BPS spectroscopy with reinforcement learning

Asa Gauntlett; Federico Carta; Finley Griffin; Yang-Hui He

REVIEW 5 major objections 6 minor 1 cited by

BPS spectroscopy with reinforcement learning

T0 review · 5 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Reinforcement learning finds every minimal BPS chamber of the SU(2) Nf=4 theory.

desk verdict Solid RL application to BPS quiver mutation with genuinely new chamber counts, but the 'all minimal chambers' claim needs an independent check. read the letter →

arxiv 2501.14863 v2 pith:KHAIWD45 submitted 2025-01-24 hep-th

Federico Carta , Asa Gauntlett , Finley Griffin , Yang-Hui He This is my paper · ORCID

classification hep-th

keywords BPSspectrumquivermutationreinforcementlearningN=2supersymmetryminimalchambersCoulombbranchSU(2)Nf=4exceptionalquivers

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper is trying to establish that BPS spectroscopy, finding exactly which particles appear at a point on the Coulomb branch of a 4d $\mathcal{N}=2$ gauge theory, can be automated with reinforcement learning instead of brute-force mutation scans. Starting from a BPS quiver, an agent learns a sequence of quiver mutations that reaches the antiparticle quiver, and the charges recorded along the way give the full BPS spectrum. The same agent, with penalties for revisiting known chambers, is then run to enumerate all minimal chambers of a theory. The paper reports counts of inequivalent minimal chambers for SU(2) Nf=4, namely 312, 576, 1656, and 1776 for its four quivers, and identifies minimal chambers for the elliptic E-type and Derksen-Owen exceptional theories. A sympathetic reader would care because this turns a previously intractable classification task into a routine computation for complete N=2 theories.

What carries the argument

The mechanism is the BPS quiver together with the mutation method: each node of the quiver is a basis hypermultiplet, arrows encode electric-magnetic pairings, and a sequence of mutations ending at the antiparticle quiver reveals the full spectrum. The reinforcement-learning layer is a proximal-policy-optimization agent whose state is the quiver adjacency matrix alongside a Gamma matrix of charge labels; its action space is the set of valid mutation nodes after dynamic masking; and its reward, built from structural and label fitness functions $f_M$ and $f_\Gamma$, measures closeness to the antiparticle quiver and gives a terminal reward when it is reached. The discount factor in the value function encourages short sequences, which is what biases the agent toward minimal chambers. This setup turns BPS spectroscopy into a path-finding problem in the quiver mutation graph.

What would settle it

Run an independent, exhaustive enumeration of mutation sequences for each of the four SU(2) Nf=4 BPS quivers, with a certified bound on the size of each mutation class, and compare the number of inequivalent minimal chambers with the reported 312, 576, 1656, and 1776; one extra chamber would refute the completeness claim.

Watch

Extended reading notes

Core claim

The central claim is that the full BPS spectrum of a complete four-dimensional $\mathcal{N}=2$ quantum field theory in a finite chamber can be obtained automatically by training a reinforcement-learning policy to generate quiver mutations, and that the same policy, run with exploration penalties, recovers all minimal chambers. The paper demonstrates this by reconstructing the known twelve-state minimal chamber of SU(2) Nf=4 and then enumerating minimal chambers for the four BPS quivers of that theory, obtaining 312, 576, 1656, and 1776 inequivalent chambers up to quiver automorphisms. It also reports minimal chambers for the exceptional elliptic E-type quivers and for the Derksen-Owen quiver X6, and confirms that X7 admits no finite chamber. The discovery is both a method and a set of concrete data: mutation sequences and chamber lists that were previously out of reach of exhaustive computer scans.

Load-bearing premise

The load-bearing premise is that stopping the search after about a million time-steps with no new minimal chamber actually means the list is complete; the paper supplies no independent upper bound that the explored mutation graph was fully covered.

Editorial extensions

If this is right

The BPS spectrum at any finite chamber of a complete N=2 theory becomes a routine RL task: run the agent, read the mutation sequence, read the charges.
The reported mutation sequences and chamber lists can be fed into index computations such as the Schur index, giving a direct route from quiver data to protected observables.
The four-count census of minimal chambers for SU(2) Nf=4 gives a target list for wall-crossing and quantum-dilogarithm identities, since every chamber should correspond to a mutation-path identity.
The method should extend to non-minimal or maximal finite chambers by changing the reward, potentially yielding a full chamber classification.
For exceptional theories, the discovered minimal chambers and the confirmed absence for X7 fix a concrete data point for the classification of complete N=2 theories.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: the same policy trained on one quiver may transfer to mutation-equivalent quivers, so one agent could certify spectra across all chambers of a theory rather than being retrained per chamber.
Editorial inference: the reward design could be adapted to estimate the number of finite chambers, not just minimal ones, by assigning rewards to new finite chambers discovered at any length.
Editorial inference: a formal proof of completeness might come from relating the RL-explored state space to known mutation-class invariants, for example cluster-algebra mutation-type classifications, which would upgrade the census from heuristic to theorem.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Bottom line: this is a useful and mostly believable application of PPO to BPS quiver mutation, with new minimal-chamber counts for SU(2) Nf=4. The main caveat is that the headline completeness claim rests on a heuristic stopping rule, and there is a caption/text inconsistency about X6/X7 that needs fixing.

The paper does something real. It frames quiver mutation as a Markov decision process, uses PPO with dynamic action masking, and shows the agent reliably reaches the antiparticle quiver on random cyclic quivers and on the SU(2) Nf=4 theory. The 12-state chamber they find matches ref. [23], and the code is on GitHub, so the core algorithm is checkable. The genuinely new data is the minimal-chamber census: 312, 576, 1656, 1776 up to symmetry for the four SU(2) Nf=4 quivers. As far as I can tell, those counts are not in the literature, and they are exactly the kind of concrete output people in N=2 classification would use.

The soft spots are real but local. The phrase "all minimal chambers" is stronger than the evidence. Section 4.1 stops when no new chambers appear in O(10^6) steps; that is saturation, not exhaustion. For 6-node quivers with finite mutation classes, an independent deterministic enumeration should be feasible, and the paper should either do that or explicitly frame the counts as "found by this search." Figure 2 claims orders of magnitude speedup over a smart walker, but there are no error bars, and the 9- and 10-node SW points are projections, so that number is softer than the text implies. And there is a flat inconsistency: the main text says X7 has no finite chamber, while the Fig. 5 caption says X6. The table itself says X7, so I suspect a caption typo, but it must be corrected.

The math itself looks sound: mutation rules are standard, the reward function is a sensible proximity measure to the antiparticle quiver, and there is no circularity—the output is a mutation path, not a fitted spectrum. The hyperparameters (λ, α, γ, network size, stopping threshold) are free, and a sensitivity analysis would make the "minimal chamber" preference claim more convincing.

Who is this for? Anyone working on BPS quiver spectra, wall-crossing, or N=2 classification. It is not a landmark, but it is a useful tool with concrete data. I would send it to a serious referee, and I would ask that referee to push on the completeness claim and the missing error bars. If the authors can turn the census into a verified enumeration, the paper becomes quite solid.

Referee Report

5 major / 6 minor

Summary. The paper applies Proximal Policy Optimisation (PPO), a model-free reinforcement learning method, to the quiver-mutation problem of BPS spectroscopy. For a BPS quiver of a complete 4d N=2 QFT, the agent's goal is to find a sequence of quiver mutations that transforms the initial quiver into its antiparticle quiver; the charges encountered along the path constitute the BPS spectrum of the corresponding finite chamber, and rewards are based on a fitness measuring proximity to the antiparticle quiver (§3, Eq. (8)). The authors test the agent on random cyclic quivers of 3–10 nodes and report convergence orders of magnitude faster than a smart-walk baseline (§4, Fig. 2). The main applications are: a claimed complete census of minimal chambers for the four SU(2) Nf=4 BPS quivers, reported as 312, 576, 1,656 and 1,776 essentially-inequivalent chambers (§4.1, Fig. 3); and single minimal chambers for the exceptional quivers Ê6, Ê7, Ê8 and X6, together with the statement that X7 has no finite chamber (§4.2, Fig. 5). Code and chamber lists are provided on GitHub.

Significance. If the census is complete and the chamber lists are correct, this is the first complete enumeration of minimal chambers for SU(2) Nf=4, a concrete reference set for wall-crossing and Schur-index computations, and a convincing demonstration that RL can steer the mutation method for quivers whose brute-force scanning grows as (number of nodes)^(sequence length). I credit the paper for using the standard mutation rules, for defining the terminal state (antiparticle quiver) entirely from the initial quiver so the reward is not circular, for reproducing the known 12-state chamber of SU(2) Nf=4 (Example 2), and for making code and chamber lists publicly available, which makes every reported chamber independently checkable (e.g., with Keller's applet). The main risk is the completeness claim for the census, which currently rests on an empirical stopping rule rather than an exhaustion argument, and the internal contradiction concerning which Derksen-Owen quiver lacks a finite chamber. Both issues are concrete and fixable within the paper's scope.

major comments (5)

[§4.1, Fig. 4] The headline 'all minimal chambers' for the four SU(2) Nf=4 quivers rests on the stopping rule 'we run the searches until no new chambers are found in O(10^6) time-steps.' This is an empirical saturation criterion, not an exhaustion argument: a stochastic search with a negative penalty on known chambers can miss rare chambers even after long plateaus, and no bound on the relevant state space is given. Because Q1–Q4 are six-node quivers of finite mutation type (triangulations of the four-punctured sphere), an independent deterministic closure of the mutation classes is feasible and should be supplied to certify the numbers 312, 576, 1,656 and 1,776. Additionally, the paper never defines 'minimal chamber' and does not describe any procedure that certifies each counted chamber as minimal rather than merely finite.
[§3, §4.1] There is a quantitative tension between the action space defined in §3 and the reported chamber counts. §3 declares illegal any node 'that has already been mutated on in the action history', so from a fixed six-node initial quiver every successful episode is a permutation of the six nodes, giving at most 6! = 720 distinct mutation sequences and hence at most 720 distinct chambers; the same bound holds across all episodes because the set of possible sequences is the same for every episode. The reported 1,656 and 1,776 chambers for Q3 and Q4 exceed this bound. Please clarify whether the no-repeat mask is enforced in the §4.1 searches, whether episodes restart from quivers other than the initial one, and how exactly a 'chamber' is defined and counted; if the mask is relaxed, the 'any finite chamber' claim from the Introduction should be re-examined as well.
[§4.2, Fig. 5] The manuscript contains a direct internal contradiction about which Derksen-Owen quiver has no finite chamber. The main text states 'the X7 quiver has no finite chamber', the Fig. 5 table lists a 10-state minimal chamber for X6 and 'No finite chamber' for X7, but the Fig. 5 caption states 'we find that the X6 Derksen-Owen quiver has no finite chamber.' These statements cannot all be true; the caption, table and text must be brought into agreement and checked against reference [4]. Relatedly, the evidentiary basis of the negative claim should be made explicit: failure of an RL search to reach the antiparticle quiver is not a proof of absence of a finite chamber, so either the result should be attributed to the literature or an exhaustive-search argument must be provided.
[§4, Fig. 2] The quantitative claim that the RL model converges 'various orders of magnitude faster' than the smart walker is presented without the statistics needed to support it. Fig. 2 shows single curves with no error bars; the text does not report the number of quivers per node count, the number of trials or seeds per quiver, or the per-trial success rate; and the 9- and 10-node smart-walk values are extrapolated ('projected') rather than measured. Please report means, spreads, trial counts and runtimes, define precisely what is being averaged ('convergence rate'), and label the extrapolated points clearly so that the speedup comparison is reproducible.
[§3, Eq. (8), Appendix A] The core method is under-specified in the manuscript. The fitness functions fM and fΓ entering the reward R(s) in Eq. (8) are never defined; the negative penalty applied upon visiting a known chamber in §4.1 is not quantified; and the ε schedule for the ε-greedy exploration described in Appendix A.2 is not given. Without these definitions the algorithm cannot be reproduced from the paper even with the stated architecture and hyperparameters (α, γ, layer sizes). Please provide the missing formulas or pseudocode, and state the values used for λ, ε and the penalty.

minor comments (6)

[References] Reference [27] lists the arXiv identifier '9703166', which appears malformed for a paper published in 2016; please verify and correct it.
[References] Reference [28] gives the journal volume as '580 (580)'; the duplicated volume number should be corrected.
[§4.1] The notation 'O(10^6 ∼ 10^7)' should be written as 'O(10^6)–O(10^7)'.
[§4.1] In the sentence 'the four possible triangulations of the fourth-punctured sphere', 'fourth-punctured' should be 'four-punctured'.
[Abstract, §1] The statement that a complete census was 'previously intractable with computer scanning' is overstated for the six-node SU(2) Nf=4 quivers, whose finite mutation classes can be inspected with Keller's applet; the relevant point is that the complete census was absent in the literature, not that any single scan was intractable.
[§4.2, Fig. 5] Please state explicitly which of the displayed chambers for Ê6, Ê7, Ê8 and X6 are new and which reproduce results already contained in reference [4], so that the abstract's 'discover new minimal chambers' is substantiated.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the RL reward targets the antiparticle quiver defined a priori from the initial quiver; BPS spectra are computed mutation paths checked against external benchmarks, with the §4.1 saturation stopping rule and the X6/X7 caption discrepancy being completeness/consistency concerns, not circular steps.

full rationale

The derivation chain is self-contained and non-circular. The RL reward R(s) = 1 − λ fM(s) − (1 − λ) fΓ(s) measures proximity to the antiparticle quiver, which is defined purely from the initial quiver ('the final quiver produced will have all charges flipped compared to the original quiver and all arrows inverted'); the target is not defined in terms of the output spectrum. The 'full BPS spectrum' is read off as 'all the charges that appear at all nodes in every intermediate steps' of the mutation path actually computed under the external mutation rules of [23], so the output is a computed path, not a fitted input. Example 2's 12-state chamber is checked against the independent result of [23], and the exceptional-theory findings are checked against [4], so the load-bearing claims rest on external literature, not self-citation; the few self-citations present ([8], [15], [25]) are contextual ML background and carry no argumentative weight. The genuine weaknesses are non-circular: (i) the completeness of the SU(2) Nf=4 census (312, 576, 1656, 1776) rests only on the §4.1 heuristic 'we run the searches until no new chambers are found in O(10^6) time-steps', an empirical saturation criterion with no exhaustion proof or independent enumeration; (ii) the 'no finite chamber' status for X7 (main text) versus X6 (Fig. 5 caption) is internally inconsistent, and the negative finding is inferred from RL/SW failure to find a chamber rather than a classified search; these are correctness/completeness risks, not reductions-by-construction, since no model parameter encodes the reported chamber counts or spectra.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The RL computation is built on the standard BPS quiver mutation framework and on several hand-chosen algorithmic parameters and stopping heuristics. No new physical entities are introduced, so the ledger lists the domain assumptions and free parameters the results depend on.

free parameters (6)

Reward mixing parameter lambda = lambda = 0.1 (Fig. 1); no sensitivity analysis
Chosen by hand in Eq. (8); balances structural vs label fitness.
PPO learning rate alpha = 0.5
Chosen by hand; controls policy update speed.
PPO discount factor gamma = 0.995
Chosen by hand; encourages shorter mutation sequences.
Neural network hidden layer sizes = 252, 504, 252
Chosen by hand; no ablation study is reported.
Stopping threshold for chamber census = O(10^6) timesteps with no new chambers
Heuristic completeness criterion in Section 4.1; the claimed 'all chambers' counts depend on this threshold.
Epsilon-greedy exploration schedule = not stated
Appendix A.2 introduces epsilon-greedy exploration but does not report the schedule; it affects exploration and results.

assumptions (4)

domain assumption The mutation method of [23] determines the full BPS spectrum and termination at the antiparticle quiver signals a finite chamber.
Invoked throughout Section 2.2 and used as the terminal condition in the RL reward; it is a theorem from the physics literature, not proved here.
domain assumption Complete N=2 theories [3,4] have BPS quivers with a fixed number of nodes equal to the rank of the charge lattice.
Used to define the state space and the n x 2n matrix representation in Section 3; provided by the completeness classification.
ad hoc to paper PPO with dynamic action masking converges to an effective policy for the mutation search problem.
No convergence guarantee is given; the paper relies on empirical results on random quivers in Section 4.
ad hoc to paper Running the search until no new chambers are seen for about 10^6 steps is sufficient to conclude all minimal chambers have been found.
Used in Section 4.1 to justify the claimed complete census for SU(2) Nf=4; this is a heuristic stopping criterion, not a proof.

how reviews work

0 comments

Cite this review

Pith. "Pith review of BPS spectroscopy with reinforcement learning." pith.science (2026). https://pith.science/paper/KHAIWD45

@misc{pith2026250114863,
  author       = {Pith},
  title        = {Pith review of: BPS spectroscopy with reinforcement learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KHAIWD45}},
  note         = {Machine review of arXiv:2501.14863}
}

abstract

We apply reinforcement learning (RL) to establish whether at a given position in the Coulomb branch of the moduli space of a 4d $\mathcal{N} = 2$ quantum field theory (QFT) the BPS spectrum is finite. If it is, we furthermore determine the full BPS spectrum at such point in moduli space. We demonstrate that using a RL model one can efficiently determine the suitable sequence of quiver mutations of the BPS quiver that will generate the full BPS spectrum. We analyse the performance of the RL model on random BPS quivers and show that it converges to a solution various orders of magnitude faster than a systematic brute-force scan. As a result, we show that our algorithm can be used to identify all minimal chambers of a given $\mathcal{N}=2$ QFT, a task previously intractable with computer scanning. As an example, we recover all minimal chambers of the $\text{SU}(2)$ $N_f = 4$ gauge theory, and discover new minimal chambers for theories that can be realized by IIB geometric engineering.

Figures

Figures reproduced from arXiv: 2501.14863 by the authors.

**Figure 1.** Cumulative reward (λ = 0.1) against mutation for a successful training episode on the BPS quiver Q1 of the SU(2) Nf = 4 SCFT, from [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Plot of of the average convergence rate to a finite ch [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 4.** Plot of the total number of distinct minimal chambe [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

Figures from the paper (2 more)

**Figure 3.** Figure 3: The four BPS quivers of the SU(2) Nf = 4 SCFT and the number of minimal chambers identified by the RL model. We note that for each finite chamber, there exists a set of equivalent chambers obtained from the symmetries of the quiver. One such symmetry is permutation of …

**Figure 5.** Figure 5: The 5 exceptional complete N = 2 theories and their BPS quivers, and their minimal chamber as identified by the RL model. In agreement with the literature, we find that the X6 Derksen-Owen quiver has no finite chamber. • The three theories identified with the elliptic …

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Learning to Trace Seiberg Dualities
hep-th 2026-07 accept novelty 6.0 of 10

Hybrid graph-transformer networks guiding A* and beam search find Seiberg-duality paths between ~10-node quivers more efficiently than BFS or pure physics heuristics, with a measured complexity breaking point.

Reference graph

Works this paper leans on

30 extracted references · 12 canonical work pages · cited by 1 Pith paper

[23]

M. Alim, S. Cecotti, C. Cordova, S. Espahbodi, A. Rastog i, C. V afa,N = 2 quantum ﬁeld theories and their BPS quivers, Adv. Theor. Ma th. Phys. 18 (1) (2014) 27–127. arXiv:1112.3984

arXiv 2014
[4]

M. Alim, S. Cecotti, C. Cordova, S. Espahbodi, A. Rastogi , C. V afa, BPS Quivers and Spectra of Complete N =2 Quantum Field Theories, Com- mun. Math. Phys. 323 (2013) 1185–1227. arXiv:1109.4941

arXiv 2013
[1]

Seiberg, E

N. Seiberg, E. Witten, Electric - magnetic duality, mono pole condensa- tion, and conﬁnement in N = 2 supersymmetric Y ang-Mills theory, Nucl. Phys. B 426 (1994) 19–52, [Erratum: Nucl.Phys.B 430, 485–48 6 (1994)]. arXiv:hep-th/9407087

arXiv 1994
[2]

Seiberg, E

N. Seiberg, E. Witten, Monopoles, duality and chiral sym metry break- ing in N = 2 supersymmetric QCD, Nucl. Phys. B 431 (1994). arXiv:hep-th/9408099

arXiv 1994
[3]

Cecotti, C

S. Cecotti, C. V afa, Classiﬁcation of complete N =2 supersymmetric the- ories in 4 dimensions (3 2011). arXiv:1103.5832

arXiv 2011
[5]

Gaiotto, N =2 dualities, JHEP 08 (2012) 034

D. Gaiotto, N =2 dualities, JHEP 08 (2012) 034. arXiv:0904.2715

arXiv 2012
[6]

New Graphs of Finite Mutation Type

H. Derksen, T. Owen, New Graphs of Finite Mutation Type, T he Elec- tronic Journal of Combinatorics 15 (2008). arXiv:0804.0787

work page Pith review arXiv 2008
[7]

H. A. Kissinger, E. Schmidt, D. Huttenlocher, The age of A I: and our human future, Hachette UK, 2021

work page 2021

Show all 30 references

[8]

He, Deep-Learning the Landscape, Phys

Y .-H. He, Deep-Learning the Landscape, Phys. Lett. B 774 (2017) 564–

2017
[9]

Kreﬂ, R.-K

D. Kreﬂ, R.-K. Seong, Machine Learning of Calabi-Y au V ol umes, Phys. Rev. D 96 (6) (2017) 066014. arXiv:1706.03346

2017 arXiv
[10]

Cariﬁo, J

J. Cariﬁo, J. Halverson, D. Krioukov, B. D. Nelson, Mach ine Learning in the String Landscape, JHEP 09 (2017) 157. arXiv:1707.00655

2017 arXiv
[11]

Ruehle, Evolving neural networks with genetic algor ithms to study the String Landscape, JHEP 08 (2017) 038

F. Ruehle, Evolving neural networks with genetic algor ithms to study the String Landscape, JHEP 08 (2017) 038. arXiv:1706.07024

2017 arXiv
[12]

He, The Calabi–Y au Landscape: From Geometry, to P hysics, to Machine Learning, Lecture Notes in Mathematics, 2021

Y .-H. He, The Calabi–Y au Landscape: From Geometry, to P hysics, to Machine Learning, Lecture Notes in Mathematics, 2021. arXiv:1812.02893

2021 arXiv
[13]

Ruehle, Data science applications to string theory, Phys

F. Ruehle, Data science applications to string theory, Phys. Rept. 839 (2020) 1–117

2020
[14]

Gukov, J

S. Gukov, J. Halverson, F. Ruehle, Rigor with machine le arning from ﬁeld theory to the poincar´ e conjecture, Nature Reviews Physics(2024) 1–10

2024
[15]

He, AI-driven research in pure mathematics and th eoretical physics, Nature Rev

Y .-H. He, AI-driven research in pure mathematics and th eoretical physics, Nature Rev. Phys. 6 (9) (2024) 546–553. arXiv:2405.19973

2024 arXiv
[16]

Chen, Y .-H

H.-Y . Chen, Y .-H. He, S. Lal, M. Z. Zaz, Machine Learning Etudes in Conformal Field Theories (6 2020). arXiv:2006.16114

2020 arXiv
[17]

Kantor, V

G. Kantor, V . Niarchos, C. Papageorgakis, Conformal bo otstrap with reinforcement learning, Physical Review D 105 (2022). arXiv:2108.09330

2022 arXiv
[18]

J. Bao, S. Franco, Y .-H. He, E. Hirst, G. Musiker, Y . Xiao , Quiver muta- tions, seiberg duality, and machine learning, Physical Rev iew D 102 (8) (2020). arXiv:2006.10783

2020 arXiv
[19]

Cheung, P .-P

M.-W. Cheung, P .-P . Dechant, Y .-H. He, E. Heyes, E. Hirs t, J.-R. Li, Clustering cluster algebras with clusters, Adv. Theor. Mat h. Phys. 27 (3) (2023) 797–828. arXiv:2212.09771

2023
[20]

J. He, H. Jenne, H. Chau, D. Brown, M. Raugas, S. C. Billey , H. Kvinge, Machines and mathematical mutations: Using GNNs to charact erize quiver mutation classes, in: The 4th Workshop on Mathematic al Rea- soning and AI at NeurIPS’24, 2024

2024
[21]

Gukov, R.-K

S. Gukov, R.-K. Seong, Machine learning BPS spectra and the gap con- jecture, Phys. Rev. D 110 (4) (2024) 046016. arXiv:2405.09993

2024 arXiv
[22]

C´ ordova, S.-H

C. C´ ordova, S.-H. Shao, Schur indices, bps particles, and argyres- douglas theories, Journal of High Energy Physics 2016 (1) (2 016). arXiv:1506.00265

2016 arXiv
[24]

Fomin, A

S. Fomin, A. Zelevinsky, Cluster algebras i: foundatio ns, Journal of the American mathematical society 15 (2) (2002) 497–529

2002
[25]

B. Feng, A. Hanany, Y .-H. He, D-brane gauge theories fro m toric singularities and toric duality, Nucl. Phys. B 595 (2001) 16 5–200. arXiv:hep-th/0003085

2001 arXiv
[26]

Tang, C.-H

C.-Y . Tang, C.-H. Liu, W.-K. Chen, S. D. Y ou, Implementi ng action mask in proximal policy optimization (PPO) algorithm, ICT Express 6 (3) (2020) 200–203

2020
[27]

Xie, BPS spectrum, wall crossing and quantum dilogar ithm identity, Adv

D. Xie, BPS spectrum, wall crossing and quantum dilogar ithm identity, Adv. Theor. Math. Phys. 20 (3) (2016) 405–524. arXiv:9703166

2016
[28]

Brustle, G

T. Brustle, G. Dupont, M. Perotin, On Maximal Green Sequ ences, Journal of Algebra 580 (580) 399–422. arXiv:1205.2050

arXiv
[29]

Keller, Quiver Mutation in Java, https://webusers.imj-prg.fr/~ bernhard.keller/quivermutation/ (2006)

B. Keller, Quiver Mutation in Java, https://webusers.imj-prg.fr/~ bernhard.keller/quivermutation/ (2006)

2006
[30]

Saito, Extended a ﬃne root systems I (Coxeter transformations), Pub- lications of the Research Institute for Mathematical Scien ces 21 (1) 75– 179

K. Saito, Extended a ﬃne root systems I (Coxeter transformations), Pub- lications of the Research Institute for Mathematical Scien ces 21 (1) 75– 179. 8

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [23]

M. Alim, S. Cecotti, C. Cordova, S. Espahbodi, A. Rastog i, C. V afa,N = 2 quantum ﬁeld theories and their BPS quivers, Adv. Theor. Ma th. Phys. 18 (1) (2014) 27–127. arXiv:1112.3984

arXiv 2014

[2] [4]

M. Alim, S. Cecotti, C. Cordova, S. Espahbodi, A. Rastogi , C. V afa, BPS Quivers and Spectra of Complete N =2 Quantum Field Theories, Com- mun. Math. Phys. 323 (2013) 1185–1227. arXiv:1109.4941

arXiv 2013

[3] [1]

Seiberg, E

N. Seiberg, E. Witten, Electric - magnetic duality, mono pole condensa- tion, and conﬁnement in N = 2 supersymmetric Y ang-Mills theory, Nucl. Phys. B 426 (1994) 19–52, [Erratum: Nucl.Phys.B 430, 485–48 6 (1994)]. arXiv:hep-th/9407087

arXiv 1994

[4] [2]

Seiberg, E

N. Seiberg, E. Witten, Monopoles, duality and chiral sym metry break- ing in N = 2 supersymmetric QCD, Nucl. Phys. B 431 (1994). arXiv:hep-th/9408099

arXiv 1994

[5] [3]

Cecotti, C

S. Cecotti, C. V afa, Classiﬁcation of complete N =2 supersymmetric the- ories in 4 dimensions (3 2011). arXiv:1103.5832

arXiv 2011

[6] [5]

Gaiotto, N =2 dualities, JHEP 08 (2012) 034

D. Gaiotto, N =2 dualities, JHEP 08 (2012) 034. arXiv:0904.2715

arXiv 2012

[7] [6]

New Graphs of Finite Mutation Type

H. Derksen, T. Owen, New Graphs of Finite Mutation Type, T he Elec- tronic Journal of Combinatorics 15 (2008). arXiv:0804.0787

work page Pith review arXiv 2008

[8] [7]

H. A. Kissinger, E. Schmidt, D. Huttenlocher, The age of A I: and our human future, Hachette UK, 2021

work page 2021