Pith. sign in

REVIEW 5 major objections 8 minor 24 references

Large-scale Neural Network Quantum States for ab initio Quantum Chemistry Simulations on Fugaku

T0 review · 5 major / 8 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper argues that neural network quantum state training can scale to 1,536 nodes at 95.8% parallel efficiency while reproducing FCI ground-state energies for N2, PH3, and LiCl.

desk verdict A serious HPC-NQS engineering paper whose headline scaling claim is real but needs a terminology fix and more load-balance evidence before I'd trust the 95.8% number. read the letter →

arxiv 2506.23809 v1 pith:NHBOILOB submitted 2025-06-30 cs.DC

classification cs.DC
keywords neuralnetworkquantumstatesabinitiochemistryparallelefficiencydynamicloadbalancingvariationalMonteCarlotransformeransatzcacheoptimizationFugakusupercomputer
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 tries to show that the scalability barriers in training neural network quantum states (NQS) for ab initio chemistry — explosive sample growth, unpredictable memory footprints, and costly local energy evaluation — can be broken with system-level optimizations. The framework it presents, QChem-Trainer, scales NQS training to 1,536 nodes of the Fugaku supercomputer at up to 95.8% parallel efficiency and cuts per-iteration time by up to 8.41x compared with the baseline. On N2, PH3, and LiCl in a minimal basis, its transformer ansatz reproduces full configuration interaction (FCI) ground-state energies to within 0.0002 Hartree. A sympathetic reader would care because the exponential sampling cost has kept NQS methods on small test systems, and the paper claims a route to making them practical on exascale machines.

What carries the argument

The central object is the density metric $d = \mathrm{sample\_unique}/\mathrm{sample\_counts}$, computed at each split layer during auto-regressive sampling. It expresses the ratio of unique electron configurations to total drawn samples; because neural network parameters change smoothly between training iterations, $d$ is treated as approximately stable across iterations, letting the scheduler multiply each process group's known sample count by the latest $d$ to predict its future unique-sample workload. Around this metric the paper builds multi-stage workload partitioning (vertical groups for splitting, horizontal groups for balancing), a hybrid BFS/DFS sampling scheme that caps peak memory by switching to depth-first traversal once unique samples exceed a threshold, and cache-centric transformer optimization so the transformer ansatz's memory footprint stays flat at scale.

What would settle it

Run the density-aware load balancer on a molecular system whose wavefunction changes sharply during training, such as N2 at dissociation, and track the max-to-mean ratio of unique samples per rank across all epochs; if at some epoch that ratio grows by more than a factor of two relative to the Fe2S2 epoch-4 measurement, the continuity premise is falsified in a regime the paper does not cover.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that NQS training need not be memory-unstable or communication-bound. The sampling phase can be reorganized as a multi-stage quadtree partition in which processes are grouped into vertical groups for splitting and horizontal groups for load balancing; a density metric $d = \mathrm{sample\_unique}/\mathrm{sample\_counts}$, updated by local all-reduce within each horizontal group, predicts how many unique samples each slice of the current sample count will produce because network parameters evolve smoothly between iterations; and a hybrid BFS-DFS sampling scheme with a fixed-size cache pool keeps peak memory flat even when the number of samples grows by orders of magnitude. Combined with MPI/OpenMP/SVE parallelization of local energy evaluation and cache pooling, selective recomputation, and lazy expansion for transformer Key/Value caches, the paper reports that the framework reaches FCI-quality ground-state energies for N2, PH3, and LiCl, sustains 95.8% parallel efficiency in weak scaling to 1,536 nodes, and accelerates end-to-end training by up to 8.41x.

Load-bearing premise

The entire load-balancing strategy rests on the assumption that the distribution of sampled electron configurations changes only slowly between consecutive training steps, so yesterday's ratio of unique to repeated samples predicts today's; if that ratio jumps between iterations, the predicted partitions go wrong and the 95.8% efficiency number would not generalize.

Editorial extensions

If this is right

  • If the density-aware load balancing holds across training, NQS can grow from single-node demonstrations to thousands of nodes; the paper demonstrates this for the H50 chain up to 1,536 nodes at 95.8% weak-scaling efficiency.
  • Memory-stable sampling lets NQS run with sample counts as high as 1.024e7, three orders of magnitude beyond the plain KVCache version, which runs out of memory at 2e4.
  • The multi-level local energy parallelism generalizes beyond NQS: qubit-packing, branch elimination, and SVE vectorization apply to any Slater-Condon Hamiltonian evaluation, so other VMC or ab initio codes could reuse them.
  • Larger molecular systems benefit more, with speedups rising from 1.83x for N2 to 8.41x for C6H6, so the optimization targets the regime where NQS is currently most infeasible.

Reading between the lines

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

  • The continuity assumption that underpins $d$ could be validated directly by measuring the distribution shift of sampled configurations between consecutive epochs; if the shift spikes at bond-breaking or under aggressive learning-rate schedules, the load balancer may need to recompute $d$ more often.
  • The density-aware partition has a natural extension to heterogeneous clusters: instead of balancing ranks by predicted unique counts, weight ranks by compute capability, using the same $d$ to convert sample counts into expected time per node.
  • The fixed-size cache pool and lazy expansion ideas transfer to any autoregressive generation workload with dynamically expanding batch sizes, such as speculative decoding in large language models.
  • If this scaling reproduces on GPU supercomputers, the practical bottleneck for NQS will shift from training throughput to ansatz accuracy and statistical error, making wavefunction architecture the next rate-limiter.
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

5 major / 8 minor

Summary. The paper presents QChem-Trainer, an HPC framework for training neural-network quantum states (NQS) for ab initio quantum chemistry on the Fugaku supercomputer. It proposes three main optimizations: (i) a sampling-parallelism strategy with multi-stage workload partitioning, density-aware dynamic load balancing, and a hybrid BFS/DFS sampling scheme; (ii) a three-level parallelization of local-energy evaluation using MPI, OpenMP, and SVE vectorization; and (iii) cache-centric optimizations for transformer-based ansatze, including fixed-size cache pooling, selective recomputation, and lazy cache expansion. Experiments report ground-state energies matching FCI for N2, PH3, and LiCl, an up-to-8.41x end-to-end speedup over the authors' base implementation, and a 95.8% parallel efficiency measured in a weak-scaling run to 1,536 nodes on the H50 system. The central claims are that these optimizations preserve accuracy, improve training throughput, and enable scalable NQS training on a large supercomputer.

Significance. If the scalability and performance claims hold up, the paper is a useful engineering contribution: it demonstrates that NQS training for ab initio chemistry can be pushed to more than a thousand nodes on a modern CPU-based supercomputer, and it provides concrete techniques (density-aware load balancing, KV-cache pooling, SIMD-friendly Slater-Condon evaluation) that could inform future NQS implementations. Strengths include direct numerical validation of energies against FCI for several molecules, performance measurements on a real large-scale machine, and an explicit separation of the authors' optimizations from a baseline implementation. However, the central scalability claim rests on the density-continuity assumption, which is not yet convincingly validated, and the reported parallel efficiency is weak scaling while the conclusion labels it strong scaling. These issues, plus the lack of comparison with prior NQS frameworks such as NNQS-transformer, mean the significance is real but the manuscript needs substantial revision before the claims can be fully accepted.

major comments (5)
  1. [Section 3.1.2, Algorithm 2, Figure 4a] The density-aware dynamic load balancing is a load-bearing component of the 95.8% parallel-efficiency claim, yet its core assumption is only demonstrated for one epoch of one system. Figure 4a shows the maximum unique-sample distribution at Epoch 4 for Fe2S2 at 256 ranks, but no time series of the density metric d, no statistics across training iterations, and no load-balance measurements at 1,536 nodes are presented. The text asserts continuity of parameter evolution and spatial locality, but does not quantify them. Since Algorithm 2 uses d from the preceding split segment (and, per the text, from previous iterations) to predict next-segment workload, the paper should either provide direct evidence that d is stable across layers and iterations, or characterize the regime in which load balance degrades. Without this, the scalability result should be treated as conditional on this assumption.
  2. [Section 4.3.4 and Conclusion] The scaling experiment is explicitly described as weak scaling: Figure 6 is labeled "Weak Scaling" and the setup text states "Nu = N * 4 * 10^3 for N nodes," meaning the workload per node is held constant. The Conclusion, however, states that "strong scaling tests show that QChem-Trainer maintains parallel efficiency of up to 95.8%." These statements are incompatible. The reported 95.8% is a weak-scaling efficiency, not a strong-scaling efficiency, and the current text gives no strong-scaling data at all. This must be corrected, because a fixed-size molecular system may not exhibit the same efficiency when scaled to 1,536 nodes.
  3. [Section 4.2, Figure 3 Right] The headline speedup of 8.41x is measured against an unlabeled "base" implementation, and no comparison is made to existing NQS codes such as NNQS-transformer [11], which is cited as a scalable NQS approach for ab initio chemistry. The strength of the speedup claim depends entirely on what is in that baseline. The paper should describe the baseline in enough detail (which parallelization, whether KV cache is used, what the sampling and energy routines include) and, if possible, provide a comparison or at least a qualitative discussion against prior work. Without this, the abstract's "up to 8.41x speedup" is potentially misleading to readers who do not know the baseline.
  4. [Table 1 and Figure 3 Left] The accuracy validation against FCI is one of the paper's most important contributions, but the reported energies have no statistical uncertainties. VMC energy estimates are stochastic, so the table values and the N2 potential-energy surface need error bars or a statement of the number of samples, the number of training iterations, and the convergence criterion used. Without this information, it is impossible to assess whether the agreement with FCI is within statistical noise or whether the reported values are from a single run.
  5. [Section 4.3.1, Figure 4b] The "memory-stable" claim is not directly supported by any reported memory-footprint measurement. Figure 4b plots iteration time versus sample count and marks OOM points, but it does not plot peak memory usage per rank or per node. The text states that the memory-stable method "can constrain the peak memory footprint," but the only evidence is the absence of OOM at tested sizes. Please report measured memory consumption as a function of sample count, and ideally per node at scale, to substantiate the memory-stability contribution.
minor comments (8)
  1. [Throughout] There are many typos and formatting errors, including "Backgroud" (Section 2 heading), "nerual" (Introduction), "prouning" (Section 2.2), "desity" (Figure 4a caption), "read line" (should be "red line" in Figure 4b caption), "mothods" (Section 4.3.4), and "1,526 nodes" versus "1,536 nodes" in the Conclusion. These should be corrected.
  2. [Table 1 section and Figure 3 Left] The text contains corrupted artifact strings such as "/uni00000013/uni00000011/..." that appear to be PDF-extraction errors. These must be removed before any final version is published.
  3. [Section 3.1.2] The explanation of density continuity says that sample distributions vary smoothly "between consecutive iterations," but Algorithm 2 updates D from the previous split segment within the same sampling pass. The text and algorithm should be aligned so the reader understands which prediction step actually uses the density metric.
  4. [Algorithm 3, lines 12-14] The branch-elimination optimization is described only informally. Please state the exact correctness condition of the predicate pred_0 and describe what Customized_function() does when pred_0 is true, so that reviewers can verify that the vectorized path is indeed equivalent to the Slater-Condon rules.
  5. [Section 2.2 and Introduction] Reference [11] (NNQS-transformer) is highly relevant prior work that also targets scalable NQS for ab initio quantum chemistry. The paper should explicitly differentiate QChem-Trainer from [11] in terms of parallelization strategy and experimental results, rather than merely citing it.
  6. [Section 4.2] The phrase "NUMA-aware MPI setting" is vague. Please specify the MPI rank-to-core mapping, any process pinning, and the communication settings used so that the experiments are reproducible.
  7. [Section 3.2] The "qubit-packing" description says that 64 orbital occupations are packed into one double-float format, but the mechanics of how 64 bits are stored in a double and how SVE operations interact with this representation is not explained clearly. A more detailed description, or a reference to a prior description, would help readers understand the vectorization.
  8. [Figure 6] The figure has two y-axes (time per iteration and parallel efficiency) but the captions do not specify which axis corresponds to which quantity. Add axis labels and make clear that the efficiency is weak-scaling efficiency.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: QChem-Trainer is an engineering report whose accuracy, speedup, and scalability claims are measured directly against baselines and FCI references, with no fitted parameter renamed as a prediction.

full rationale

The paper's central claims are empirical measurements rather than derivations. Ground-state energies are computed by variational Monte Carlo using Eq. (1) and compared directly to FCI and CCSD values in Table 1; the N2 potential energy surface provides an external benchmark. The speedup claims are based on end-to-end execution time per iteration relative to a baseline implementation, and the scalability claim is based on measured weak-scaling parallel efficiency at 1,536 nodes (Figure 6b). No fitted parameter enters these measurements, and no calibration to the target result is used. The density metric d = sample_unique / sample_counts in Section 3.1.2 is a workload-prediction heuristic, not a derived quantity; the 'continuity of parameter evolution' is explicitly stated as an assumption enabling the load-balancing strategy. Even if that assumption is only validated at one epoch for Fe2S2 (Figure 4a), that is a robustness or validation concern, not circular reasoning. The paper cites prior work, including NNQS-transformer [11], but only as a baseline and related approach; it does not invoke any self-citation as a load-bearing proof of the new claims. No step in the paper reduces, by definition or by fitted construction, to its own inputs. The derivation chain is therefore self-contained with respect to circularity, though the dynamic load-balancing premise deserves stronger empirical support.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claims depend on standard VMC formulas, the assumption of smooth sample-distribution evolution, the dominance of double excitations, and the negligible cost of selective recomputation. No new physical entities are introduced; the hand-set sampling threshold and cache pool size are implementation parameters that affect memory-stability and performance but are not fitted to data.

free parameters (2)
  • BFS/DFS switch threshold k
    Controls when sampling switches from layer-wise BFS to chunked DFS; the value is not stated and would need tuning per system, affecting the memory-stability claim.
  • cache pool size
    Fixed-size KVCache pool is a hand-set memory budget; the paper does not specify how it is chosen across systems, though it is central to the stable-memory claim.
assumptions (4)
  • standard math VMC expectation value and gradient formulas (Eq. 1-4) are correct and the stochastic estimator is unbiased
    All energy results inherit the standard VMC estimator; not derived in this paper.
  • domain assumption Sample distributions vary smoothly between consecutive iterations, allowing historical density to predict current workload
    Stated in Section 3.1.2; required by Algorithm 2 for load balance, tested only at epoch 4 in Figure 4a.
  • domain assumption Double excitations dominate the Hamiltonian evaluation, so branch-free vectorization of Slater-Condon rules remains correct
    Stated in Section 3.2; the customized check function guards correctness, but performance depends on this dominance.
  • domain assumption Recomputing discarded KV caches when switching to DFS is negligible in cost
    Stated in Section 3.3.1; used to justify the hybrid sampling and cache strategy without a measured breakdown of recomputation overhead.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Large-scale Neural Network Quantum States for ab initio Quantum Chemistry Simulations on Fugaku." pith.science (2026). https://pith.science/paper/NHBOILOB

@misc{pith2026250623809,
  author       = {Pith},
  title        = {Pith review of: Large-scale Neural Network Quantum States for ab initio Quantum Chemistry Simulations on Fugaku},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NHBOILOB}},
  note         = {Machine review of arXiv:2506.23809}
}
read the original abstract

Solving quantum many-body problems is one of the fundamental challenges in quantum chemistry. While neural network quantum states (NQS) have emerged as a promising computational tool, its training process incurs exponentially growing computational demands, becoming prohibitively expensive for large-scale molecular systems and creating fundamental scalability barriers for real-world applications. To address above challenges, we present \ours, a high-performance NQS training framework for \textit{ab initio} electronic structure calculations. First, we propose a scalable sampling parallelism strategy with multi-layers workload division and hybrid sampling scheme, which break the scalability barriers for large-scale NQS training. Then, we introduce multi-level parallelism local energy parallelism, enabling more efficient local energy computation. Last, we employ cache-centric optimization for transformer-based \textit{ansatz} and incorporate it with sampling parallelism strategy, which further speedup up the NQS training and achieve stable memory footprint at scale. Experiments demonstrate that \ours accelerate NQS training with up to 8.41x speedup and attains a parallel efficiency up to 95.8\% when scaling to 1,536 nodes.

Figures

Figures reproduced from arXiv: 2506.23809 by the authors.

Figure 1
Figure 1. (a): NQS workflows. (b): Transformers architec [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the sampling parallelsim. (a): Multi [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Left: Potential energy surfaces of N2. Righ: Overall speedup. Four molecular systems are chosen to test QChem-Trainer’s performance. From left to right, the spin orbits in molecular system increases. The figure demonstrate the normalized speedups compared to baseline implementa￾tion. To test the overall performance gains, we choose four systems with the increased spin orbits: N2, F e2S2, H50 and C6H6 with 20, 40, 10… view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Step by step energy calculation speedup with [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: We test the scaling performance in H50 system with two energy calculation methods. Figure (a) demonstrate the scaling result using sample space energy calculation. In contrast, Figure (b) employs another accurate energy calculation method. Both of them showcase the sca…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 19 canonical work pages

  1. [11]

    Nnqs-transformer: an efficient and scalable neural network quantum states approach for ab initio quantum chemistry,

    Y . Wu, C. Guo, Y . Fan, P. Zhou, and H. Shang, “Nnqs-transformer: an efficient and scalable neural network quantum states approach for ab initio quantum chemistry,” in Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis , ser. SC ’23. New York, NY , USA: Association for Computing Machinery, 2023

  2. [1]

    The configuration interaction method: Advances in highly correlated approaches,

    C. D. Sherrill and H. F. Schaefer III, “The configuration interaction method: Advances in highly correlated approaches,” in Advances in quantum chemistry . Elsevier, 1999, vol. 34, pp. 143–269

  3. [2]

    Coupled-cluster theory in quantum chemistry,

    R. J. Bartlett and M. Musiał, “Coupled-cluster theory in quantum chemistry,” Reviews of Modern Physics , vol. 79, no. 1, pp. 291–352, 2007

  4. [3]

    Møller–plesset perturbation theory: from small molecule methods to methods for thousands of atoms,

    D. Cremer, “Møller–plesset perturbation theory: from small molecule methods to methods for thousands of atoms,” Wiley Interdisciplinary Reviews: Computational Molecular Science , vol. 1, no. 4, pp. 509– 530, 2011

  5. [4]

    Wave function optimization in the variational monte carlo method,

    S. Sorella, “Wave function optimization in the variational monte carlo method,” Physical Review B—Condensed Matter and Materials Physics, vol. 71, no. 24, p. 241103, 2005

  6. [5]

    Solving the quantum many-body problem with artificial neural networks,

    G. Carleo and M. Troyer, “Solving the quantum many-body problem with artificial neural networks,” Science, vol. 355, no. 6325, pp. 602– 606, 2017

  7. [6]

    Ab initio quantum chemistry with neural-network wavefunctions,

    J. Hermann, J. Spencer, K. Choo, A. Mezzacapo, W. M. C. Foulkes, D. Pfau, G. Carleo, and F. Noé, “Ab initio quantum chemistry with neural-network wavefunctions,” Nature Reviews Chemistry , vol. 7, no. 10, pp. 692–709, Aug. 2023

  8. [7]

    Recurrent neural network wave functions,

    M. Hibat-Allah, M. Ganahl, L. E. Hayward, R. G. Melko, and J. Carrasquilla, “Recurrent neural network wave functions,” Physical Review Research, vol. 2, no. 2, p. 023358, Jun. 2020. [Online]. Avail- able: https://link.aps.org/doi/10.1103/PhysRevResearch.2.023358

Show all 24 references
  1. [8]

    From tensor-network quantum states to tensorial recurrent neural networks,

    D. Wu, R. Rossi, F. Vicentini, and G. Carleo, “From tensor-network quantum states to tensorial recurrent neural networks,” Physical Re- view Research, vol. 5, no. 3, p. L032001, 2023

  2. [9]

    Two-dimensional frustrated j 1-j 2 model studied with neural network quantum states,

    K. Choo, T. Neupert, and G. Carleo, “Two-dimensional frustrated j 1-j 2 model studied with neural network quantum states,” Physical Review B, vol. 100, no. 12, p. 125124, 2019

  3. [10]

    Variational optimiza- tion of the amplitude of neural-network quantum many-body ground states,

    J.-Q. Wang, H.-Q. Wu, R.-Q. He, and Z.-Y . Lu, “Variational optimiza- tion of the amplitude of neural-network quantum many-body ground states,” Physical Review B , vol. 109, no. 24, p. 245120, 2024

  4. [12]

    Transformer variational wave functions for frustrated quantum spin systems,

    L. L. Viteritti, R. Rende, and F. Becca, “Transformer variational wave functions for frustrated quantum spin systems,” Physical Review Letters, vol. 130, no. 23, p. 236401, 2023

  5. [13]

    A Self-Attention Ansatz for Ab-initio Quantum Chemistry,

    I. von Glehn, J. S. Spencer, and D. Pfau, “A Self-Attention Ansatz for Ab-initio Quantum Chemistry,” 11th International Conference on Learning Representations (ICLR) , 2023

  6. [14]

    Supercomputer Fugaku - Supercomputer Fugaku, A64FX 48C 2.2GHz, Tofu interconnect D | TOP500

    “Supercomputer Fugaku - Supercomputer Fugaku, A64FX 48C 2.2GHz, Tofu interconnect D | TOP500.” [Online]. Available: https://www.top500.org/system/179807/

  7. [15]

    A stochastic approximation method,

    H. Robbins and S. Monro, “A stochastic approximation method,” The annals of mathematical statistics , pp. 400–407, 1951

  8. [16]

    Adam: A method for stochastic optimiza- tion,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimiza- tion,” arXiv preprint arXiv:1412.6980 , 2014

  9. [17]

    Fixing weight de- cay regularization in adam,

    I. Loshchilov and F. Hutter, “Fixing weight de- cay regularization in adam,” in International Conference on Learning Representations , 2019. [Online]. Available: https://openreview.net/forum?id=rk6qdGgCZ

  10. [18]

    Optimizing neural networks with kronecker-factored approximate curvature,

    J. Martens and R. Grosse, “Optimizing neural networks with kronecker-factored approximate curvature,” in International confer- ence on machine learning . PMLR, 2015, pp. 2408–2417

  11. [19]

    Scalable neural quantum states architecture for quantum chemistry,

    T. Zhao, J. Stokes, and S. Veerapaneni, “Scalable neural quantum states architecture for quantum chemistry,” Machine Learning: Sci- ence and Technology , vol. 4, Jun. 2023

  12. [20]

    An efficient implementa- tion of slater-condon rules,

    A. Scemama and E. Giner, “An efficient implementa- tion of slater-condon rules,” 2013. [Online]. Available: https://arxiv.org/abs/1311.6244

  13. [21]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” in Proceedings of the 31st International Conference on Neural Informa- tion Processing Systems, ser. NIPS’17. Red Hook, NY , USA: Curran Associate...

  14. [22]

    HPCG - November 2024 | TOP500

    “HPCG - November 2024 | TOP500.” [Online]. Available: https://top500.org/lists/hpcg/2024/11/

  15. [23]

    Language models are unsupervised multitask learners,

    A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever et al., “Language models are unsupervised multitask learners,” OpenAI blog, vol. 1, no. 8, p. 9, 2019

  16. [24]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” in International Conference on Learning Representations , 2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.