magnitude of the learned 12-electron wavefunction and report a linear trimer of first-order zeros around each electron (splitting of the Laughlin third-order zero), Ψ∼r at short distance with a crossover to Ψ∼r³ near ℓB, and total phase winding 3. For λ=9 they find an angular-momentum drop (M=108→100 for N=9, 135→127 for N=10) and a radial density with one central and N−1 outer electrons, which they interpret, with explicit caveats, as a transition from the FQH droplet toward a crystal-like state (rotating Wigner molecule, Wigner crystal, or Hall crystal).","tokens_in":18858,"tokens_out":19509,"duration_ms":189877,"significance":"The paper's core methodological claim—that a single attention-based FNN can variationally solve the continuum FQH problem without LL projection, at N up to 12, and capture the l=1/l=3 short-distance crossover—is significant for the NN-VMC field if the benchmarks hold. Strengths to credit: the angular-momentum expectation value used as a convergence diagnostic is thoughtful; the SM derives the two-electron cusp condition analytically (Eq. S14); the confining-potential implementation is checked at the 0.1% level; and the hyperparameters and optimization details are reported transparently. The short-distance findings are, however, less novel than presented: the zero-splitting physics was predicted in the cited literature (Refs. [40–44]), and the cusp coefficient is an explicit input to the ansatz rather than a purely learned output. The phase-transition claim is a finite-size observation with the authors' own caveats. The demonstration is nevertheless credible and valuable: it provides the first full-Hilbert-space NN-VMC treatment of this canonical problem and opens a route to strong-LL-mixing regimes inaccessible to LLL-projected ED.","major_comments":[{"comment":"The claim in Sec. II that 'no information about FQH physics is put in by hand during NN initialization and training' is not consistent with the construction described in Sec. IV and the SM. The Jastrow factor in Eq. (3) is admitted to be motivated by the two-electron Coulomb cusp of Eq. (4), and β is initialized at 1/4, the midpoint of the window [1/7, 1/3] that presupposes the l=1/l=3 mixture whose 'discovery' is announced in Sec. VI; the SM states that 'the cusp information learned from FermiNet can then be put into our more expressive Psiformer ansatz.' I note a mitigating detail: the Jastrow is real and positive, so the zero-splitting and phase winding in Fig. 4 are produced by the learned determinants rather than literally inserted, and the optimized l=1-to-l=3 amplitude ratio and the trimer scale ξ are genuine variational outputs. Nevertheless, the qualitative short-distance content (Ψ∼r at r→0, winding l=1, crossover to winding 3 at ∼ℓB) is the input hypothesis rather than the network's discovery, and the same zero-splitting physics was already predicted and studied in Refs. [40–44]. The manuscript should (i) explicitly enumerate which elements are encoded inputs and which are learned; (ii) add a control (e.g., β initialized at 1/3 and 1/7, or the Jastrow omitted) showing that the converged energy and the short-distance structure are stable; and (iii) temper the 'no information... by hand' phrasing.","section":"Sec. II, Sec. IV, Eq. (3); SM 'Jastrow Factor and Cusp Conditions'"},{"comment":"The central benchmark claim that the FNN 'consistently attains energies lower than LL-projected ED' (Secs. VI–VII, Figs. 3(a) and 5(a)) is presented without any statistical error bars on the VMC energies, without the variance of the local energy that Sec. V itself identifies as the convergence criterion, and without a numeric table of energies, so at λ=1/3, where the advantage is described as 'slightly lower,' the reader cannot tell whether the difference is meaningful or within Monte Carlo noise. In addition, the SM section on ED specifies only the LLL basis and M=ML constraint; it does not state whether the disk confining potential Vc and the background term Vb are included in the ED Hamiltonian, which must be identical to the FNN Hamiltonian for the comparison to be valid. At large λ, 'outperforms ED significantly' is the expected behavior of any full-Hilbert-space method against a LLL-projected one, so an independent reference (multi-LL ED at N_LL=2–3, or fixed-phase DMC with the best trial state) is needed to support the accuracy claim in the strong-mixing regime. Please provide a table of energies with statistical errors and sample counts, clarify the ED Hamiltonian, and add an independent cross-check at λ≥3.","section":"Sec. VI Fig. 3(a); Sec. VII Fig. 5(a); SM 'Exact Diagonalization Calculation'"},{"comment":"The abstract's claim that 'a phase transition from FQH liquid to a crystal state is found at strong LL mixing' is materially stronger than what Sec. VII establishes, and the section itself concedes: 'it is difficult to draw a definite conclusion from the above NN-VMC studies' on a finite disk. The evidence consists of angular-momentum drops (M=108→100 for N=9; 135→127 for N=10) at λ=9 and a radial density with a single central electron; the asserted consistency across 7–11 electrons is not shown, there is no scaling of the crossover with N, no explicit comparison of converged FNN energies in the M=100 and M=108 sectors to show which sector wins, and no (rotating-frame) two-body correlation or order parameter that would discriminate a Wigner molecule, a Wigner crystal, or a Hall crystal. The 'unbiased solver' claim makes the absence of any independent cross-check at λ=9, where LLL-projected ED is no longer a useful reference, a further correctness risk. I recommend either softening the abstract and conclusion to a 'signature consistent with a crystal-like state on the disk,' or adding the scaling and order-parameter analysis required to justify the phase-transition claim.","section":"Sec. VII, Fig. 5(b)-(c); Abstract"}],"minor_comments":[{"comment":"The word 'ciruclarly' appears twice in Sec. VI and should read 'circularly'; the affiliation line 'N¨ othnitzer Straße' on the first page also renders with a misplaced diacritic.","section":"Sec. VI"},{"comment":"The running header title, 'Solving the fractional quantum Hall problem with self-attention neural network,' differs from the arXiv title 'Solving and visualizing fractional quantum Hall wavefunctions with neural network'; the submitted title should be unified.","section":"Title page"},{"comment":"The angular-momentum monitor is a genuine strength, but the reported value 197.97 for N=12 is given without an error bar; report the statistical uncertainty, since the convergence interpretation rests on proximity to the integer 198.","section":"Sec. VI"},{"comment":"The expression 1/ϵ(2|l|+1) in Eq. (4) and SM Eq. (S14) is ambiguous between [ϵ(2|l|+1)]^{-1} and (1/ϵ)(2|l|+1); add parentheses or state the intended reading in words.","section":"Sec. IV and SM Eq. (S14)"},{"comment":"The trimer-of-zeros structure is presented from a single 'typical Monte Carlo configuration' (the most probable of 5000 samples); state how the zero positions ξ and the winding structure vary across independently sampled configurations.","section":"Fig. 4 and surrounding text"},{"comment":"The SM says the cusp parameter was learned by an auxiliary FermiNet calculation and then transferred to Psiformer; this is an important methodological statement and should appear in the main text near Sec. IV, together with a note that β=1/4 is an effective cusp for the intermediate-distance window rather than the exact r→0 cusp 1/3 mentioned in the same section.","section":"Sec. IV and SM"},{"comment":"No data-availability or code-release statement is given; for a method built on public repositories (FermiNet/JAX), releasing the evaluation and training scripts would aid reproducibility.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":"The manuscript would benefit from an editorial check of the novelty framing: the short-distance zero-splitting is already present in Refs. [40–43], and the SM's explicit statement that cusp information was transferred from FermiNet into the Psiformer Jastrow directly contradicts the 'no information by hand' claims in the abstract and Sec. II. The phase-transition claim in the abstract should be aligned with the body's own caveats before acceptance. I found the computational work itself careful (convergence monitoring via angular momentum, error-bounded potential implementation, hyperparameter transparency), so the recommended path is substantive revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the LL-mixing energy results are the real content here, and they're worth taking seriously. The short-distance 'beyond Laughlin' story is not; it's baked into the Jastrow factor. The paper is a solid method demonstration with an overstated unbiased-discovery narrative.\n\nWhat's genuinely new: real-space NN-VMC for FQH without LL projection, reaching N=12 and beating LLL-projected ED at fixed M for λ=1/3, with larger gains at λ=3 and 9. The angular momentum convergence check is a nice idea, and the authors are transparent about finite-size caveats for the λ=9 transition. The cusp-condition derivation in the SM is careful and useful.\n\nWhere it falls down: Eq. (3) and the SM state plainly that β is initialized to 1/4, between the l=1 and l=3 cusp values, and that cusp information from FermiNet is 'put into' the Psiformer ansatz. So the observed Ψ~r at short distance and the splitting of the third-order zero into three first-order zeros are exactly what this ansatz with a Coulomb cusp and a Laughlin-like determinant produces (their Eq. 9). Calling this a network 'discovery' contradicts their own claim that no FQH physics is put in by hand. The paper would be more honest if it said the Jastrow imposes the cusp and the network optimizes the crossover distance. Also, there are no statistical errors on the energy comparisons, and the ED benchmark is only in the LLL at one M, so 'consistently attains lower energies' is credible but not as strong as it sounds. The λ=9 state is presented with appropriate caution, but without a scaling analysis it remains a finite-disk observation.\n\nWho this is for: people working on LL mixing in FQH and on extending NN-VMC to topological phases. They'll get a useful method and benchmark, not a proof of unbiased discovery.\n\nRecommendation: send it to peer review. It deserves a serious referee, but the authors should be pushed to release code/data, add error bars and an independent benchmark (even fixed-phase DMC at λ=1/3), and rewrite the 'unbiased' claims to state exactly what is imposed versus learned.","headline":"LL-mixing energies are a real step forward; the 'unbiased discovery' of short-distance physics is pre-encoded in the Jastrow ansatz.","tokens_in":19346,"tokens_out":2734,"would_cite":true,"duration_ms":25638,"reading_group":"yes","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["73.43.-f","71.15.-m"],"model":"deepseek-v4-flash","headline":"A single self-attention neural network, solving the ν=1/3 Coulomb gas in real space without Landau-level projection, beats projected exact diagonalization in energy and reveals short-distance wavefunction structure plus a strong-mixing…","keywords":["fractional quantum Hall effect","neural network variational Monte Carlo","self-attention wavefunction","Landau level mixing","Coulomb cusp","Laughlin wavefunction","Wigner crystal","exact diagonalization"],"falsifier":"Run the same Psiformer at ν=1/3, λ=1/3 with the Jastrow factor's β fixed to the Laughlin value 1/7 (or with the Jastrow removed entirely) and check whether the optimized wavefunction still develops a β≈1/3 cusp and three split first-order zeros around each electron; if it does not, the short-distance structure was pre-imposed by the ansatz. Separately, repeat the λ=9 calculation on a torus or with a much softer confining potential for N=7,...,20 and check whether the angular-momentum drop and the one-center-plus-edge density survive in the thermodynamic limit.","tokens_in":18275,"feed_emoji":"🧲","tokens_out":6938,"duration_ms":124633,"temperature":0.7,"pith_summary":"The paper asks whether a single, physics-agnostic neural-network wavefunction can variationally solve the fractional quantum Hall problem directly in real space, without projecting onto the lowest Landau level. The authors adapt the self-attention network Psiformer to spin-polarized electrons on a disk at filling ν=1/3 with Coulomb interaction and a neutralizing background. They find that the network consistently reaches energies below lowest-Landau-level projected exact diagonalization, spontaneously learns circular symmetry and quantized angular momentum, and produces a wavefunction with short-distance Coulomb cusp and zero-splitting structure absent from the Laughlin ansatz. At strong Landau-level mixing the network finds a transition from the fractional quantum Hall droplet to a crystal-like state. If correct, this makes neural-network variational Monte Carlo a universal, unbiased solver for strongly correlated and topological states where Landau-level projection fails.","feed_headline":"Neural-network wavefunction finds lower FQH energies than projected ED","feed_subtitle":"Unbiased real-space solver captures Coulomb cusps and a liquid-to-crystal transition at strong LL mixing.","key_machinery":"The central object is Psiformer, a self-attention fermionic neural network whose permutation-equivariant complex generalized orbitals are built from one-electron features via attention layers, aggregated into a sum of Slater determinants, and multiplied by a Jastrow factor J = ∑ −β α²/(α+|ri−rj|). The Jastrow factor carries the two-electron Coulomb cusp: the short-distance wavefunction behaves as r^l[1 + r/(ϵ(2l+1))], and the initial β is set to 1/4, between the l=1 value 1/3 and the l=3 value 1/7, so the network can interpolate between generic Coulomb behavior and Laughlin-like correlation. Total angular momentum is measured during training as a convergence diagnostic, since low-lying states differ in angular momentum sector. This machinery lets the network represent unprojected wavefunctions with infinitely many Landau levels and allows direct visualization of phase and magnitude structure near electron collisions.","core_discovery":"The central claim is that a self-attention fermionic neural network, with no Landau-level truncation and no FQH-specific design, variationally solves the ν=1/3 Coulomb electron gas in a magnetic field more accurately than LLL-projected exact diagonalization across a wide range of the Landau-level mixing parameter λ. At λ=1/3 the network's ground-state energy is lower than ED for 6–10 electrons, and the gap widens at larger λ; kinetic energy above ωc/2 directly evidences occupation of higher Landau levels. Microscopically, the learned wavefunction shows a Coulomb cusp as two electrons approach and, instead of the Laughlin m=3 zero, three first-order zeros whose combined phase winding around one electron is still 3, consistent with mixing of l=1 and l=3 relative angular momentum channels. At λ=9 the ground-state angular momentum drops from M=108 to M=100 and the charge density develops a one-electron-at-center, eight-on-the-edge shell structure, which the authors interpret as a transition from the FQH liquid to a rotating Wigner-molecule or crystal state on the disk.","pith_inferences":["A clean test of whether the short-distance structure is genuinely learned would be to fix the Jastrow β at the Laughlin value 1/7 or remove the Jastrow entirely and check whether the optimized wavefunction still develops the l=1 cusp and the split first-order zeros.","The λ=9 transition could be an artifact of the disk's confining potential stabilizing the one-center-plus-edge shell; repeating the calculation on a torus or with a much softer boundary would show whether the angular-momentum drop and density profile persist in the thermodynamic limit.","The same unprojected approach should extend naturally to even-denominator and non-Abelian states such as ν=5/2, where Landau-level mixing is thought to be crucial, and to moiré fractional Chern insulators, where the angular-momentum diagnostic would need replacement by momentum or entanglement diagnostics.","The paper notes FermiNet struggles to reach the correct angular momentum in this problem; a controlled comparison with FermiNet at identical Jastrow factor and system size would isolate whether self-attention's all-to-all equivariant weighting is the key architectural advantage."],"forward_implications":["The same Psiformer architecture solves the FQH droplet and the strong-mixing crystal state from one unbiased ansatz, without separate trial wavefunctions for each phase.","LL-projected ED energies are variational upper bounds that miss Landau-level mixing; at moderate to strong λ the FNN energies are lower and kinetic energy exceeds ωc/2, so unprojected treatment changes quantitative energetics.","Short-distance FQH pair correlations are not the Laughlin r^m form: the Coulomb cusp requires mixing of infinitely many Landau levels, and the m=3 zero splits into three first-order zeros with total phase winding 3.","At λ=9 the ground state on the disk changes from the FQH droplet (angular momentum M=108) to a crystal-like state (M=100) with one electron in the center and eight on the edge for N=9, with similar angular-momentum drops for 7–11 electrons.","The method reaches N=12 electrons on a disk without Landau-level truncation, beyond the 10-electron exact-diagonalization limit presented in the paper."],"supporting_citations":[{"why":"Supplies the Psiformer self-attention architecture and the Jastrow-plus-determinants ansatz form used throughout the paper.","marker":"[36]"},{"why":"Defines the Laughlin wavefunction at ν=1/3 whose energy, zero structure, and short-distance behavior are the benchmarks the paper compares against.","marker":"[38]"},{"why":"Introduced the generalized Slater determinant construction for fermionic neural networks that the present ansatz builds on.","marker":"[28]"},{"why":"Provides the full-LLL exact diagonalization in disk geometry whose energies serve as the baseline the FNN must beat.","marker":"[50]"},{"why":"Derives the two-electron Coulomb cusp condition used to justify the Jastrow factor and choose its initial β value.","marker":"[58–61]"},{"why":"Supplies the classical shell configurations used to interpret the λ=9 state as a rotating Wigner molecule.","marker":"[67]"}],"fun_headline_variants":["Neural net finds lower FQH energies than exact diagonalization","AI solver reveals Coulomb cusp in FQH wavefunction","Neural network finds FQH-to-crystal phase transition","Attention-based fermion net beats ED for FQH","Neural solver resolves FQH states beyond Laughlin"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The short-distance 'discovery' rests on assuming the trial wavefunction's built-in two-electron factor—whose initial strength sits between the l=1 and l=3 cusp values—did not already force the zero-splitting and cusp reported as learned, and the transition claim rests on assuming the angular-momentum drop at the strongest mixing is the thermodynamic ground state rather than a finite-disk artifact.","fun_headline_variants_meta":{"raw":{"variants":["Neural net finds lower FQH energies than exact diagonalization","AI solver reveals Coulomb cusp in FQH wavefunction","Neural network finds FQH-to-crystal phase transition","Attention-based fermion net beats ED for FQH","Neural solver resolves FQH states beyond Laughlin"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000929,"raw_usage":{"total_tokens":3977,"prompt_tokens":945,"completion_tokens":3032,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":561,"completion_tokens_details":{"reasoning_tokens":2947}},"tokens_in":561,"tokens_out":3032,"duration_ms":87089,"temperature":1.0,"reasoning_tokens":2947,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:10:23.656268+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same Psiformer at ν=1/3, λ=1/3 with the Jastrow factor's β fixed to the Laughlin value 1/7 (or with the Jastrow removed entirely) and check whether the optimized wavefunction still develops a β≈1/3 cusp and three split first-order zeros around each electron; if it does not, the short-distance structure was pre-imposed by the ansatz. Separately, repeat the λ=9 calculation on a torus or with a much softer confining potential for N=7,...,20 and check whether the angular-momentum drop and the one-center-plus-edge density survive in the thermodynamic limit.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduced the generalized Slater determinant construction for fermionic neural networks that the present ansatz builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the classical shell configurations used to interpret the λ=9 state as a rotating Wigner molecule."}],"review_version":1}