{"id":"4d9c19c5-6436-4f35-bace-90814502bfb9","arxiv_id":"2501.12978","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"The paper sketches a neurosymbolic classifier for Galois groups and presents height-bounded counts of small-degree polynomials, but provides no trained model or accuracy results.","lead":"A short paper proposes using neurosymbolic networks (neural networks combined with symbolic rules) to classify the Galois groups of polynomials. It builds small databases of cubics, quartics, and quintics with computed Galois groups and sketches a hybrid classifier, but reports no experimental results.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central efficiency claim is unverified: Section 5 describes a GaloisNetwork but reports no accuracy, baseline, or training comparison, so 'more efficient than usual neural networks' is not supported by the manuscript.","rationale":"Good-faith reading: the paper is a mix of classical Galois theory review, database construction, and a sketch of a neurosymbolic architecture. The classical sections are mostly standard, and the database counts, if reproducible, would be a useful resource. The problem is that the novel claim, the neurosymbolic classifier's efficiency advantage, is not tested anywhere in the PDF. A reader cannot distinguish 'works relatively well' from an unsupported assertion. I agree with the reader's rejection and also find the enumeration concern real (no code or data snapshot provided), but I would prioritize the missing experiment: it is the one condition that must be true for the abstract's central claim, and the manuscript provides no evidence for it. If a reproduction with the described architecture and baselines were supplied and showed a clear advantage, the claim would land; until then the appropriate verdict is rejection. This is not a claim of misconduct; it is a claim of insufficient evidence and, in the signature layer, a concrete correctness gap.","tokens_in":17594,"tokens_out":5622,"duration_ms":59125,"concrete_test":"Run the described GaloisNetwork on the paper's quartic and quintic databases (or a fresh Sage enumeration of P_h^4 and P_h^5 for h=10/20) with the same 64x3 ReLU MLP and Adam/100-epoch setup, with and without the symbolic post-processing layers, and report test accuracy and training cost on an identical train/validation split. If the symbolic version does not beat the plain MLP by a statistically significant margin, or if the experiment cannot be run because the database and code are unavailable, the abstract's efficiency claim remains unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5 (Neurosymbolic Networks) is the only place the abstract's core claim could be evidenced, but its 'Implementation and efficiency' subsection contains no experimental numbers whatsoever. It states that the authors 'implement this approach and test it for quartics and quintics' and that evaluations 'demonstrate the effectiveness,' yet no accuracy, training curve, confusion matrix, wall-clock time, parameter count, or comparison to a plain feedforward network appears. The architecture described is a standard 64-64-64 ReLU MLP with softmax plus a post-processing rule; the abstract's 'more efficient than usual neural networks' therefore has no measurable meaning in the text. The supporting code in Listing 1 is also not a correct signature layer as written: it never checks that the chosen primes do not divide the discriminant (Dedekind's theorem requires p∤Δ_f), and it records only a set of factor degrees, discarding multiplicity information; thus the symbolic layer's correctness is itself unverified. Even granting the database enumeration counts in Section 4, the central efficiency claim would remain unsupported. Load-bearing because the paper's novelty claim is precisely this efficiency comparison.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a neurosymbolic architecture for classifying Galois groups of irreducible integer polynomials of degrees 3–5, combining a feedforward ReLU network with symbolic layers based on real-root counts, reductions modulo primes, and discriminant checks. It also reports enumeration counts and distributions for height-bounded cubic, quartic, and quintic datasets, including a table of the twenty quintics of height at most 10 with Galois group C5. The theoretical sections review classical criteria for Galois groups of cubics, quartics, and quintics, and Section 5 sketches an implementation called GaloisNetwork. The central efficiency claim—that the neurosymbolic network is more efficient than plain neural networks—is not supported by any reported experiments.","tokens_in":17809,"tokens_out":7467,"duration_ms":82337,"significance":"If the claimed database counts and invariant formulas were fully established, they would provide useful concrete test data for questions around Malle's conjecture and for comparing naive and moduli heights. The explicit enumeration of height-bounded quintics with Galois group C5 in Table 4 is a concrete, potentially checkable contribution. However, the paper's main advertised novelty, the neurosymbolic efficiency gain, has no empirical support, and a load-bearing proof in the solvability criterion is incomplete. The paper does not ship reproducible code or a data snapshot, so the exact counts cannot be independently verified. Credit is due for the explicit invariant formulas and the attempt to combine Dedekind's theorem with learned features, but as written these are classical ingredients assembled into an untested sketch rather than a demonstrated result.","major_comments":[{"comment":"The abstract's central claim that the neurosymbolic network is 'more efficient than usual neural networks' is not supported by any measurement in the paper. Section 5, 'Implementation and efficiency', describes a 64-64-64 ReLU MLP with a softmax output and a post-processing rule, but it reports no accuracy, loss curves, confusion matrices, wall-clock time, parameter counts, or any baseline comparison to a plain neural network. Section 1 is even more explicit, calling the system 'not a fully developed product yet' and saying there 'might not be a compelling reason to use AI models' for degrees 3, 4, and 5; this directly contradicts the definitive efficiency claim in the abstract.","section":"Section 1 / Section 5"},{"comment":"The proof of Lemma 8 infers Delta_g != 0 from Delta_f != 0 via the displayed product formula Delta_g = ((alpha1-alpha2)(alpha3-alpha4)(alpha4-alpha5)(alpha3-alpha5))^4 * Delta_f * I_2^2 * I_3 * I_4^2 * I_6^2, but the nonvanishing of the invariants I_2, I_3, I_4, and I_6 is never established. If any of these invariants vanishes on a polynomial with distinct roots, the product formula does not imply Delta_g != 0, and the lemma is not proven. The formula is introduced as a result of 'computational algebra' without a derivation, an implementation, or a reference, so the reader cannot verify it.","section":"Section 2.3, Lemma 8"},{"comment":"The enumeration counts and the training data depend on the Sage call 'PP.rational_points(h)' followed by an irreducibility filter, but the manuscript provides no code, data snapshot, or commit hash, and the exact normalization, duplicate-removal, and projective-representative conventions are not fully specified. Consequently, claims such as 1,178,856 irreducible cubics, 5,676 non-S4 quartics, and exactly 20 quintics with Galois group C5 cannot be checked. An error in enumeration, scaling, or class labeling would invalidate both the distribution discoveries and every trained example in Section 5.","section":"Section 4, Lemmas 11-13"},{"comment":"The signature layer as implemented in Listing 1 does not check that the chosen primes satisfy p does not divide Delta_f, which is required for the validity of Dedekind's theorem. It also stores only factor degrees as a set, discarding multiplicities and treating repeated factors of the same degree as a single entry. The resulting 'signature' can therefore disagree with the group signatures defined in Section 3.2, so the symbolic layer's correctness is not established.","section":"Section 5, Listing 1"},{"comment":"The manuscript states that Galois labels are computed by an external computer algebra system and that the features are classical invariants (discriminant, real-root count, mod-p factorization). This means the model itself does not discover or define the target; the claimed advantage over a plain network must be demonstrated by a controlled comparison. Since no such comparison appears, the paper's title and abstract overstate what is currently a proposal plus a database, not a validated method.","section":"Section 4.2 / Section 5"}],"minor_comments":[{"comment":"The statement of Lemma 7 contains the Albanian words 'dhe' and 'për' in an otherwise English mathematical text; these should be replaced with 'and' and 'for'.","section":"Section 2.3, Lemma 7"},{"comment":"The phrase 'the ration of the weighted height over the naive height' should read 'the ratio of the weighted height to the naive height' in both occurrences.","section":"Sections 4.2 and 4.3"},{"comment":"Table 5 has duplicate row numbers and several rows without an explicit discriminant value, so the list of forty cubics with Galois group C3 is not usable as a clean data table.","section":"Appendix A.1, Table 5"},{"comment":"Table 1 lists the number of transitive subgroups of S_n for n up to 45 but omits n = 32, which is needed for the claimed range and for checking the completeness of the enumeration.","section":"Section 3.1, Table 1"},{"comment":"The weighted moduli height, the set P_h^n, and the invariants used in the database are defined by reference to [16] and [17] rather than in the paper, making Section 4 not self-contained and the exact height convention ambiguous.","section":"Section 4, definitions"},{"comment":"In Section 4.2, J6 is first defined as Delta(f,x) and then stated to satisfy J6 = (4*J2^3 - J3^2)/27; the relationship between these two definitions should be stated explicitly to avoid confusion.","section":"Section 4.2, invariants"}],"recommendation":"reject","confidential_remarks":"The editor may wish to check whether the linked repository at [18] actually contains the promised data and code, since the load-bearing definitions and many of the claimed counts are hosted on the authors' own RISAT site rather than in the paper. The manuscript reads as a preliminary position report: the core efficiency claim is untested, the proof of Lemma 8 has a gap, and the database is not reproducible from the text. A future submission with real experiments, a repaired proof, and a versioned data release could be worth reconsidering."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the bottom line: the paper's headline claim—that a neurosymbolic network classifies Galois groups more efficiently than a plain neural network—is not supported by anything in the text. Section 5 describes the architecture and says it was implemented and tested, but gives no numbers, no training curves, no baseline comparison, no wall-clock time. That is not a minor omission; it's the entire stated contribution.\n\nWhat the paper does well is the classical background. The cubic and quartic resolvent criteria are correctly presented and properly attributed, and the quintic solvability discussion via the sextic resolvent is accurate as far as it goes. The idea of creating height-bounded databases of irreducible polynomials with their Galois groups is reasonable, and the reported counts (e.g., 20 quintics of height ≤ 10 with C5) would be genuinely useful if the enumeration were reproducible. The tables of transitive subgroups are also handy.\n\nThe soft spots are substantial. First, the enumeration is unverifiable: no code, no data snapshot, no commit hash. The description relies on Sage's rational_points(h) without spelling out the normalization from projective points to integer polynomials, and the tables contain duplicated entries (e.g., Table 5). Second, the proof of Lemma 8 is not valid as written. The displayed formula for Δ_g includes factors I_2, I_3, I_4, I_6 that could vanish even when Δ_f ≠ 0, so Δ_f ≠ 0 does not imply Δ_g ≠ 0. The lemma may be true classically, but this proof doesn't establish it. Third, the Appendix code for the signature layer is not a correct implementation of Dedekind's theorem: it never checks that the primes don't divide the discriminant, and it stores a set of factor degrees, discarding multiplicities that are required.\n\nNone of these issues are fatal to the underlying idea—a neurosymbolic classifier for Galois groups could be useful, especially for degrees ≥ 6. But the paper as written is a survey plus a proposal, not a demonstrated result. The audience that would get value is someone wanting a quick refresher on the classical criteria and a cautionary example of how an ML paper can overclaim. For a serious referee, it's not ready: the main claim is untested and the data is not checkable. My recommendation is to desk reject, with a note that adding real experiments, a reproducible data release, and a fixed proof of Lemma 8 would make a resubmission worth considering.","headline":"The abstract's efficiency claim is never tested: Section 5 describes a GaloisNetwork but reports no accuracy, baselines, or training metrics, so the paper's central novelty is unsubstantiated.","tokens_in":18355,"tokens_out":3236,"would_cite":false,"duration_ms":34158,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["11R32","68T07","12Y05","11R09"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a neurosymbolic network—a feedforward classifier wrapped in symbolic layers for mod-$p$ signatures, discriminants, and real-root counts—classifies Galois groups of low-degree polynomials more efficiently than a plain…","keywords":["Galois theory","Neurosymbolic AI","Neurosymbolic networks","Galois group classification","polynomial databases","bounded height","invariant theory","solvability by radicals"],"falsifier":"Recompute the bounded-height enumeration with an independent implementation: for degree 5, height at most 10, list all irreducible integer quintics with projective height $\\leq 10$, compute their Galois groups by a resolvent-based method, and compare the counts. The paper predicts exactly 20 polynomials with Galois group $C_5$ in three $\\mathbb{Q}$-isomorphism classes, 480 with $F_5$, 900 with $D_5$, and 1,146 with $A_5$; for quartics it predicts 5,676 irreducible polynomials of height $\\leq 10$ not having Galois group $S_4$. Any mismatch in these exact totals would show the database, and therefore the training set, is unreliable.","tokens_in":17303,"feed_emoji":"🧮","tokens_out":14293,"duration_ms":131016,"temperature":0.7,"pith_summary":"This paper sets out to show that Galois groups of polynomials can be classified by a neurosymbolic network: a standard feedforward classifier surrounded by classical symbolic layers for reduction modulo $p$, discriminant square-testing, and real-root counting. The authors build exact databases of irreducible cubics, quartics, and quintics of bounded height whose Galois groups are known, then use them to train and test the hybrid model. The databases yield a concrete number-theoretic discovery: cyclic Galois groups are extremely rare at small height—for example, exactly 20 irreducible quintics of height at most 10 have Galois group $C_5$, in three $\\mathbb{Q}$-isomorphism classes. If the central claim is correct, this architecture is a template for pushing Galois-group computation to degrees where pure symbolic methods slow down and pure learning cannot see rare groups.","feed_headline":"Hybrid AI finds Galois groups, and cyclic quintics are rare","feed_subtitle":"Adding symbolic layers to a neural net predicts Galois groups and counts how rare cyclic C5 quintics really are.","key_machinery":"The central object is the GaloisNetwork, a neurosymbolic classifier that combines a feedforward network (input layer, three hidden layers of 64 ReLU units, softmax output) with symbolic layers that encode classical Galois theory. The signature layer reduces the polynomial modulo $p = 2,3,5,7$, records the factorization degree pattern, and compares it against the precomputed list of transitive-subgroup signatures, fixing the group when exactly one candidate remains. The discriminant layer checks whether $\\Delta_f$ is a square in $\\mathbb{Q}$, separating subgroups of $A_d$ from those that are not. The real-root layer counts real roots by sign variations and applies the theorem that a prime-degree polynomial with few non-real roots must have Galois group $A_p$ or $S_p$. The training databases are built by enumerating projective points of bounded height, keeping irreducible forms, and labelling each by its invariants and Galois group; for quintics the invariants $J_4, J_8, J_{12}, J_{18}$ and the resolvent quantities $d_1,\\dots,d_6$ carry the solvability criterion.","core_discovery":"The paper's central claim is that a hybrid network whose neural predictions are filtered through symbolic checks outperforms a plain neural network for classifying Galois groups of irreducible polynomials over $\\mathbb{Q}$ in small degrees. The symbolic checks are the factorization signature of $f(x) \\bmod p$ for $p = 2,3,5,7$, the question of whether the discriminant $\\Delta_f$ is a square, and the number of real roots obtained by sign-variation counting; the paper reports that these layers, combined with the invariants of binary forms, resolve the Galois group directly in many cases and leave only genuinely ambiguous polynomials to the neural head. The supporting databases are exact enumerations of projective points of bounded height, filtered to irreducible polynomials and labelled by Galois group. On the basis of those databases the paper states exact counts, including 5,676 irreducible quartics of height $\\leq 10$ with Galois group not isomorphic to $S_4$ (5162 $D_4$, 184 $A_4$, 222 $V_4$, 108 $C_4$) and, for quintics of height $\\leq 10$, exactly 20 with Galois group $C_5$, 480 with $F_5$, 900 with $D_5$, and 1,146 with $A_5$, with the $C_5$ examples falling into three $\\mathbb{Q}$-isomorphism classes.","pith_inferences":["A natural next experiment is a head-to-head accuracy-versus-training-size comparison with a plain network on identical splits; the efficiency claim in the paper is architectural and such a curve would make it quantitative.","If the enumeration is exact, the 20-quintic count implies that a purely statistical learner would need to see nearly the whole height-$\\le 10$ population to collect even a handful of $C_5$ examples; the symbolic layers supply exactly the inductive bias that makes such rare classes learnable.","The same modular recipe—cheap symbolic necessary conditions plus a neural head—transfers naturally to other classification problems in algebraic geometry where the paper's moduli-height machinery already applies, such as automorphism groups or isogeny classes.","A direct scaling test would be to run the same architecture on sextics and septics with identical height bounds, since the paper identifies those degrees as the point where symbolic computation alone begins to slow down."],"forward_implications":["When the mod-$p$ signature is unique to one transitive subgroup, the symbolic layer fixes the Galois group before learning is needed, so the neural component only has to handle the remaining ambiguous polynomials.","The exact rarity counts give a concrete check for distribution conjectures about bounded-height polynomials, such as the expected frequency of cyclic groups and recent bounded-invariant counts for quartic forms.","For degrees 6 and 7, where transitive subgroup lists are much larger and signatures are no longer unique, the same three-layer architecture is the paper's proposed route for keeping classification tractable.","Using absolute invariants to collapse $\\mathbb{Q}$-equivalent polynomials shrinks the effective dataset and, according to the paper, makes the learning problem easier rather than harder."],"supporting_citations":[{"why":"Defines the naive height, weighted moduli height, and invariants of binary forms that the paper uses to build and compare the polynomial databases.","marker":"[16]"},{"why":"Identifies the public database of labelled polynomials that serves as both the training data and the source of the exact enumeration counts.","marker":"[18]"},{"why":"Provides the preceding construction of polynomial datasets with Galois labels and implementation details for the deep-learning components.","marker":"[17]"},{"why":"Gives the theorem that a prime-degree polynomial with few non-real roots must have Galois group A_p or S_p, which anchors the real-root symbolic layer.","marker":"[3]"},{"why":"Supplies the bounded-invariant theory of binary quartic forms used to validate the quartic database and to compare invariant height with naive height.","marker":"[2]"},{"why":"Supplies the explicit expressions for the resolvent coefficients d_1,...,d_6 in terms of J_4, J_8, J_12, used for the quintic solvability criterion.","marker":"[8]"},{"why":"Provides the invariant-theoretic background and the discriminant relation for quintics (Delta = J_4^2 - 128 J_8) used in the feature and discriminant layers.","marker":"[6]"},{"why":"Lists the invariants of binary forms used in the proof that the resolvent discriminant is nonzero, which supports the quintic solvability criterion.","marker":"[4]"}],"fun_headline_variants":["Neurosymbolic net beats plain AI at Galois groups","Cyclic quintics are rare, says neurosymbolic AI","Hybrid AI: Galois groups solved, cyclic quintics scarce","AI with symbolic checks outperforms neural nets on Galois","Rarity of cyclic quintics revealed by neurosymbolic net"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole edifice rests on the assumption that the computer enumeration of bounded-height irreducible polynomials is complete and that every Galois label attached to a polynomial is correct; a single wrong count or wrong label would corrupt the rarity discovery and every training example built from that data.","fun_headline_variants_meta":{"raw":{"variants":["Neurosymbolic net beats plain AI at Galois groups","Cyclic quintics are rare, says neurosymbolic AI","Hybrid AI: Galois groups solved, cyclic quintics scarce","AI with symbolic checks outperforms neural nets on Galois","Rarity of cyclic quintics revealed by neurosymbolic net"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000491,"raw_usage":{"total_tokens":2409,"prompt_tokens":938,"completion_tokens":1471,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":1385}},"tokens_in":554,"tokens_out":1471,"duration_ms":11138,"temperature":1.0,"reasoning_tokens":1385,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:34:35.640355+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the bounded-height enumeration with an independent implementation: for degree 5, height at most 10, list all irreducible integer quintics with projective height $\\leq 10$, compute their Galois groups by a resolvent-based method, and compare the counts. The paper predicts exactly 20 polynomials with Galois group $C_5$ in three $\\mathbb{Q}$-isomorphism classes, 480 with $F_5$, 900 with $D_5$, and 1,146 with $A_5$; for quartics it predicts 5,676 irreducible polynomials of height $\\leq 10$ not having Galois group $S_4$. Any mismatch in these exact totals would show the database, and therefore the training set, is unreliable.","supporting_citations":[{"cited_title":"Machine learning for moduli space of genus two curves and an application to isogeny based cryptography","cited_arxiv_id":"2403.17250","evidence_quote":"Defines the naive height, weighted moduli height, and invariants of binary forms that the paper uses to build and compare the polynomial databases."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Identifies the public database of labelled polynomials that serves as both the training data and the source of the exact enumeration counts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the preceding construction of polynomial datasets with Galois labels and implementation details for the deep-learning components."},{"cited_title":"Bialostocki and T","cited_arxiv_id":null,"evidence_quote":"Gives the theorem that a prime-degree polynomial with few non-real roots must have Galois group A_p or S_p, which anchors the real-root symbolic layer."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the bounded-invariant theory of binary quartic forms used to validate the quartic database and to compare invariant height with naive height."},{"cited_title":"Bruce King","cited_arxiv_id":null,"evidence_quote":"Supplies the explicit expressions for the resolvent coefficients d_1,...,d_6 in terms of J_4, J_8, J_12, used for the quintic solvability criterion."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the invariant-theoretic background and the discriminant relation for quintics (Delta = J_4^2 - 128 J_8) used in the feature and discriminant layers."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Lists the invariants of binary forms used in the proof that the resolvent discriminant is nonzero, which supports the quintic solvability criterion."}],"review_version":1}