Pith. sign in

REVIEW 4 major objections 5 minor 76 references

Exploring Structures in Physics Problems: Can AI Agents Discover Statistical Mechanical Mappings?

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read LLM-based agents can recover exact finite-size partition functions under numerical feedback, yet can pass every numeric check while naming the wrong solvability class — so numeric agreement alone cannot certify a statistical-mechanical mapp

desk verdict Core observation is real and supported by the table itself, but the paper's quantitative scaffolding is too thin to trust: single trials, a buggy complexity audit, and no released code make the rankings and honesty scores unverifiable. read the letter →

arxiv 2607.26367 v1 pith:4JFFLN5N submitted 2026-07-29 cs.AI

Wanyu Zhao , Wanbing Zhao This is my paper
classification cs.AI
keywords LLMagentsscientificdiscoverystatisticalmechanicspartitionfunctionexactlysolvablemodelstransfermatrixPfaffianverification
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

This paper asks whether LLM-based agents can rediscover statistical-mechanical mappings: given a raw partition function, can an agent transform it into a known tractable representation? To test this, it builds a six-problem benchmark of Ising-type partition functions, covering transfer-matrix methods, gauge-removable disorder, planar/Pfaffian structure, and a negative control, and runs a simple propose-verify-revise agent across eight LLM configurations. The main finding is two-sided: numerical feedback from a brute-force verifier often repairs code bugs and leads agents to compute correct finite-size partition functions, but agents sometimes pass all numerical probes while mislabeling the underlying tractable class or claiming a lower computational complexity than their code actually has. The upshot is a design argument: for AI discovery agents in theoretical physics, verification must include symbolic and structural checks, not just number matching.

What carries the argument

The load-bearing object is the propose-verify-revise agent loop: a proposer that classifies a raw partition function into one of nine tractable classes and writes Python code for Z_efficient; a brute-force verifier that checks the candidate against exact enumeration on small instances; and a feedback string that reports failures without revealing Z_brute. The verifier's pass condition is purely numerical (relative error below 1e-6), while a separate post-hoc audit checks whether the claimed complexity label matches the submitted code. The benchmark's six tasks instantiate the tractability library — transfer_matrix, gauge, pfaffian, no_efficient_known — so that a correct discovery must name t

What would settle it

Take one of the reported passing solutions that carried a wrong tractable-class label and run a symbolic verifier on it — for example, check whether the submitted code actually implements a valid Pfaffian reduction for the planar grid. If that solution truly passes all numerical probes while the symbolic check fails, the paper's central claim is confirmed; if no such solution exists across the benchmark, the claim is undercut.

Watch

Extended reading notes

Core claim

The paper's central claim is that a feedback loop driven only by numerical equality is sufficient to recover correct partition-function values on textbook Ising tasks, but insufficient to certify that an agent has discovered the right mapping. Across the six tasks, most agents reach 100% numerical pass rates within the five-round loop, often by round one; the verifier successfully drives repairs such as swapping a broken Pfaffian attempt for a row-transfer implementation. Yet the same agents can submit an exponential row-transfer computation under a polynomial label like 'Pfaffian / O(n^3)', and the numeric verifier will report success because all probed values match brute force. The paper r

Load-bearing premise

Every reported success rate and model ranking rests on a single trial per model-variant-task; since LLM sampling is stochastic and the benchmark has only six tasks, a single re-run could shift the percentages by about 17 points, so the ranking and failure-mode frequencies are not statistically grounded.

Editorial extensions

If this is right

  • With numerical feedback, LLM agents can repair sign and implementation errors and compute exact finite-size partition functions on the six textbook Ising tasks, so verifier loops are a viable repair mechanism.
  • Because agents can pass every numeric probe while mislabeling the tractable class or understating complexity, any mapping benchmark that scores only final numerical equality will overstate discovery ability.
  • The refinement loop is effective for numerical repair but nearly blind to class-only errors: once a solution is numerically right with the wrong label, the existing feedback gives no signal to correct it.
  • Paraphrased problem descriptions change behavior: reasoning-oriented models are largely paraphrase-invariant, while some models lose classification accuracy under paraphrase; paraphrase can also improve zero-shot numerical success by sidestepping memorized but unreliable labels.
  • The negative control confirms the intended behavior: given a random 3-regular Ising glass with no known efficient method in the library, agents honestly report no_efficient_known and fall back to brute force rather than invent a tractable class.

Reading between the lines

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

  • A natural next test is to score agents on the explanation rather than the number: requiring a symbolic derivation (transfer-matrix elements, gauge transformation, or a valid Pfaffian construction) before numeric checks would directly penalize the class-only errors the paper found.
  • The paper's tier taxonomy implies that the real bar for AI discovery lies in Tier 2/3 problems — quantum-code thresholds and measurement-induced transitions — where the mapping is not in the training distribution; Tier 1 success should be read as retrieval plus repair, not discovery.
  • One cheap design fix the paper leaves implicit: feed the complexity audit's mismatch back into the loop so that a numerically passing solution with an overstated label triggers revision; the current feedback string only reports value mismatches.
  • If the numeric verifier were extended to MCMC or tensor-network oracles at larger n, the same benchmark could test whether agents find mappings that generalize beyond the probed sizes — a stricter criterion than matching at n up to 16.
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.

Referee Report

4 major / 5 minor

Summary. This paper investigates whether LLM-based agents can discover statistical-mechanical mappings for partition functions. It introduces StatMechBench-v0, a six-task benchmark of Ising-type models (1D transfer matrix, 2D torus transfer matrix, planar spin glass/Pfaffian, Mattis gauge, and a random 3-regular glass negative control), and a propose-verify-revise agent that classifies each task into a tractable class and provides an executable Z_efficient function checked against brute force. Eight LLM configurations are tested on canonical and paraphrased problem descriptions. The main reported findings are (i) numerical verifier feedback repairs many implementation errors, (ii) some agents pass all probed numerical checks while assigning an incorrect tractable class or understating computational complexity, and (iii) numerical verification alone is therefore insufficient; future agents should use symbolic and structural verification. The paper also proposes a three-tier taxonomy and future design directions.

Significance. If the findings are robust, the paper makes a useful early contribution: it defines a concrete benchmark for a mathematically well-specified discovery task and provides a documented agent/verifier setup that future work can build on. The central design implication—finite numerical agreement cannot certify tractability class or complexity—is logically sound and important, and the paper's appendices give enough detail to reproduce the prompt templates and verifier. The paper also deserves credit for explicitly labeling the negative control and for acknowledging a false positive in its honesty metric. However, the empirical support is not yet solid: the evaluation runs a single trial per condition, the complexity-honesty metric has a known regex failure and is not fully specified, and the concrete failure-mode examples are not released in a form that can be inspected. These issues are load-bearing for the paper's empirical claims, not mere presentation.

major comments (4)
  1. [§3.3, Table 3, Fig. 2] All quantitative results are based on one run per model-variant-task, with no random seeds or confidence intervals. With only six tasks, a single task flip changes every percentage by about 17 points, so statements such as "Opus 4.7 performs best" and "three models remain honest" are not statistically grounded. Since LLM outputs are stochastic, a single trial cannot represent a model's performance. Please run multiple independent trials (e.g., different temperatures/seeds) and report confidence intervals or at least per-task success matrices; otherwise the numerical comparisons and failure-mode frequencies in §3.3 cannot be assessed.
  2. [Table 3 footnote and §3.3] The complexity-honesty score O is computed by an unspecified regex audit that the footnote itself shows to be wrong: complexity_O = O(n^2 2^{√n}) is algebraically equivalent to the gold O(L^2 2^L) with n=L^2, yet the audit penalized it because the regex anchors on the literal character 'L'. This false positive means the O scores, the "three models remain honest" claim, and the reported rate of complexity understating are unverified. The audit is not described in App. C; a reproducible O metric is needed, ideally a semantic comparison of the claimed complexity with the actual algorithm's scaling, plus human adjudication of ambiguous cases.
  3. [§3.3 and App. C] The central existence claim—that agents can pass the brute-force verifier while misidentifying the tractable class or understating complexity—is supported only by summary sentences ("exponential row-transfer constructions under polynomial-time labels such as pfaffian/O(n^3)") and a post-hoc audit that is unreliable. No round-by-round logs, candidate code, or final JSON outputs are released, so the specific cases cannot be inspected. App. C describes the verifier as a numerical check on a cooperative proposer and explicitly does not audit complexity. Please release the full trajectories (or at least one complete example per failure mode) and the O-audit implementation; without them, the existence claim is not independently verifiable.
  4. [Table 2 note and §3.2] The benchmark assigns exactly one gold tractable class per task despite the authors' acknowledgment that multiple mechanisms can apply. For example, P03 (2D Ising on a torus) is labeled transfer_matrix, but a free-fermion/Jordan–Wigner representation is also a valid exact solution; P04 (planar Ising) is labeled pfaffian, but a row-transfer construction also computes the partition function at the probed sizes. TC0 therefore counts valid alternative mappings as errors, and the conclusion that agents "misidentify the underlying tractable class" may conflate genuine class-only errors with benchmark label choices. Please either provide multi-label gold annotations or an expert-validated argument that the chosen label is the uniquely intended one for each task.
minor comments (5)
  1. [App. B vs §3.3] App. B's fixed rewrite rule conflicts with §3.3's statement that both canonical and paraphrased descriptions are "generated using LLMs". Please clarify how the descriptions were actually produced.
  2. [Fig. 2] The legend uses "DeepSeek-Reasoner" while Table 3 uses "DeepSeek V4-FlashThink". Use consistent model names across the paper.
  3. [Table 3 footnote] The notation O(n2√n) is ambiguous; please format as O(n^2 2^{√n}) or similar.
  4. [App. B, P06] The canonical P06 task card contains an explicit instruction to consider no_efficient_known ("If you determine, after analysing... state so explicitly"), which may prime the expected negative label. This should be disclosed or removed if P06 is meant as an unbiased negative control.
  5. [§3.3] The metric label "O" is undefined before first use. Define it or spell out "complexity honesty" in the text.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the benchmark's ground truth is external brute-force enumeration and textbook results, and no prediction reduces to a fitted parameter or self-citation.

full rationale

This paper is an empirical evaluation, not a derivation: it builds StatMechBench-v0 from textbook tractable classes (Table 2), verifies candidate partition functions against external brute-force enumeration (§3.2, App. C), and reports LLM-agent outcomes. The central claim—that agents can pass numerical checks while mislabeling the tractable class or understating complexity—is an observed behavioral pattern, not a quantity derived from fitted inputs. The tractable-class labels are supplied in the prompt and then used as the evaluation rubric; this is a benchmark design choice, not a circular reduction, since the labels are fixed external conventions (e.g., pfaffian, transfer_matrix) rather than outputs of the model being tested. The flagged O-audit regex false positive (Table 3 footnote) is a measurement/reproducibility concern, explicitly acknowledged by the authors, and does not make any claimed result true by construction. No load-bearing self-citation chain, uniqueness theorem, or ansatz-smuggling appears; references to prior work are context and inspiration (e.g., FunSearch, ERA), not the justification for the paper's empirical findings. Weaknesses such as single-trial measurements and unreleased logs affect statistical robustness but are correctness risks, not circularity. Honest non-finding is therefore appropriate.

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

The benchmark's validity rests on brute-force enumeration as ground truth at n≤16, on the correctness of the six gold labels in Table 2, and on hand-set verifier tolerances and probed sizes. The evaluation also assumes a single LLM trial per condition is representative. No physical entities are introduced; StatMechBench-v0 is an evaluation artifact. The principal free parameters are experimental hyperparameters (tolerance, round budget, sample count, system sizes) rather than fitted physical constants.

free parameters (4)
  • verifier tolerance epsilon_rel = 1e-6
    Hand-set pass criterion; a stricter tolerance could alter pass rates and change repair-loop behavior.
  • probed system sizes n per task = {6,10,14}, {9,16}, {8,10}
    Chosen small enough for brute-force ground truth; at n≤16, exponential base 2^sqrt(n) is tiny, so numerical checks cannot separate polynomial from sub-exponential implementations.
  • refinement rounds budget k = 5
    Hand-set budget; saturation curves show most trials finish by round 1, so the k* metric depends on this arbitrary cap.
  • parameter samples per probed n = 5
    Hand-set; more samples would reduce the false-accept risk of the brute-force verifier.
assumptions (6)
  • standard math Partition function is Z=sum_s exp(-beta H(s)) (Eq. 1); mappings are transformations to known solvable representations.
    Sec 2.1-2.2; this defines the task object for the whole benchmark.
  • domain assumption Brute-force enumeration over 2^n configurations yields exact ground-truth partition functions at n≤16.
    Appendix C verifier; if brute-force precision is wrong, all pass/fail labels are wrong.
  • domain assumption Gold tractable-class labels and complexity bounds in Table 2 correctly characterize P01-P06.
    Appendix B task cards; if any gold label is wrong, benchmark conclusions shift.
  • domain assumption Random 3-regular ±1 Ising glass has no efficient exact mechanism within the considered library.
    P06 negative control, Table 2; if a hidden tractable class applies, the negative-control interpretation fails.
  • domain assumption Paraphrased task descriptions carry the same information as canonical ones.
    Appendix B rewrite rule; if paraphrases dropped or added constraints, canonical-vs-paraphrase comparisons are invalid.
  • ad hoc to paper One trial per model-variant-task is representative of that LLM's performance.
    Sec 3.3 and Table 3; no repeated seeds are reported, so this is the paper's weakest evaluation assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Exploring Structures in Physics Problems: Can AI Agents Discover Statistical Mechanical Mappings?." pith.science (2026). https://pith.science/paper/4JFFLN5N

@misc{pith2026260726367,
  author       = {Pith},
  title        = {Pith review of: Exploring Structures in Physics Problems: Can AI Agents Discover Statistical Mechanical Mappings?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4JFFLN5N}},
  note         = {Machine review of arXiv:2607.26367}
}
read the original abstract

An important skill in theoretical physics is to recognize when a new problem can be transformed into a known model. We study this skill as an AI-agent task: can LLM-based agents discover statistical mechanical mappings from a raw partition function to a tractable representation? To probe this question, we introduce StatMechBench-v0, a benchmark of six Ising-type problems covering transfer-matrix methods, gauge-removable disorder, and planar/Pfaffian structure. We evaluate a simple propose-verify-revise agent across multiple LLMs and problem phrasings. The results show that numerical feedback often helps agents repair code and recover correct partition functions. However, agents can also pass the numerical checks while misidentifying the underlying tractable class or understating computational complexity. This both reveals limitations in current LLM reasoning and calls for a verification stack that goes beyond numerical agreement, incorporating, for example, symbolic checks and structural invariants. Our study provides an early evaluation and design directions for AI agents aimed at structural discovery in theoretical physics.

Figures

Figures reproduced from arXiv: 2607.26367 by the authors.

Figure 1
Figure 1. General logic underlying the exact evaluation of a partition function [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Cumulative numerical evaluation pass rates across self-refinement rounds, with eight LLMs [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

76 extracted references · 2 linked inside Pith

  1. [1]

    Journal of Mathematical Physics , volume =

    Topological quantum memory , author =. Journal of Mathematical Physics , volume =. 2002 , publisher =

    2002

  2. [2]

    Annales de l'Institut Henri Poincar

    Statistical mechanical models for quantum codes with correlated noise , author =. Annales de l'Institut Henri Poincar. 2021 , publisher =

    2021

  3. [3]

    Physical Review X , volume =

    Measurement-Induced Phase Transitions in the Dynamics of Entanglement , author =. Physical Review X , volume =. 2019 , publisher =

    2019

  4. [4]

    Physical Review B , volume =

    Measurement-induced criticality in random quantum circuits , author =. Physical Review B , volume =. 2020 , publisher =

    2020

  5. [5]

    Physical Review B , volume =

    Measurement-driven entanglement transition in hybrid quantum circuits , author =. Physical Review B , volume =. 2019 , publisher =

    2019

  6. [6]

    Physical Review B , volume =

    Critical properties of the measurement-induced transition in random quantum circuits , author =. Physical Review B , volume =. 2020 , publisher =

    2020

  7. [7]

    PRX Quantum , volume =

    Measurement-Induced Phase Transitions on Dynamical Quantum Trees , author =. PRX Quantum , volume =. 2023 , publisher =

    2023

  8. [8]

    Physical Review B , volume =

    Statistical mechanics of quantum error correcting codes , author =. Physical Review B , volume =. 2021 , publisher =

    2021

Show all 76 references
  1. [9]

    2011 , eprint =

    Fault-tolerant quantum computing with color codes , author =. 2011 , eprint =

    2011

  2. [10]

    Confinement-

    Wang, Chenyang and Harrington, Jim and Preskill, John , journal =. Confinement-. 2003 , publisher =

    2003

  3. [11]

    Physical Review Letters , volume =

    Solving Statistical Mechanics Using Variational Autoregressive Networks , author =. Physical Review Letters , volume =. 2019 , publisher =

    2019

  4. [12]

    2019 , eprint =

    Solving Statistical Mechanics on Sparse Graphs with Feedback Set Variational Autoregressive Networks , author =. 2019 , eprint =

    2019

  5. [13]

    Physical Review Letters , volume =

    Tropical Tensor Network for Ground States of Spin Glasses , author =. Physical Review Letters , volume =. 2021 , publisher =

    2021

  6. [14]

    and Colwell, Lucy , year =

    Brenner, Michael P. and Colwell, Lucy , year =. Solving an Open Problem in Theoretical Physics using. 2603.04735 , archivePrefix =

  7. [15]

    Polynomial-time approximation algorithms for the

    Jerrum, Mark and Sinclair, Alistair , journal =. Polynomial-time approximation algorithms for the. 1993 , publisher =

    1993

  8. [16]

    Physical Review Letters , volume =

    Tensor Renormalization Group Approach to Two-Dimensional Classical Lattice Models , author =. Physical Review Letters , volume =. 2007 , publisher =

    2007

  9. [17]

    and Shaham, Jacob , address =

    Feynman, Richard P. and Shaham, Jacob , address =. Statistical mechanics : a set of lectures , year =. Statistical mechanics : a set of lectures , isbn =

  10. [18]

    Novikov, Alexander and others , title=

  11. [19]

    Pawan and Dupont, Emilien and Ruiz, Francisco J

    Romera-Paredes, Bernardino and Barekatain, Mohammadamin and Novikov, Alexander and Balog, Matej and Kumar, M. Pawan and Dupont, Emilien and Ruiz, Francisco J. R. and Ellenberg, Jordan S. and Wang, Pengming and Fawzi, Omar and Kohli, Pushmeet and Fawzi, Alhussein , title=. Natu...

    doi

  12. [20]

    arXiv preprint arXiv:2509.06503 , year=

    An AI system to help scientists write expert-level empirical software , author=. arXiv preprint arXiv:2509.06503 , year=

    arXiv

  13. [21]

    Pawan and Dupont, Emilien and Ruiz, Francisco J

    Romera-Paredes, Bernardino and Barekatain, Mohammadamin and Novikov, Alexander and Balog, Matej and Kumar, M. Pawan and Dupont, Emilien and Ruiz, Francisco J. R. and Ellenberg, Jordan S. and Wang, Pengming and Fawzi, Omar and Kohli, Pushmeet and Fawzi, Alhussein , title =. Nat...

  14. [22]

    and Wu, Yuhuai and Le, Quoc V

    Trinh, Trieu H. and Wu, Yuhuai and Le, Quoc V. and He, He and Luong, Thang , title =. Nature , volume =

  15. [23]

    Ruiz, Francisco J

    Fawzi, Alhussein and Balog, Matej and Huang, Aja and Hubert, Thomas and Romera-Paredes, Bernardino and Barekatain, Mohammadamin and Novikov, Alexander and R. Ruiz, Francisco J. and Schrittwieser, Julian and Swirszcz, Grzegorz and Silver, David and Hassabis, Demis and Kohli, Pu...

  16. [24]

    2024 , howpublished =

    2024

  17. [25]

    Madaan, Aman and Tandon, Niket and Gupta, Prakhar and Hallinan, Skyler and Gao, Luyu and Wiegreffe, Sarah and Alon, Uri and Dziri, Nouha and Prabhumoye, Shrimai and Yang, Yiming and Gupta, Shashank and Majumder, Bodhisattwa Prasad and Hermann, Katherine and Welleck, Sean and Y...

  18. [26]

    Shinn, Noah and Cassano, Federico and Berman, Edward and Gopinath, Ashwin and Narasimhan, Karthik and Yao, Shunyu , title =. Proc. NeurIPS , year =

  19. [27]

    Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan , title =. Proc. ICLR , year =

  20. [28]

    Chen, Bei and Zhang, Fengji and Nguyen, Anh and Zan, Daoguang and Lin, Zeqi and Lou, Jian-Guang and Chen, Weizhu , title =. Proc. ICLR , year =

  21. [29]

    arXiv:2009.03393 , year =

    Polu, Stanislas and Sutskever, Ilya , title =. arXiv:2009.03393 , year =

    2009 arXiv

  22. [30]

    Yang, Kaiyu and Swope, Aidan and Gu, Alex and Chalamala, Rahul and Song, Peiyang and Yu, Shixing and Godil, Saad and Prenger, Ryan and Anandkumar, Anima , title =. Proc. NeurIPS Datasets and Benchmarks , year =

  23. [31]

    Zheng, Kunhao and Han, Jesse Michael and Polu, Stanislas , title =. Proc. ICLR , year =

  24. [32]

    Schick, Timo and Dwivedi-Yu, Jane and Dess\`i, Roberto and Raileanu, Roberta and Lomeli, Maria and Zettlemoyer, Luke and Cancedda, Nicola and Scialom, Thomas , title =. Proc. NeurIPS , year =

  25. [33]

    Gao, Luyu and Madaan, Aman and Zhou, Shuyan and Alon, Uri and Liu, Pengfei and Yang, Yiming and Callan, Jamie and Neubig, Graham , title =. Proc. ICML , year =

  26. [34]

    and Inala, Jeevana Priya and Wang, Chenglong and Gao, Jianfeng and Solar-Lezama, Armando , title =

    Olausson, Theo X. and Inala, Jeevana Priya and Wang, Chenglong and Gao, Jianfeng and Solar-Lezama, Armando , title =. 2023 , note =

    2023

  27. [35]

    and Robson, Esme Sutherland and Kohli, Pushmeet and de Freitas, Nando and Kavukcuoglu, Koray and Vinyals, Oriol , title =

    Li, Yujia and Choi, David and Chung, Junyoung and Kushman, Nate and Schrittwieser, Julian and Leblond, R\'emi and Eccles, Tom and Keeling, James and Gimeno, Felix and Lago, Agustin Dal and Hubert, Thomas and Choy, Peter and de Masson d'Autume, Cyprien and Babuschkin, Igor and ...

  28. [36]

    Zhang, Yanbo and Khan, Sumeer A. and Mahmud, Adnan and Yang, Huck and Lavin, Alexander and Levin, Michael and Frey, Jeremy and Dunnmon, Jared and Evans, James and Bundy, Alan and Dzeroski, Saso and Tegner, Jesper and Zenil, Hector , title=. npj Artificial Intelligence , year=....

    doi

  29. [37]

    2025 , eprint=

    Evaluating Large Language Models in Scientific Discovery , author=. 2025 , eprint=

    2025

  30. [38]

    Crystal Statistics. I. A Two-Dimensional Model with an Order-Disorder Transition , author =. Physical Review , volume =. 1944 , publisher =

    1944

  31. [39]

    Mathematical Proceedings of the Cambridge Philosophical Society , volume =

    Some Generalized Order-Disorder Transformations , author =. Mathematical Proceedings of the Cambridge Philosophical Society , volume =. 1952 , publisher =

    1952

  32. [40]

    Part I , author =

    Statistics of the Two-Dimensional Ferromagnet. Part I , author =. Physical Review , volume =. 1941 , publisher =

    1941

  33. [41]

    1982 , isbn =

    Exactly Solved Models in Statistical Mechanics , author =. 1982 , isbn =

    1982

  34. [42]

    arXiv preprint arXiv:2512.16654 , year=

    Scalable tests of quantum contextuality from stabilizer-testing nonlocal games , author=. arXiv preprint arXiv:2512.16654 , year=

  35. [43]

    and Wigner, E

    Jordan, P. and Wigner, E. , journal =. 1928 , doi =

    1928

  36. [44]

    Physical Review Letters , volume =

    Jordan-Wigner Transformation for Quantum-Spin Systems in Two Dimensions and Fractional Statistics , author =. Physical Review Letters , volume =. 1989 , publisher =

    1989

  37. [45]

    The Number of Dimer Arrangements on a Quadratic Lattice , author =

    The Statistics of Dimers on a Lattice: I. The Number of Dimer Arrangements on a Quadratic Lattice , author =. Physica , volume =. 1961 , doi =

    1961

  38. [46]

    Physical Review , volume =

    Statistical Mechanics of Dimers on a Plane Lattice , author =. Physical Review , volume =. 1961 , publisher =

    1961

  39. [47]

    Physical Review B , volume=

    Theory of the phase transition in random unitary circuits with measurements , author=. Physical Review B , volume=. 2020 , publisher=

    2020

  40. [48]

    Zeitschrift f

    Ising, Ernst , title =. Zeitschrift f. 1925 , doi =

    1925

  41. [49]

    Journal of Mathematical Physics , volume =

    Lenard, Andrew , title =. Journal of Mathematical Physics , volume =. 1961 , doi =

    1961

  42. [50]

    , title =

    Baxter, Rodney J. , title =. Mathematical Proceedings of the Cambridge Philosophical Society , volume =. 1963 , doi =

    1963

  43. [51]

    and Hemmer, Per C

    Kac, Mark and Uhlenbeck, George E. and Hemmer, Per C. , title =. Journal of Mathematical Physics , volume =. 1963 , doi =

    1963

  44. [52]

    and Hemmer, Per C

    Uhlenbeck, George E. and Hemmer, Per C. and Kac, Mark , title =. Journal of Mathematical Physics , volume =. 1963 , doi =

    1963

  45. [53]

    and Kac, Mark and Uhlenbeck, George E

    Hemmer, Per C. and Kac, Mark and Uhlenbeck, George E. , title =. Journal of Mathematical Physics , volume =. 1964 , doi =

    1964

  46. [54]

    , title =

    Montroll, Elliott W. , title =. Il Nuovo Cimento , volume =

  47. [55]

    and Kac, Mark , title =

    Berlin, Theodore H. and Kac, Mark , title =. Physical Review , volume =. 1952 , doi =

    1952

  48. [56]

    Eugene , title =

    Stanley, H. Eugene , title =. Physical Review , volume =. 1968 , doi =

    1968

  49. [57]

    and Wannier, Gregory H

    Kramers, Hendrik A. and Wannier, Gregory H. , title =. Physical Review , volume =. 1941 , doi =

    1941

  50. [58]

    , title =

    Lieb, Elliott H. , title =. Physical Review Letters , volume =. 1967 , doi =

    1967

  51. [59]

    , title =

    Lieb, Elliott H. , title =. Physical Review , volume =. 1967 , doi =

    1967

  52. [60]

    , title =

    Baxter, Rodney J. , title =. Physical Review Letters , volume =. 1971 , doi =

    1971

  53. [61]

    and Wu, Fa-Yueh , title =

    Baxter, Rodney J. and Wu, Fa-Yueh , title =. Physical Review Letters , volume =. 1973 , doi =

    1973

  54. [62]

    Thirty-seventh Conference on Neural Information Processing Systems , year=

    Self-Refine: Iterative Refinement with Self-Feedback , author=. Thirty-seventh Conference on Neural Information Processing Systems , year=

  55. [63]

    2025 , url=

    Seyed Iman Mirzadeh and Keivan Alizadeh and Hooman Shahrokhi and Oncel Tuzel and Samy Bengio and Mehrdad Farajtabar , booktitle=. 2025 , url=

    2025

  56. [64]

    Proceedings of the 42nd International Conference on Machine Learning , pages =

    A Machine Learning Approach to Duality in Statistical Physics , author =. Proceedings of the 42nd International Conference on Machine Learning , pages =. 2025 , editor =

    2025

  57. [65]

    Leverrier, Anthony and Z. Quantum. Proceedings of the 63rd IEEE Annual Symposium on Foundations of Computer Science (FOCS) , pages =. 2022 , doi =

    2022

  58. [66]

    , title =

    Lenz, W. , title =. Physikalische Zeitschrift , volume =

  59. [67]

    PRX Quantum , volume=

    Statistical mechanical mapping and maximum-likelihood thresholds for the surface code under generic single-qubit coherent errors , author=. PRX Quantum , volume=. 2025 , publisher=

    2025

  60. [68]

    Introducing Claude Haiku 4.5 , year =

  61. [69]

    Introducing Claude Sonnet 4.6 , year =

  62. [70]

    Introducing Claude Opus 4.7 , year =

  63. [71]

    DeepSeek-V4-Flash , year =

  64. [72]

    Thinking Mode , year =

  65. [73]

    DeepSeek-V4-Pro , year =

  66. [74]

    Gemini 3.1 Pro , year =

  67. [75]

    GPT-5.4 mini , year =

  68. [76]

    2026 , eprint=

    PRL-Bench: A Comprehensive Benchmark Evaluating LLMs' Capabilities in Frontier Physics Research , author=. 2026 , eprint=

    2026

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.