{"id":"49b52fbc-b561-419d-95f5-bb775c3305d8","arxiv_id":"2607.02905","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":5.5,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"Lecture notes define neural-network field theory and survey how it recovers known QFT/string results plus applied AI techniques for string problems.","lead":"These lecture notes teach neural networks to string theorists through a field-theory lens, define field theories via network architectures plus parameter densities (NN-FT), and survey results that recover Liouville structure constants, string amplitudes, the BKT transition, and a mode-counting derivation of D=26. They also cover agentic workflows, PINNs for Calabi-Yau metrics, RL on landscapes and knots, and conjecture generation.","discovery_kind":"review","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the OS-axioms caveat already flagged by the reader.","rationale":"The reader correctly identifies both the strongest claim (definitional NN-FT plus concrete recoveries) and the weakest assumption (OS axioms beyond free/QM cases). Because the notes are an exposition that surveys already-arXived results rather than a single new theorem, the OS gap is a known open problem rather than a hidden flaw that collapses the central claim. Analytic recoveries (string amplitudes, D=26 mode count) and numerical checks (DOZZ, BKT) are transparent and match external formulas; no circularity or internal contradiction is present. Reproducibility is moderate but adequate for lecture notes. Consequently the ACCEPT / HIGH-confidence verdict needs no adjustment. The concrete test above simply re-verifies the most distinctive analytic claim (critical dimension) under an independent zeta regularization, which is a useful sanity check rather than a threat to the argument.","tokens_in":47784,"tokens_out":546,"duration_ms":5465,"concrete_test":"Independently recompute the regularized Jacobian mode counts of §3.2.5 (Eqs. 193–194) for the free boson and bc ghosts on S^{2}, confirming that ζ_X(0) = −2/3 and ζ_bc(0) = −5/3 still yield B_reg = −(D−26)/3; if the arithmetic fails under a different zeta prescription the D=26 claim weakens, otherwise the analytic recovery stands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The manuscript is lecture notes whose central claim is definitional and synthetic: a field theory may be specified by (architecture, parameter density), free theories arise by CLT + spectrum shaping, interactions by CLT violation, and the cited constructions recover known results (DOZZ, Virasoro-Shapiro/Veneziano, BKT, Ward/anomaly identities, D=26 mode count). Those recoveries are either analytic (parameter-space Gaussian integrals, zeta-regularized Jacobians) or numerical against external formulas, and the notes do not claim a new theorem beyond the already-posted papers. The only structural gap that could undermine the “QFT” reading is precisely the one the reader already isolates: reflection positivity and cluster decomposition are verified only for free cases and quantum mechanics (§3.1.3), so the Euclidean ensembles are not yet guaranteed to continue to unitary Lorentzian theories. That gap is openly stated and does not affect the Euclidean constructions or the pedagogical value of the notes. No further load-bearing inconsistency appears in the argument.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"These lecture notes present a three-day course on AI for string theorists. Day 1 develops neural-network essentials (expressivity, statistics via NNGP and 1/N non-Gaussianities, dynamics via NTK and µP) with a field-theoretic vocabulary. Day 2 defines Neural Network Field Theory (NN-FT) as the data of an architecture ϕ_θ together with a density P(θ), with correlators computed from the parameter-space partition function; free theories arise by CLT plus spectrum shaping, interactions by CLT violation, and recent constructions recover Liouville DOZZ constants, Virasoro-Shapiro/Veneziano amplitudes, BKT phenomenology with vortices, Ward identities/anomalies, and a mode-counting derivation of D=26. Day 3 surveys applied ML for strings (agents, PINNs for Calabi-Yau metrics, RL on vacua and unknotting, conjecture generation). The notes synthesize and pedagogically reorganize results largely already posted as research papers.","tokens_in":48020,"tokens_out":1234,"duration_ms":11141,"significance":"If the NN-FT definition and constructions hold as stated, the notes supply a coherent, field-theoretic language for ensembles of networks and a concrete engineering toolkit (CLT free theories, independence-breaking interactions, absorption for symmetries, discrete latents for topology) that recovers standard results of 2d CFT, string theory, and topological phase transitions. Strengths include analytic recovery of string amplitudes from Gaussian parameter integrals, percent-level numerical DOZZ agreement with reported error bars, an explicit zeta-regularized Jacobian count yielding D=26, and open acknowledgment that OS axioms beyond free/QM cases remain open. As lecture notes they are valuable for training and for consolidating a rapidly growing literature; they do not claim a single new theorem beyond the cited papers.","major_comments":[{"comment":"§3.1.3 (and the OS checklist): the claim that NN ensembles define quantum field theories rests on Osterwalder–Schrader reconstruction. The notes correctly state that reflection positivity and cluster decomposition are verified only for free cases and quantum mechanics, and remain open for the interacting constructions (Liouville, strings, BKT). This is load-bearing for any Lorentzian/unitary reading of Day 2. The Euclidean constructions and amplitude recoveries stand independently, but the manuscript should either (i) restrict the “QFT” language more carefully to Euclidean correlators throughout §§3.2.2–3.2.5, or (ii) supply at least a sketch of how RP/cluster are expected to hold (or fail) for the deformed densities used in Liouville and ϕ^4.","section":null},{"comment":"§3.2.3, Eqs. (156)–(165) and the D=26 count in §3.2.5, Eqs. (189)–(194): the free-boson and ghost architectures are engineered so that the two-point functions match the standard worldsheet propagators (spectrum shaping + Grassmann Gaussians). Once that is fixed, Virasoro-Shapiro/Veneziano and the critical-dimension count follow by standard Gaussian integrals and zeta regularization of mode Jacobians. This is design, not circular derivation of the target numbers, but the notes should state more explicitly that the physical content is the choice of architecture/density that realizes the free theory, after which the amplitudes and anomaly are recovered rather than independently predicted. A short paragraph distinguishing “engineering the free theory” from “deriving the amplitude” would prevent misreading.","section":null}],"minor_comments":[{"comment":"Introduction and §5: the notes are dated July 2026 and cite contemporaneous results (IMO gold, Erdős counterexamples, agentic systems). For archival publication, a brief note on the snapshot nature of the “current AI landscape” paragraphs would help future readers.","section":null},{"comment":"Fig. 3 (Liouville DOZZ): the caption reports L=30, 10×50k runs, and error bars smaller than markers. Adding a short statement of the Monte-Carlo estimator variance or a reference to the companion paper’s numerical appendix would strengthen reproducibility claims.","section":null},{"comment":"§2.4.3 (µP scaling): the one-parameter family and the unique solution with η∼O_N(1) are dense. A small table of (a_ℓ,b_ℓ,c,d) for the standard NTK vs µP regimes would aid readers who skip the index algebra.","section":null},{"comment":"§4.1 (agents): the GPD pipeline and “order-of-magnitude drop in implementation barrier” are useful but anecdotal. A single concrete before/after example (e.g., time to reproduce a Day-2 baseline) would make the claim more falsifiable.","section":null},{"comment":"Typos and polish: the disclaimer already notes residual typos from rapid posting; a pass for consistency of notation (e.g., θ vs θ(t), G^{(n)} vs G_c^{(4)}) and for missing cross-references between Day-1 NTK and Day-3 metric flows would improve readability.","section":null}],"recommendation":"minor_revision","confidential_remarks":"These are lecture notes synthesizing work largely already on arXiv (many coauthored by the lecturer). The journal should confirm that the venue accepts pedagogical/survey notes of this length and that the novelty bar is appropriate for “notes” rather than a research article. The OS-axioms caveat is already stated honestly; I do not see a hidden load-bearing error. Fit for a hep-th or mathematical-physics lecture-notes series is good; for a pure research journal the contribution is more consolidative than transformative."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"These are lecture notes, not a research paper claiming a first result. The punchline is that Halverson has a clean, field-theoretic organization of neural-network theory (expressivity / statistics / dynamics) and then a working definition of NN-FT: a field theory is an architecture plus a density on its parameters. With that definition the notes recover free theories by CLT + spectrum shaping, interactions by CLT violation, and concrete constructions that match DOZZ, Virasoro-Shapiro/Veneziano, BKT phenomenology, Ward identities, and a mode-counting derivation of D=26. All of those recoveries already appear in the cited 2025–26 papers; the notes re-present them for a Pre-Strings audience.\n\nWhat works: the Day-1 scaffolding is clear and usable for hep-th students. The free-theory constructions are transparent. The Liouville Monte-Carlo reports percent-level agreement with DOZZ and error bars. The string amplitudes come out of ordinary Gaussian integrals over parameters; the D=26 count is a zeta-regularized Jacobian mismatch. Circularity is low: once the free two-point function is engineered, the non-trivial checks are against external formulas. The agent-centric framing of Day 3 is practical and honest about verification.\n\nSoft spots are real but proportionate. Reflection positivity and cluster decomposition are verified only for free cases and quantum mechanics; the notes state this openly. Code is not shipped. There is no independent new theorem here. None of that undercuts the Euclidean constructions or the pedagogical purpose.\n\nWho it is for: string theorists and students who want a single readable entry point to the NN-FT program and to current ML-for-strings practice. It deserves a serious referee if submitted as lecture notes or a review; it is not a research article that needs to invent a new claim. I would bring it to reading group for the Day-2 survey and the D=26 derivation, and I would cite the notes when pointing people at the program.","headline":"Solid lecture notes that package the 2025–26 NN-FT papers into a usable three-day course; the value is synthesis and pedagogy, not a new theorem.","tokens_in":48686,"tokens_out":497,"would_cite":true,"duration_ms":6687,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A field theory can be defined by a neural-network architecture and a density on its parameters, recovering free fields, Liouville, string amplitudes, BKT, and D=26.","keywords":["neural network field theory","NNGP","Liouville theory","bosonic string","critical dimension","BKT transition","Ward identities","Calabi-Yau metrics"],"falsifier":"Construct an explicit finite-width NN-FT whose four-point and higher correlators can be computed exactly, check whether the Osterwalder–Schrader reflection-positivity inequalities hold after analytic continuation, and compare the resulting Lorentzian spectrum or S-matrix against a known continuum QFT or string amplitude.","tokens_in":48600,"feed_emoji":"🧠","tokens_out":1031,"duration_ms":8854,"temperature":0.7,"pith_summary":"These lecture notes argue that the data of a neural network—its architecture together with a probability density on its parameters—already define a field theory, because correlators can be computed by integrating over parameters rather than over field configurations. Infinite-width networks become free fields by the central limit theorem; interactions appear when that theorem is broken by finite width or by deforming the density. With this definition the notes reconstruct Liouville structure constants, the Virasoro–Shapiro and Veneziano amplitudes, the BKT transition via discrete topological latents, Ward identities and anomalies on parameter space, and a mode-counting derivation of the bosonic-string critical dimension D=26. They also survey how the same tools reverse: physics-informed networks for Calabi–Yau metrics, reinforcement learning over string vacua and knots, and interpretable models aimed at conjecture generation. A sympathetic reader cares because the construction supplies both a new language for field theory and concrete computational routes to quantities that usually require path integrals or lattice methods.","feed_headline":"Neural nets define field theories, recover D=26","feed_subtitle":"Architecture plus parameter density yields free fields, Liouville, string amplitudes, and BKT","key_machinery":"Neural Network Field Theory (NN-FT): the data (ϕ_θ, P(θ)) whose parameter-space partition function Z[J] = ∫ dθ P(θ) exp(∫ J ϕ_θ) defines all correlators. Free limits follow from the CLT; interactions from finite N or density deformations; symmetries and anomalies from flows on parameter space that leave or break that integral.","core_discovery":"The central claim is that a neural-network field theory is the pair (architecture, parameter density), with the partition function an integral over parameters. From this definition one engineers free theories by the CLT and spectrum shaping, interactions by 1/N corrections or independence-breaking deformations, and quantum theories by clearing Osterwalder–Schrader conditions. Concrete ensembles then recover Liouville DOZZ constants, string amplitudes, BKT phenomenology, Ward identities, and a Jacobian mode-count that yields D=26.","pith_inferences":["If reflection positivity can be engineered systematically, NN-FT becomes a constructive route to interacting QFTs that bypasses the need to write an action first.","The same parameter-space Ward identity machinery that produced D=26 should apply to other anomaly coefficients (central charges, chiral anomalies) once the relevant flows are identified.","Agentic verification loops wrapped around NN-FT correlator codes could turn the Monte-Carlo Liouville and BKT checks into automated, continuously refined numerical proofs.","Architecture-robust free-boson constructions suggest that finite-N non-Gaussian corrections to string amplitudes are themselves architecture-dependent observables worth classifying."],"forward_implications":["Any Euclidean QFT in the constructive sense admits an NN-FT description with countably many parameters (universality).","Liouville three-point structure constants can be obtained by Monte Carlo sampling of a conditioned zero-mode density plus free spherical-harmonic modes.","Virasoro–Shapiro and Veneziano amplitudes arise as exact finite-dimensional integrals over network parameters once free bosons and ghosts are realized as infinite-width ensembles.","Topological sectors enter as discrete latent variables, allowing vortex unbinding and the BKT jump to be read off from network correlators.","The critical dimension D=26 follows from a regularized mismatch between bosonic and Grassmann Jacobians under a Weyl flow on parameter space."],"fun_headline_variants":["Neural nets define field theories recover D=26","NN architecture plus density yields free fields Liouville strings","Field theories from nets recover string amplitudes and D=26","Neural network ensembles engineer Liouville BKT and critical D=26","Architecture-density pairs define NN-FT recovering bosonic D=26"],"cache_read_input_tokens":32896,"weakest_assumption_plain":"That the Euclidean correlators of these engineered ensembles continue to unitary Lorentzian quantum field theories, which requires reflection positivity and cluster decomposition that are verified only in free and quantum-mechanics cases so far.","fun_headline_variants_meta":{"raw":{"variants":["Neural nets define field theories recover D=26","NN architecture plus density yields free fields Liouville strings","Field theories from nets recover string amplitudes and D=26","Neural network ensembles engineer Liouville BKT and critical D=26","Architecture-density pairs define NN-FT recovering bosonic D=26"]},"model":"grok-4.5","effort":"low","cost_usd":0.005136,"raw_usage":{"total_tokens":1420,"prompt_tokens":750,"num_sources_used":0,"completion_tokens":68,"cost_in_usd_ticks":51360000,"prompt_tokens_details":{"text_tokens":750,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":602,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":750,"tokens_out":68,"duration_ms":6539,"temperature":1.0,"reasoning_tokens":602,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T06:13:12.695379+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Construct an explicit finite-width NN-FT whose four-point and higher correlators can be computed exactly, check whether the Osterwalder–Schrader reflection-positivity inequalities hold after analytic continuation, and compare the resulting Lorentzian spectrum or S-matrix against a known continuum QFT or string amplitude.","supporting_citations":[],"review_version":1}