Pith. sign in

REVIEW 3 major objections 3 minor 36 references

Cycle inconsistency is not cohomology; certification requires a frozen comparison complex and full structural gates.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 09:06 UTC pith:ZAV4REJA

load-bearing objection A careful, honest paper that makes a useful conservative point—cycle inconsistency is not cohomology without a frozen comparison complex—but the positive certification branch still lacks the natural-data ingredient needed to apply it. the 3 major comments →

arxiv 2607.20887 v1 pith:ZAV4REJA submitted 2026-07-23 cs.LG cs.AImath.AG

TwistedMerge: Certified Higher-Order Diagnostics and Abstention for Model Merging

Ting Gong , Shitan Xu This is my paper
classification cs.LG cs.AImath.AG
keywords model merginggauge synchronizationhigher-order diagnosticscohomologyholonomyabstentiondescent theorycertification pipeline
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

TwistedMerge's central claim is that a nonzero cycle residual in model merging—three pairwise alignments failing to compose—does not by itself constitute a cohomological obstruction. A higher-obstruction diagnosis is meaningful only after a comparison complex is frozen before residual inspection, transition residuals are certified central and closed in a fixed coefficient system, and a repair model is stated. The paper proves error-control theorems for the resulting three-way decision (trivial, nontrivial on the complex, uncertified) and a no-go showing constant edge matrices cannot realize the tetrahedral obstruction witness. Its experiments show a planted alignment defect is removed by synchronization, naive low-rank factor averaging is gauge-dependent while global synchronization is stable, and natural checkpoint collections yield no certified central or period-index class. If correct, the practical upshot is that cycle scores should not be used as obstruction predictors without the full certification, and that conservative abstention is the safe default.

Core claim

The paper establishes that a nonzero class in H2(K; A) on a finite comparison complex K rules out only the tested A-valued edge repair on that stated K, and that no natural Brauer or period-index class is certified in the present experiments. It formalizes a finite descent instance in which checkpoints are vertices, alignments are edge transitions, and triangle products are residuals, and it separates four regimes: fixed-chart averaging, synchronization-removable gauge inconsistency, a certified central obstruction on a frozen complex, and nonabelian holonomy. The central theoretical results are the frozen-complex three-way error-control theorem, the predeclared-family error-control theorem,

What carries the argument

The engine of the framework is the frozen comparison complex K = Φ(DK), a finite simplicial complex constructed from data that are independent of the residuals and transition fits, together with the distance-to-coboundaries obstruction norm Def(u) on central 2-cochains. The pipeline computes triangle residuals, applies inverse-consistency, centrality, projection-fidelity, and closure gates, and then returns trivial, nontrivial on K, or uncertified based on confidence intervals for Def(bc) relative to the coboundary subspace. The error-control mechanism is Theorem 2.2.9, which uses the 1-Lipschitz stability of the distance map to prevent false trivial and false nontrivial declarations. The no

Load-bearing premise

The whole scheme depends on deciding in advance which model-overlap contexts exist, before any residual is inspected; the paper admits no such principled rule yet exists for natural checkpoints, and without a pre-frozen comparison complex a certified 'nontrivial' verdict can be manufactured post hoc.

What would settle it

Run the full frozen-complex protocol on a natural checkpoint collection and find one certified nontrivial central class in H2(K; A); the paper's natural-data conclusion says no such class is certified, so a single certified natural instance would falsify that claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Cycle residuals alone should not be used to claim higher-order merge failure; any such claim requires a pre-registered comparison complex, coefficient identification, centrality, and closure.
  • In natural checkpoint collections, cycle residual is not a reliable predictor of merge degradation (negative held-out R^2), so validation-loss information is more trustworthy.
  • Naive averaging of low-rank adapter factors is gauge-dependent; stable merging requires global factor synchronization or dense-delta SVD, not direct factor averaging.
  • The three-way decision rule with abstention can control error rates under estimated transitions, so certified nontrivial status is meaningful only when the frozen-complex gates pass.
  • Conservative abstention—returning an ordinary or synchronized fallback—is the default when structural certificates are absent; this prevents unsupported rank or branch lifts.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The framework implies a methodological rule for the field: any future claim of a cohomological obstruction in real models should be accompanied by the frozen-complex and coefficient-system certificate; otherwise it is unfalsifiable by this paper's standard.
  • The no-go result suggests that if genuine higher obstructions exist in neural checkpoints, they must arise from overlap-dependent alignments rather than fixed matrices; a concrete next experiment is to search for such overlap-dependent transition sections in trained transformer stacks.
  • The abstention design may generalize beyond merging: any diagnostic that can be gamed by post hoc structure selection can adopt the same trivial/nontrivial/uncertified triage to keep claims honest.
  • If natural Brauer classes are ever certified, the period-index gate predicts that only ranks divisible by the class index can support a compatible lift; an apparent successful lift at a non-index rank would indicate a flaw in the lift construction.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 3 minor

Summary. This paper introduces TwistedMerge, a certification and abstention framework for model merging. It formalizes finite descent data on a comparison complex K, separates raw cycle inconsistency from synchronization-removable gauge defects, central H2(K;A) obstructions, and nonabelian holonomy, and returns a three-way decision (trivial, nontrivial on K, uncertified). The main theoretical results are a constant-edge no-go proposition for the tetrahedral boundary, a frozen-complex three-way error-control theorem, simultaneous control over a predeclared family of complexes, a refinement-persistence proposition, and period/index rank gates for projective representations. The experimental program includes a causal planted-alignment defect, a LoRA gauge-invariance audit, controlled tetrahedral and finite-Heisenberg systems, robust calibration under noise, and natural-data studies. The paper explicitly reports negative natural results: cycle residuals do not predict merge degradation and no natural Brauer or period-index class is certified.

Significance. If the claims are accepted, the paper makes a valuable negative and methodological contribution: it shows that a nonzero cycle residual alone is not a higher-order obstruction, and it specifies the conditions under which a cohomological interpretation is warranted. The frozen-complex protocol and the three-way abstention rule are genuinely useful safeguards, and the separation of confirmatory from exploratory claims, plus the honest reporting of null results, is exemplary. The controlled finite-Heisenberg experiments verify concrete projective-representation rank thresholds, and the LoRA audit gives a crisp demonstration of gauge dependence in factor averaging. Reproducibility is a clear strength: the code, audits, and report snapshots are referenced. The main weakness is that the certified positive branch is only operational when K and the coefficient system are specified by design; for natural checkpoints, the required ex ante comparison complex is admitted to be unavailable, so the practical contribution currently reduces to abstention and negative diagnostics.

major comments (3)
  1. [§4, first limitation; Definition 2.2.8] The frozen comparison complex K = Φ(DK) is load-bearing for the certified branch. The manuscript explicitly states that the present natural checkpoint collections do not yet supply an application-grounded higher-overlap rule, so no natural K is available. Without such a rule, any K chosen on natural data would risk being selected after residual inspection, and Theorem 2.2.9 would not apply. The positive branch is therefore demonstrated only on prescribed synthetic complexes. This is not an internal contradiction, but it is an unmet precondition for the paper's central certified-diagnostics claim. The revision should either provide a concrete predeclared Φ on at least one natural checkpoint collection, or explicitly narrow the contribution to a synthetic-control/no-go framework whose certified branch is not yet operational for real model merging.
  2. [Theorem 2.2.9; Algorithm 1] The unconditional probability bound in Theorem 2.2.9 relies on DK being independent of Dcert. The manuscript does not describe how this independence is achieved in the 120-collection natural study. Availability scores, transition-map estimates, and certification residuals are often computed from overlapping data (the same checkpoints, datasets, or preprocessing), so the independence assumption is nontrivial. The paper should specify the actual data splits for DK, Dalign, Dcert, Dselect, and Dtest in the natural experiments, or restate the guarantee as conditional on the independence assumption. This is directly relevant to whether the certified branch can be transferred from controlled constructions to real model-merging problems.
  3. [§3.4 and §3.5; Tables 10–13] The empirical support for the nontrivial-certificate branch is entirely internal to prescribed synthetic systems: the tetrahedral complex is given as ∂∆3, the coefficient groups are fixed by construction, and the finite-Heisenberg systems supply their own commutation relations. The causal planted benchmark in §3.3 is a useful bridge, but it uses exact functional copies with one corrupted edge. The natural-data section then shows that all structural gates fail. The paper's central claim is thus supported only in the negative direction for natural checkpoints. A major revision should either add a semi-natural or natural setting in which the full certified branch is exercised with a predeclared Φ and coefficient identification, or make the absence of such a setting an explicit, prominent scope limitation in the abstract and introduction rather than only in Section 4.
minor comments (3)
  1. [Abstract and text] There are several formatting glitches: 'Appendix Appendix B' in Section 2, 'P .R. China' in the author affiliation, and 'arXi v:2511.21437' in Reference [8]. These should be corrected.
  2. [Code/data availability] The audited repository commit is given as b0e1ac4 in the text, while Reference [12] points to commit 7a0620bb19dffba97012350b6ffd20684bcbe220. The version used for the numerical claims should be unambiguous; please reconcile the two identifiers.
  3. [Table 8] The table reports 'memory figures are analytical counts' and notes that the fixtures are scaled from trained factors. This is stated clearly, but the caption could also say that the timing comparison is illustrative and not a benchmark claim for the method as a whole.

Circularity Check

1 steps flagged

Central no-go, abstention, and natural-negative results are self-contained; only the controlled oracle 'recovery' experiments are mildly circular because the witnesses are generated from the same theorems that define the predictions.

specific steps
  1. self definitional [Section 3.4.1, tetrahedral H2(μ2) witness; cf. Abstract ('Controlled central systems recover the predicted non-coboundary and projective-rank behavior')]
    "Let K=∂∆3 and let the face signs be (c012,c013,c023,c123)=(−1,+1,+1,+1). By Theorem 2.3.19, this cocycle is not a coboundary."

    The 'predicted' non-coboundary certificate is exactly Theorem 2.3.19 applied to an oracle cochain that was constructed to have one negative face, i.e. to satisfy the same product criterion the theorem uses. Table 10's 'coboundary? no' output is therefore entailed by the construction rather than independently corroborated. The same holds for the finite-Heisenberg rank thresholds: the clock–shift systems realize precisely the relation whose rank divisibility is proved, so 'recovering' d^k is a consistency check of the implementation. The paper explicitly labels these cases as prescribed oracle/controlled witnesses, so this is a minor self-consistency circularity and is not load-bearing for the paper's main no-go or natural-data conclusions.

full rationale

The paper's load-bearing derivation chain is largely self-contained. Definition 2.2.8 freezes K=Φ(DK) before residual inspection, and Theorem 2.2.9 is an interval-arithmetic consequence of the 1-Lipschitz property of distance to coboundaries; it does not fit the certified class from the data. Proposition 2.3.6 and Theorem 2.3.19 are proved from the definitions, with no circular dependence on the experiments. The headline claim is explicitly conservative: no natural Brauer or period-index class is certified, and Section 4 admits that no application-grounded higher-overlap rule yet exists for natural checkpoint collections. That is an unfulfilled precondition for the positive certified branch, not a circular step. The self-citations ([12] code repository, [21] period-index background pointer) are not load-bearing; the actual period-index fact used is cited to the external textbook [19]. The only notable circularity is presentational: the controlled experiments that 'recover predicted non-coboundary and projective-rank behavior' use oracle witnesses constructed from the same theorems that define the predictions, so they verify the implementation rather than independently validating the framework. Because the central no-go, abstention, and natural negative results do not reduce to a fit or to a self-citation chain, the overall circularity score is low.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 3 invented entities

The central claim rests on a handful of modeling choices that are not independently grounded: the frozen complex K, the coefficient system A, and the statistical concentration bound ε. These are exactly the quantities the paper leaves unestimated for natural data, which is why the certificate's practical reach is currently limited to controlled constructions.

free parameters (3)
  • three-way decision thresholds (τ0, τ1 / τcent, τphase) = τcent = τphase = 3×10^-4, confidence margin = 0.25 in the selected policy; Table 12 sweeps 1e-5..3e-4.
    Hand-chosen and swept to produce favorable coverage-abstention numbers in the controlled calibration; no natural-data calibration is provided.
  • comparison-complex availability thresholds {λ_dim} = not specified
    Definition 2.2.8 requires predeclared thresholds to decide which simplices enter K. No concrete values or natural estimation procedure are given.
  • noise-tolerance ε and confidence δ in Theorem 2.2.9 = not estimated
    The error-control intervals require an a priori bound ||b_c−c||≤ε with probability 1−δ. The paper does not estimate ε on real or controlled certification data, so the three-way certificate's guarantee is conditional.
axioms (5)
  • domain assumption The learning site C_learn with a Grothendieck topology and a stack of local models with effective strict descent exists for model merging.
    Section 2.2 postulates this dictionary; if real checkpoints do not form such a stack, the descent theory is only an analogy.
  • domain assumption K = Φ(DK) is chosen independently of certification residuals.
    Definition 2.2.8 and Theorem 2.2.9 require this independence for no-post-hoc-complex selection. The paper admits no application-grounded K exists for natural checkpoints.
  • domain assumption Transition estimates satisfy ||b_c−c||≤ε with confidence 1−δ for the relevant cochain norm.
    Theorem 2.2.9 and Proposition 2.3.11 rely on this bound; ε is never estimated in the experiments.
  • domain assumption The central coefficient group A is abelian and lies in the center of the gauge group for the certified central branch.
    Definition 2.3.2 and Theorem 2.2.15 require centrality; centrality itself must be certified, and no natural A is identified.
  • standard math Equating finite H2(K;A) with site-level cohomology requires an appropriate comparison theorem/acyclicity.
    Remark 2.2.5 warns that identification with H2(C_learn;A) is not automatic. The paper is careful here, but the claim relies on the ambient theory when site-level language is used.
invented entities (3)
  • finite comparison complex K no independent evidence
    purpose: The frozen simplicial complex on which residuals are evaluated and certified.
    Central to the pipeline, but no application-grounded construction exists for natural checkpoints (Section 4); hence no external falsifiable handle.
  • central coefficient system A with projection π_A no independent evidence
    purpose: Recodes triangle residuals into an abelian cochain before H2 certification.
    The choice of A is not identified from data; controlled experiments choose A by construction, and natural data yield no certified A.
  • learning site C_learn / stack of local models no independent evidence
    purpose: Conceptual ambient structure for the descent interpretation.
    Postulated analogy; without a concrete site and descent datum, it provides no falsifiable prediction outside the paper's framing.

pith-pipeline@v1.3.0-alltime-deepseek · 41618 in / 14920 out tokens · 162781 ms · 2026-08-01T09:06:01.705590+00:00 · methodology

0 comments
read the original abstract

Model merging combines independently trained or fine-tuned models, but pairwise alignability does not imply globally consistent alignment. We formulate merging as a finite descent problem in which checkpoints are local objects, alignment maps are transitions, and cycle products are residuals. TwistedMerge is a conservative certification pipeline that separates fixed-chart averaging, synchronization-removable gauge inconsistency, a certified central obstruction on a specified comparison complex, and nonabelian holonomy. A residual is promoted to a cohomology class only after inverse-consistency, coefficient-identification, centrality, and closure tests; otherwise the method abstains and returns an ordinary or synchronized fallback. We prove a constant-edge no-go result, frozen-complex three-way and predeclared-family error-control theorems, and a refinement test for comparison-complex sensitivity. A planted neural alignment defect is removed by cycle-consistent synchronization, showing that a nonzero cycle score alone is not a higher obstruction. Controlled central systems recover the predicted non-coboundary and projective-rank behavior, while noisy estimates move from certification to abstention without false lifts on the tested controls. A trained low-rank-adapter audit shows that naive factor averaging depends on the chosen GLr representative, whereas global factor synchronization and dense-delta SVD are stable. On natural checkpoint collections, cycle residuals do not predict merge degradation and no natural central or period-index class is certified. The results position descent theory as a falsifiable certification and abstention framework.

Figures

Figures reproduced from arXiv: 2607.20887 by Shitan Xu, Ting Gong.

Figure 1
Figure 1. Figure 1: Core certification flow. The comparison complex is frozen before transition fitting and residual inspection. Structural branches are admitted only after their [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Causal parameter-level alignment inconsistency. The vertical axis is [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗
Figure 5
Figure 5. Figure 5: Robust period-index certification and abstention. Intermediate noise is [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Natural-data held-out prediction boundary. The horizontal axis is leave [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

36 extracted references · 10 linked inside Pith

  1. [1]

    S. K. Ainsworth, J. Hayase, S. Srinivasa, Git re-basin: Merging models modulo permutation symmetries (2022). arXiv:2209.04836. URLhttps://arxiv.org/abs/2209.04836

    Pith/arXiv arXiv 2022

  2. [2]

    Crisostomi, et al., c2m3: Cycle-consistent multi-model merging, in: Advances in Neural Information Processing Systems, 2024.arXiv:2405.17897

    D. Crisostomi, et al., c2m3: Cycle-consistent multi-model merging, in: Advances in Neural Information Processing Systems, 2024.arXiv:2405.17897

    Pith/arXiv arXiv 2024

  3. [3]

    Ilharco, et al., Editing models with task arithmetic, in: International Conference on Learning Representations, 2023

    G. Ilharco, et al., Editing models with task arithmetic, in: International Conference on Learning Representations, 2023

    2023

  4. [4]

    Yadav, et al., Ties-merging: Resolving interference when merging models, in: Advances in Neural Information Pro- cessing Systems, V ol

    P. Yadav, et al., Ties-merging: Resolving interference when merging models, in: Advances in Neural Information Pro- cessing Systems, V ol. 36, 2023, pp. 7093–7115

    2023

  5. [5]

    Yu, et al., Language models are super mario: Absorb- ing abilities from homologous models as a free lunch, in: International Conference on Machine Learning, 2024

    L. Yu, et al., Language models are super mario: Absorb- ing abilities from homologous models as a free lunch, in: International Conference on Machine Learning, 2024

    2024

  6. [6]

    Shoemake, Animating rotation with quaternion curves, ACM SIGGRAPH Computer Graphics 19 (3) (1985) 245– 254

    K. Shoemake, Animating rotation with quaternion curves, ACM SIGGRAPH Computer Graphics 19 (3) (1985) 245– 254

    1985

  7. [7]

    E. Yang, L. Shen, G. Guo, X. Wang, X. Cao, J. Zhang, D. Tao, Model merging in LLMs, MLLMs, and beyond: Methods, theories, applications and opportunities, arXiv preprint arXiv:2408.07666Revised December 2025 (2024). doi:10.48550/arXiv.2408.07666. URLhttps://arxiv.org/abs/2408.07666

    Pith/arXiv arXiv doi:10.48550/arxiv.2408.07666 2025

  8. [8]

    O. K. Hitit, L. Girrbach, Z. Akata, A systematic study of in-the-wild model merging for large language models, Transactions on Machine Learning Research (2026). arXi v:2511.21437. URLhttps://arxiv.org/abs/2511.21437

    arXiv 2026

  9. [9]

    Zheng, C

    D. Zheng, C. Allen-Blanchette, Beyond pairwise: Diagnos- ing higher-order merge failures via hodge decomposition, in: ICML 2026 Workshop on Weight-Space Symmetries, 2026. URL https://openreview.net/forum?id=5IsDyR zKks

    2026

  10. [10]

    Karuturi, K

    S. Karuturi, K. S. Bukkapatnam, L. Patel, T. A. Shastry, The GL(r) gauge symmetry of LoRA: Principal bundle structure, loss landscape geometry, and implications for adapter merging, in: ICML 2026 Workshop on Weight- Space Symmetries, 2026. URL https://openreview.net/forum?id=YMsbYX Dtqw

    2026

  11. [11]

    Javidnia, A gauge theory of superposition: Toward a sheaf-theoretic atlas of neural representations (2026).arXi v:2603.00824,doi:10.48550/arXiv.2603.00824

    H. Javidnia, A gauge theory of superposition: Toward a sheaf-theoretic atlas of neural representations (2026).arXi v:2603.00824,doi:10.48550/arXiv.2603.00824. URLhttps://arxiv.org/abs/2603.00824

    doi:10.48550/arxiv.2603.00824 2026

  12. [12]

    Gong, Twistedmerge: Twisted sheaves, descent the- ory, and learning theory, GitHub repository, commit 7a0620bb19dffba97012350b6ffd20684bcbe220 (2026)

    T. Gong, Twistedmerge: Twisted sheaves, descent the- ory, and learning theory, GitHub repository, commit 7a0620bb19dffba97012350b6ffd20684bcbe220 (2026). URL https://github.com/tinggong9/TwistedMe rge

    2026

  13. [13]

    Giraud, Cohomologie non abélienne, Springer, 1971

    J. Giraud, Cohomologie non abélienne, Springer, 1971

    1971

  14. [14]

    Lieblich, Moduli of twisted sheaves, Duke Mathemati- cal Journal 138 (1) (2007) 23–118

    M. Lieblich, Moduli of twisted sheaves, Duke Mathemati- cal Journal 138 (1) (2007) 23–118

    2007

  15. [15]

    Dodziuk, Finite-difference approach to the hodge the- ory of harmonic forms, American Journal of Mathematics 98 (1) (1976) 79–104

    J. Dodziuk, Finite-difference approach to the hodge the- ory of harmonic forms, American Journal of Mathematics 98 (1) (1976) 79–104

    1976

  16. [16]

    Horak, J

    D. Horak, J. Jost, Spectra of combinatorial laplace oper- ators on simplicial complexes, Advances in Mathematics 244 (2013) 303–336

    2013

  17. [17]

    H. H. Bauschke, P. L. Combettes, Convex Analysis and Monotone Operator Theory in Hilbert Spaces, 2nd Edition, Springer, 2017

    2017

  18. [18]

    R. A. Horn, C. R. Johnson, Matrix Analysis, 2nd Edition, Cambridge University Press, 2012

    2012

  19. [19]

    Gille, T

    P. Gille, T. Szamuely, Central Simple Algebras and Galois Cohomology, Cambridge University Press, 2006

    2006

  20. [20]

    Lieblich, Twisted sheaves and the period–index prob- lem, Compositio Mathematica 144 (1) (2008) 1–31

    M. Lieblich, Twisted sheaves and the period–index prob- lem, Compositio Mathematica 144 (1) (2008) 1–31

    2008

  21. [21]

    Gong, Moduli of vector bundles onµn-gerbes over genus 2 curves and the period–index problem (2025)

    T. Gong, Moduli of vector bundles onµn-gerbes over genus 2 curves and the period–index problem (2025). arXiv: 2512.03417. URLhttps://arxiv.org/abs/2512.03417

    arXiv 2025

  22. [22]

    A. J. de Jong, The period–index problem for the brauer group of an algebraic surface, Duke Mathematical Journal 123 (1) (2004) 71–94

    2004

  23. [23]

    Hotchkiss, A

    J. Hotchkiss, A. Perry, The period–index conjecture for abelian threefolds and donaldson–thomas theory (2024). arXiv:2405.03315. URLhttps://arxiv.org/abs/2405.03315

    Pith/arXiv arXiv 2024

  24. [24]

    Huybrechts, The period–index problem for hyperkähler manifolds (2024).arXiv:2411.17604

    D. Huybrechts, The period–index problem for hyperkähler manifolds (2024).arXiv:2411.17604. URLhttps://arxiv.org/abs/2411.17604

    Pith/arXiv arXiv 2024

  25. [25]

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chen, Lora: Low-rank adaptation of large language models (2021).arXiv:2106.09685. URLhttps://arxiv.org/abs/2106.09685

    Pith/arXiv arXiv 2021

  26. [26]

    M. Wortsman, et al., Model soups: Averaging weights of multiple fine-tuned models improves accuracy without increasing inference time, in: International Conference on Machine Learning, 2022.arXiv:2203.05482

    Pith/arXiv arXiv 2022

  27. [27]

    V . N. Vapnik, Statistical Learning Theory, Wiley- Interscience, 1998. 33

    1998

  28. [28]

    Shalev-Shwartz, S

    S. Shalev-Shwartz, S. Ben-David, Understanding Machine Learning: From Theory to Algorithms, Cambridge Univer- sity Press, 2014

    2014

  29. [29]

    G. L. Marchetti, et al., Algebra unveils deep learning: An invitation to neuroalgebraic geometry, in: International Conference on Machine Learning, 2025. arXiv:2501.1 8915

    2025

  30. [30]

    Kileel, M

    J. Kileel, M. Trager, J. Bruna, On the expressive power of deep polynomial neural networks (2019). arXiv:1905.1 2207. URLhttps://arxiv.org/abs/1905.12207

    Pith/arXiv arXiv 2019

  31. [31]

    Zhang, G

    L. Zhang, G. Naitzat, L.-H. Lim, Tropical geometry of deep neural networks (2018).arXiv:1805.07091. URLhttps://arxiv.org/abs/1805.07091

    Pith/arXiv arXiv 2018

  32. [32]

    Brandenburg, G

    M.-C. Brandenburg, G. Loho, G. Montúfar, The real tropical geometry of neural networks (2024). arXiv: 2403.11871. URLhttps://arxiv.org/abs/2403.11871

    Pith/arXiv arXiv 2024

  33. [33]

    Grothendieck, Le groupe de Brauer

    A. Grothendieck, Le groupe de Brauer. I, II, III, North- Holland, 1968

    1968

  34. [34]

    Edidin, B

    D. Edidin, B. Hassett, A. Kresch, A. Vistoli, Brauer groups and quotient stacks, American Journal of Mathematics 123 (4) (2001) 761–777

    2001

  35. [35]

    Colliot-Thélène, A

    J.-L. Colliot-Thélène, A. N. Skorobogatov, The Brauer– Grothendieck Group, Springer, 2021

    2021

  36. [36]

    URLhttps://stacks.math.columbia.edu 34

    The Stacks Project Authors, Stacks project (2026). URLhttps://stacks.math.columbia.edu 34

    2026