{"id":"0dfcbffe-df77-4e80-90fd-ba1a98f6a167","arxiv_id":"2501.09571","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A neural network that learns matrix-valued group representations by design achieves near-perfect accuracy and length extrapolation on group prediction tasks, outperforming fixed-representation, MLP, LSTM, and Transformer baselines.","lead":"MatrixNet is a neural network that learns its own matrix representations of group element inputs, such as rotations or braids, instead of using fixed mathematical representations. The paper shows this learned representation beats standard baselines on order prediction for symmetric groups and on a braid group task tied to open math problems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Soft, unablated relation loss leaves the length-extrapolation claim causally unsupported.","rationale":"MatrixNet is a clean and promising architecture: the matrix block provably gives a free-group representation, the parameter count is competitive, and the empirical results on S10 and B3 are strong and reproducible in principle. However, the central scientific claim—that relation regularization enables extrapolation to longer words—is exactly the kind of claim that needs a causal test. The paper's own Table 4 shows the best variants still have relation error 0.066–0.071; the limitation statement admits non-exact homomorphism. Naively, such an error could compound over a word of length 8–10 and produce large matrix discrepancies, yet the observed MSE is minuscule. This discrepancy leaves open at least two alternative explanations: (1) the test set's long words reduce mostly to short training words, so the free-group product alone suffices; (2) the MLP output is insensitive to the residual error, so the model is not truly 'respecting relations' in the sense claimed. Without an ablation removing L_rel or varying its weight, and without measuring relation error on the actual test-length words, the causal claim is unsubstantiated. This does not impugn the empirical results, which are impressive; it means the abstract's assertion should be softened or, better, backed by the proposed ablation. The reader's CONDITIONAL verdict is appropriate; my concern is the same as the reader's weakest assumption, so no verdict change is needed.","tokens_in":16419,"tokens_out":8882,"duration_ms":90927,"concrete_test":"Train MatrixNet-LN and MatrixNet-NL on the B3 Jordan-Hölder dataset with the relation loss weight λ set to 0, 0.01, 0.1, and 1.0 (using the same every-10-batches schedule), and compare extrapolation MSE on length 8, 9, and 10 words. Simultaneously, for each trained model, compute the relation error on random words of length up to 20 by comparing M_w for w and its braid-normal-form reduction (e.g., using Garside normal form). If extrapolation MSE does not degrade substantially when λ=0, the length generalization is not attributable to relation learning; if relation error grows with word length while test MSE stays low, the claim needs an explanation of how the MLP tolerates the error. This isolates the causal role of the relation loss.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim—generalization to group elements of greater word length via relation regularization—depends on the learned matrices M_g being an exact (or nearly exact) homomorphism from the group to GL(n). The architecture guarantees a free-group representation exactly (Prop. 1-2), but the quotient structure is enforced only by the auxiliary loss L_rel (Sec. 4.3), which penalizes a finite set of relation words and is added every 10 training batches with no reported weight. For B3, the single braid relation σ1σ2σ1 = σ2σ1σ2 is enforced only in this soft manner. Table 4 reports relational error ~0.07 (Frobenius norm of M_{σ1σ2σ1} - M_{σ2σ1σ2}) for the best variants, while the non-relational difference is ~5; this is not negligible, yet the extrapolation MSE (Fig. 2) is near 0.001. The paper provides no analysis of how this error propagates as word length grows, nor any ablation showing that extrapolation fails without L_rel. It is possible that the test words of length 8–10 largely reduce to words of length ≤7 seen in training, making the 'extrapolation' a consequence of the free-group product rather than of learned relations; or that the MLP is insensitive to the residual relation error. Without ablating the relation loss or measuring relational error on longer words, the causal claim that relation regularization drives extrapolation is unsupported. The Limitations section itself concedes 'the homomorphism property is not exact,' which directly contradicts the abstract's assertion that MatrixNet 'respects group relations allowing generalization.'","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MatrixNet, a neural architecture that learns matrix representations of group generators from sequences, with the product of generator matrices representing a group element. The matrix block uses a matrix exponential so that the map from the free group to GL(n) is a homomorphism by construction, while group relations are encouraged through an auxiliary loss that penalizes deviation of relation words from the identity. The authors evaluate MatrixNet on order prediction in finite symmetric and product groups and on Jordan-Hölder multiplicity prediction for the braid group B3, reporting that the best variants outperform MLP, LSTM, and Transformer baselines and maintain near-zero error on braid words of length 8-10 after training only on words up to length 7. The central claim is that learned matrix representations, aided by relation regularization, generalize to group elements of greater word length than seen in training.","tokens_in":16653,"tokens_out":3811,"duration_ms":39151,"significance":"If the central claim holds, MatrixNet offers a parameter-efficient and interpretable way to learn functions on discrete groups with a learned representation, and the connection to an open problem in categorical braid group actions is a compelling application. The paper has several concrete strengths: Propositions 1 and 2 correctly establish that the matrix block defines a free-group representation; the code is released; the experiments compare against multiple baselines; and Table 4 provides a diagnostic (relational error) that is correlated with extrapolation performance. However, the headline length-extrapolation claim is supported only by a soft, finite-set regularization loss with no ablation, no error-propagation analysis, and no control for test words that reduce to shorter words. The significance is therefore conditional on additional experiments that directly test the causal role of relation regularization.","major_comments":[{"comment":"The causal claim that relation regularization is what enables length extrapolation is not established, because no ablation removes the auxiliary relation loss L_rel. The near-zero MSE of MatrixNet-LN and MatrixNet-NL on words of length 8-10 could in principle arise from the free-group matrix product structure and the MLP's interpolation, rather than from learned relations. Please report the same extrapolation experiment for models trained with the relation-loss weight set to zero, and show whether relational error and extrapolation MSE degrade.","section":"§5.2, Figure 2"},{"comment":"L_rel penalizes only a finite set of relation words, yet the extrapolation claim requires the approximate homomorphism property to hold for all words of length 8-10. The paper reports relational error of about 0.066-0.071 for the best variants, and the Limitations section states that the homomorphism property is not exact, but no analysis is given of how this residual error compounds with word length or how it translates into MSE on the length-8, length-9, and length-10 test sets. Please measure relational error directly on those test words, e.g., the Frobenius norm of M_w1 - M_w2 for equivalent words of each test length, and provide a bound or empirical error-propagation analysis.","section":"§4.3, Appendix B.2, Table 4"},{"comment":"The length extrapolation experiment does not report the fraction of test words of length 8-10 that reduce, via the braid relation, to group elements representable by words of length at most 7. If a large fraction of these test words are equivalent to shorter words already seen in training, the result would be partly explained by invariance to known relations rather than by genuine composition of longer words. Please report the distribution of minimal word lengths for the test sets, and show results separately for words that are irreducible to length at most 7.","section":"§5.2, Length Extrapolation Results"}],"minor_comments":[{"comment":"The regularization details state that the braid relation used is \"σ1σ2σ2 = σ2σ1σ2\", which appears to be a typo; the correct braid relation is σ1σ2σ1 = σ2σ1σ2, and the inverse relation should be σ1^{-1}σ2^{-1}σ1^{-1} = σ2^{-1}σ1^{-1}σ2^{-1}. Please correct this.","section":"Appendix B.2"},{"comment":"The sentence \"Figure 2 shows visual plots of the learned representations\" appears to refer to Figure 3, since Figure 2 contains the length extrapolation plots.","section":"§5.3"},{"comment":"The phrase \"over the several finite groups\" should be \"over several finite groups\".","section":"Abstract"},{"comment":"The presentation of D4 as \"r4 = f2 = f rf r\" is incomplete or ambiguous; it should be written with relations such as r^4 = f^2 = (fr)^2 = 1.","section":"§3.1"},{"comment":"The model name \"MatrixNet-Nonlinear\" in Table 1 is inconsistent with the name \"MatrixNet-NL\" used elsewhere.","section":"§5.1"},{"comment":"The regularization loss weight is not reported; the paper only states that the term is added every 10 training batches. Please specify the weight and whether it is constant or scheduled.","section":"Appendix B.2"}],"recommendation":"major_revision","confidential_remarks":"The paper's abstract and contribution list emphasize length extrapolation via relation regularization, so the missing ablations and error-propagation analysis are not cosmetic; they are central to the main claim. I would be willing to reconsider after the authors add the proposed experiments and analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the core idea is genuinely fresh: instead of feeding a fixed representation of a group element to a network, MatrixNet learns a homomorphism from the free group on the generators to GL(n) via matrix exponentials, then uses the product matrix as features, with an auxiliary loss that penalizes relation violations. That combination is new and it works surprisingly well on the two tasks: order prediction in S10 and Jordan-Hölder multiplicities for B3. The matrix-block construction is sound; Propositions 1 and 2 are correct. Second, the headline claim that the model 'respects group relations allowing generalization to greater word length' is not actually established. The relation loss is soft—added every 10 batches, weight unspecified—and Table 4 shows residual relational error ~0.07 for the best variants. There is no ablation without the relation loss, no analysis of how the error compounds with word length, and the correct braid relation is misprinted in Appendix B.2 (σ1σ2σ2 = σ2σ1σ2 should be σ1σ2σ1 = σ2σ1σ2). The extrapolation result might be real, but as written the causal story is under-supported.\n\nWhat's solid: the empirical wins over baselines are large and consistent across variants, the finite-group experiments are sensible, and the code is public. What's soft: the finite-group tables don't report per-trial variance or trial counts; the regularization schedule is underspecified; and the 'automatic generalization' claim rests on a soft constraint. The reader's conditional verdict is fair. The stress-test note is right that the extrapolation claim needs an ablation and an error-propagation estimate. I'd ask the authors for: (1) the relation-loss weight and schedule, (2) an ablation without L_rel, (3) relational error on words of length 8–10, not just the generators' relation, (4) a corrected braid relation, and (5) trial counts in the finite-group tables.\n\nBottom line: worth a serious referee. The architecture is worth building on and the evidence, though incomplete, points in a plausible direction. I'd accept it for review and ask for revisions.","headline":"Learned group representations are a real idea, but the length-extrapolation claim needs an ablation and a corrected relation before it holds.","tokens_in":17250,"tokens_out":1693,"would_cite":true,"duration_ms":18110,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","20F36","20C30"],"pacs":[],"model":"deepseek-v4-flash","headline":"MatrixNet learns its own matrix representations of group elements, and the learned relations let it generalize to longer words.","keywords":["group representation learning","matrix exponential","braid group","group relations","signed one-hot encoding","Jordan-Hölder multiplicities","sample efficiency","length extrapolation"],"falsifier":"Train MatrixNet-LN or MatrixNet-NL on words up to length 7 as described, then evaluate the relation error $\\|M_{\\sigma_1\\sigma_2\\sigma_1\\sigma_2\\sigma_1\\sigma_2} - M_{\\sigma_2\\sigma_1\\sigma_2\\sigma_1\\sigma_2\\sigma_1}\\|$ and the MSE on length-12 words. If the relation error grows with word length while the non-relational difference stays bounded, or if the length-12 MSE jumps well above training-level MSE, the claim that relation regularization confers automatic generalization to longer words is not supported.","tokens_in":16146,"feed_emoji":"🧮","tokens_out":5497,"duration_ms":55433,"temperature":0.7,"pith_summary":"The paper proposes an architecture, MatrixNet, that turns group elements into matrices learned from data rather than fixed representation-theoretic ones. It tests this on predicting element order in finite groups and on predicting Jordan-Hölder multiplicities under a braid group action. Across those tasks the learned representations beat or match MLP, LSTM, and Transformer baselines with comparable parameter counts. The paper's strongest claim is that a soft relation loss makes the learned representation respect group relations, so performance holds on braid words longer than any seen in training.","feed_headline":"Learned group matrices beat fixed ones and reach unseen word lengths","feed_subtitle":"MatrixNet maps generators to learned matrices, outperforming baselines and generalizing to longer braid words.","key_machinery":"The matrix block: each generator $g_i$ is encoded by a signed one-hot vector $v_{g_i}$, reshaped to a matrix $A_i$, and exponentiated to $M_{g_i} = \\exp(A_i)$, with $M_{g^{-1}} = \\exp(-A_i) = M_{g_i}^{-1}$ and $M_1 = I$. A word $g = g_{i_1} \\cdots g_{i_\\ell}$ is represented by the matrix product $M_g = M_{g_{i_1}} \\cdots M_{g_{i_\\ell}}$. Because the matrices are invertible by construction, the map is a homomorphism from the free group; a relation-loss term $L_{\\text{rel}} = \\sum_{r \\in R} \\|M_r - I\\|$ (Frobenius norm) nudges the representation to satisfy the group's relations, and the flattened matrix feeds a task MLP.","core_discovery":"The central discovery is that group elements can be encoded by matrices learned from a presentation of the group, and that these learned representations carry enough algebraic structure to improve both accuracy and length generalization. Concretely, MatrixNet maps each generator to an invertible matrix via a signed one-hot encoding, a reshape, and a matrix exponential; a group element is represented by the product of its generator matrices. Group axioms (identity, inverses) hold by construction, and group relations are encouraged by a regularization term that pushes the matrix of each relation word toward the identity. On the braid group task, the best variants attain mean squared error near 0.001 and about 99.9% accuracy, compared with errors of 0.075 to 2.895 for the baselines; after training on words up to length 7, they keep near-zero MSE on lengths 8 through 10 while baselines degrade sharply.","pith_inferences":["The length-extrapolation claim rests on a finite set of relation words in the loss, so a natural stress test is to measure $\\|M_w - I\\|$ for longer words that are products or conjugates of the braid relation; if the error grows with word length, the automatic-generalization result is limited to the regularized relations.","If learned representations can be driven to near-exact homomorphisms, MatrixNet could serve as a computational probe for open questions in representation theory, generating candidate outputs or outliers for mathematicians to verify; the paper stops short of making this claim.","The block-diagonal variant suggests a path toward incorporating known irreducible structure while still learning the specific representation, which could be extended to continuous groups by replacing the matrix exponential with a Lie-algebra generator formulation."],"forward_implications":["Higher sample efficiency: with roughly equal parameter counts, MatrixNet variants reach accuracy levels that MLP, LSTM, and Transformer baselines do not reach on the braid-group prediction task.","Length extrapolation: after training only on braid words up to length 7, MatrixNet-LN and MatrixNet-NL keep near-zero MSE on words of length 8, 9, and 10, showing that relation regularization can replace data augmentation for unseen words.","Parameter efficiency: because only generator matrices are learned, the parameter count does not scale with the size of the group, making the method applicable to infinite groups such as the braid group.","Flexibility across group structures: the same architecture handles symmetric groups, product groups, abelian groups, and the infinite braid group without task-specific representations.","Interpretable learned representations: the learned matrices for equivalent braid words become nearly identical even when the relation was not directly regularized, suggesting the representation captures group structure rather than memorizing words."],"supporting_citations":[{"why":"Supplies the categorical braid action and the Jordan-Hölder multiplicity setup that defines the main benchmark task.","marker":"[44]"},{"why":"Establishes that spherical twist functors satisfy braid relations, providing the group action used in the task.","marker":"[45]"},{"why":"Gives the complete answer for the 3-strand braid group case, providing ground truth for the braid-action predictions.","marker":"[46]"},{"why":"Provides the algorithm used to generate additional Jordan-Hölder multiplicity examples for longer braid words.","marker":"[47]"},{"why":"Supplies the maximum word length formula for adjacent transpositions used to construct the symmetric group dataset.","marker":"[49]"},{"why":"Maschke's theorem underlies the block-diagonal decomposition that motivates the MatrixNet-MC variant.","marker":"[41]"},{"why":"Provides the classification of symmetric group representations that underlies the fixed-representation baseline.","marker":"[43]"}],"fun_headline_variants":["MatrixNet learns matrix reps, beating fixed ones and longer words","Learned matrices for group elements outperform fixed reps and reach new lengths","MatrixNet encodes group elements via learned matrices, boosting length generalization","MatrixNet turns generators into learned matrices, generalizing to unseen word lengths"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The auxiliary relation loss is applied only to a finite set of relation words, so the learned matrices must be close enough to an exact homomorphism that small relation errors do not compound over longer unseen words; if they do compound, the length-extrapolation result weakens.","fun_headline_variants_meta":{"raw":{"variants":["MatrixNet learns matrix reps, beating fixed ones and longer words","Learned matrices for group elements outperform fixed reps and reach new lengths","MatrixNet encodes group elements via learned matrices, boosting length generalization","MatrixNet turns generators into learned matrices, generalizing to unseen word lengths"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000916,"raw_usage":{"total_tokens":3868,"prompt_tokens":819,"completion_tokens":3049,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":435,"completion_tokens_details":{"reasoning_tokens":2975}},"tokens_in":435,"tokens_out":3049,"duration_ms":20942,"temperature":1.0,"reasoning_tokens":2975,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T19:51:57.333620+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train MatrixNet-LN or MatrixNet-NL on words up to length 7 as described, then evaluate the relation error $\\|M_{\\sigma_1\\sigma_2\\sigma_1\\sigma_2\\sigma_1\\sigma_2} - M_{\\sigma_2\\sigma_1\\sigma_2\\sigma_1\\sigma_2\\sigma_1}\\|$ and the MSE on length-12 words. If the relation error grows with word length while the non-relational difference stays bounded, or if the length-12 MSE jumps well above training-level MSE, the claim that relation regularization confers automatic generalization to longer words is not supported.","supporting_citations":[{"cited_title":"Picard groups for derived module categories","cited_arxiv_id":null,"evidence_quote":"Gives the complete answer for the 3-strand braid group case, providing ground truth for the braid-action predictions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the algorithm used to generate additional Jordan-Hölder multiplicity examples for longer braid words."},{"cited_title":"Humphreys","cited_arxiv_id":null,"evidence_quote":"Supplies the maximum word length formula for adjacent transpositions used to construct the symmetric group dataset."},{"cited_title":"Ueber den arithmetischen charakter der coefficienten der substitutionen endlicher linearer substitutionsgruppen","cited_arxiv_id":null,"evidence_quote":"Maschke's theorem underlies the block-diagonal decomposition that motivates the MatrixNet-MC variant."}],"review_version":1}