Pith. sign in

REVIEW 3 major objections 6 minor 25 references

Improving Generative Inverse Design of Rectangular Patch Antennas with Test Time Optimization

T0 review · 3 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Spending more inference-time compute improves generated rectangular patch antenna designs, without retraining.

desk verdict A plausible two-stage generative pipeline for patch antenna inverse design whose accuracy claim rests on an unvalidated surrogate scorer; the one full-wave failure suggests the concern is real. read the letter →

arxiv 2505.18188 v2 pith:FH4LBPKV submitted 2025-05-19 eess.SP cs.AIcs.LGcs.SYeess.SY

Beck LaBash , Shahriar Khushrushahi , Fabian Ruehle This is my paper · ORCID
classification eess.SPcs.AIcs.LGcs.SYeess.SY
keywords inversedesignrectangularpatchantennatest-timecomputevariationalautoencoderconditionalgenerationadversarialdisentanglementS11reflectioncoefficientelectromagneticsimulation
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 establish that for coaxial-fed rectangular patch antennas, design accuracy can be bought at inference time rather than with more training data. The proposed pipeline first trains a variational autoencoder on measured reflection-coefficient (S11) curves, then trains an adversarially conditioned generative model that maps a target response curve to antenna dimensions (patch length, width, feed position). At test time the paper searches the response-curve latent space, samples many candidate designs, and gradient-optimizes the design latent code against geometric constraints. Across three target specifications, simulated S11 curves improve as the number of sampled curves and designs per curve grows, and the paper argues this behavior parallels test-time compute scaling in other domains.

What carries the argument

The load-bearing mechanism is the two-stage latent pipeline: a response variational autoencoder that maps full S11 curves into a low-dimensional latent space, followed by an adversarially disentangled conditional variational autoencoder that maps antenna dimensions (L,W,p) into a separate latent code and decodes geometries from that code together with the response curve. The adversarial predictor is what makes the second stage controllable: it tries to recover the response curve from the design latent code, so the encoder is trained to strip response information out of that code and the decoder must use the supplied curve to reconstruct the geometry. At test time the same latents are treated as search spaces: gradient steps move the response latent toward in-distribution curves that match the masked target, and move the design latent to satisfy geometric constraints without changing the response.

What would settle it

Run full electromagnetic simulation over the complete candidate pools from the test-time scaling experiments and compare the design selected by the lowest surrogate score with the design having the lowest true simulated error for each target. If the surrogate-selected designs do not also achieve the lowest simulated error, or if the accuracy gap between small and large search budgets disappears under true simulation, the central claim is refuted.

Watch

Extended reading notes

Core claim

The central claim is that a two-stage generative architecture becomes more accurate as more compute is spent during inference, without retraining. Stage 1 compresses 1000-point S11 response curves into a 64-dimensional latent space with a beta-weighted variational autoencoder. Stage 2 trains a conditional variational autoencoder whose decoder maps a design latent code together with a response curve to (L,W,p); an adversarial predictor tries to infer the response curve from the design latent code, and the encoder is penalized for making that possible. This forces the decoder to rely on the explicit conditioning input. At test time, gradient search over the response latent finds in-distribution curves matching the masked target bands, best-of-N sampling over both latent spaces grows the candidate pool, and gradient optimization of the design latent against geometric penalties improves manufacturability while preserving the response. The paper demonstrates on three targets that larger search budgets produce simulated responses closer to the target, with the 5.0 GHz target still missing its depth requirement even after extensive search.

Load-bearing premise

The load-bearing premise is that the neural surrogate scorer ranks candidate designs the same way full electromagnetic simulation would; if the surrogate over-ranks designs that fail in simulation, the reported test-time gains could be an artifact.

Editorial extensions

If this is right

  • Increasing the number of sampled response curves from 1 to 20 lowers the average lowest surrogate score for both random and nearest-neighbor latent initialization.
  • Increasing the number of sampled designs per curve also lowers the lowest surrogate score, and optimizing the design latent against geometric penalties keeps that accuracy while yielding more manufacturable geometries.
  • Because the pipeline is probabilistic, multiple distinct geometries can satisfy the same response target, and best-of-N sampling exploits that one-to-many structure.
  • The test-time gains do not require expanding the training dataset or retraining the models, so the method turns inference compute directly into design accuracy.

Reading between the lines

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

  • The same two-stage latent-search recipe should transfer to other inverse-design problems with expensive forward simulators, such as optics, acoustics, or materials, whenever a learnable manifold of feasible responses can be searched.
  • The 5.0 GHz failure suggests the response autoencoder's latent manifold may lack curves with enough depth at that frequency and bandwidth; adding training data near that region is a concrete test of whether search is limited by the manifold's coverage.
  • An ablation that trains the conditional generator without the Stage 1 latent search, conditioning directly on the idealized target curve, would isolate how much of the accuracy gain comes from the response search versus the controllable decoder.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes a two-stage deep learning framework for inverse design of rectangular patch antennas. Stage 1 trains a VAE to encode S11 frequency response curves into a latent space; at test time, a latent-space search finds an in-distribution response matching a user-specified target (a product of Lorentzian notches). Stage 2 trains a conditionally disentangled CVAE that maps a target response and a latent code to antenna geometry (L, W, p). The central claim is that test-time compute - specifically best-of-N sampling of curves and designs and gradient-based optimization of the conditional latent code - improves the accuracy of generated designs without additional training data, while also enabling auxiliary objectives such as manufacturability. The experiments report scaling curves (Figure 3) using a neural surrogate scorer, and show full-wave openEMS simulations for three target responses at two search budgets (Figure 4). The paper concludes that the approach generalizes to more complex design spaces and provides public code and data.

Significance. If established, the claim that test-time search and optimization can substantially improve inverse-design accuracy with limited training data would be practically valuable, given the cost of EM simulation datasets. The paper has notable strengths: it uses an external full-wave solver (openEMS) for final validation, it explicitly acknowledges a failure case (Sec. 6.3), and it releases the dataset, code, and trained models. The two-stage architecture (latent-space response search plus conditioned geometry generation) is a sensible way to handle out-of-distribution targets and one-to-many mappings. However, the quantitative evidence for the central scaling claim is largely self-referential: Figure 3 ranks candidates with a surrogate trained on the same simulation data, with no demonstrated correlation to true EM error, and the only independent full-simulation check (Figure 4) includes one failure that contradicts the monotone-improvement narrative. These issues are fixable with additional validation, but they are load-bearing for the paper's main conclusion.

major comments (3)
  1. [Section 6.2, Figure 3] The scaling evidence in Figure 3 is based entirely on the 'Surrogate Scorer', but the score is not defined anywhere in the paper. Section 5.5 mentions that the surrogate uses a beta-NLL loss and uncertainty estimates, but no exact formula for the score or the frequency masking is given. More importantly, the manuscript never validates that the surrogate score correlates with the true openEMS error on generated candidates. Best-of-N sampling will monotonically lower any scorer's value as the pool grows, so the monotone curves in Figure 3 are expected even if the surrogate is a poor proxy for EM performance. To support the claim that test-time compute improves design accuracy, the authors should (a) define the surrogate score explicitly, (b) report its correlation with openEMS error on a held-out set of generated designs, and (c) re-plot Figure 3 using full-simulation error (or a validated proxy) for the selected designs at each budget. Alternatively, a direct-optimization baseline that optimizes (L, W, p) with a surrogate or with openEMS would show whether the improvement is specific to the proposed framework.
  2. [Section 6.3, Figure 4] The full-simulation validation covers only three target curves, and one of them - the 5.0 GHz target with d = -10 dB - fails to meet the depth requirement even with the larger 10x20 search budget. This failure is acknowledged in the text, but it directly undercuts the claim that 'devoting more compute to search ... yields a more accurate result' in all cases. The paper should report quantitative metrics for all three targets at both budgets (e.g., minimum |S11| within the target band, bandwidth achieved, and MSE against the target curve). It should also analyze why the 5.0 GHz target fails: is the target out of the distribution of realizable responses, does the surrogate over-rank the selected design, or is the search budget still insufficient? Without such analysis, the reader cannot assess whether the failure is a fundamental limitation or a fixable artifact.
  3. [Section 5.5, Oracle Scorer] The Oracle Scorer, which runs full EM simulation and computes MSE against the target, is defined in Section 5.5 but never used in any experiment. Since the oracle is the ground truth for design quality, comparing the surrogate-ranked selections with oracle-ranked selections on the same three target curves would directly address whether the surrogate over-ranks designs (as the 5.0 GHz failure suggests). Even if the oracle is computationally expensive, a small-scale comparison (e.g., for the 200 designs in the 10x20 budget) would be feasible and would substantially strengthen the paper's evidence. The authors should add this comparison or explicitly justify why the oracle is not used for validation.
minor comments (6)
  1. [Abstract] The abstract says 'adversarial training process', but the method uses an adversarial predictor (a small network) that attempts to infer the condition from the latent code, not a generative adversarial network. Rephrase to 'adversarially disentangled training' or similar to avoid ambiguity.
  2. [Section 6.2, Figure 3] The shaded regions in Figure 3 indicate variability across runs, but the number of runs is not reported. State how many independent runs were averaged and whether the same target set was used for each run.
  3. [Table 1] For the 5.0 GHz target, the generated patch dimensions change substantially between budgets (L = 15.7 mm for 1x1 vs L = 29.0 mm for 10x20, and W from 25.5 mm to 46.9 mm). The paper does not discuss whether either design is physically reasonable or manufacturable, despite the stated auxiliary objective of manufacturability. A brief comment would be helpful.
  4. [Section 6.3] The sentence 'In each case, it seems that devoting more compute ... yields a more accurate result' uses the hedge 'seems'. Since the preceding sentence and Figure 4 show only three examples, please either provide quantitative support or temper the claim to match the evidence.
  5. [Section 4] The dataset description says the grid is 'augmented' with an algorithm to sample additional triplets inside the convex hull, but no details of this algorithm or the resulting distribution are given. Please clarify how the final 1292 designs are split into training/validation/test sets, if at all.
  6. [References] Reference [9] is incomplete: the conference name and page numbers are missing. Please complete the bibliographic entry.

Circularity Check

1 steps flagged · score 5.0 of 10

Figure 3 uses the surrogate score as both the selection objective and the accuracy metric, so best-of-N improvements are tautological; the openEMS checks in Figure 4 provide only partial independent support.

  1. fitted input called prediction [Section 6.2, Figure 3; Section 6.3, first paragraph]
    "In both investigations, we consider three distinct target response functions, and at each search configuration, record the lowest Surrogate Scorer score from the pool of generated antenna designs... We find that as the pool of antenna designs ˜x grows through the amount of candidate frequency response curves, both initialization strategies yield better and more consistent predictions. ... 10 candidate curves are generated, from which 20 design curves are sampled (pool of 200), and the design with lowest Surrogate Scorer score is chosen."

    The test-time selection rule is argmin over the surrogate score ('the design with lowest Surrogate Scorer score is chosen'), and the reported scaling metric is the minimum of that same surrogate score over an expanding pool. For any fixed scoring function, the minimum over a larger random pool is stochastically non-increasing, so the monotone curves in Figure 3 are guaranteed by the selection procedure itself and do not independently show that generated designs are electromagnetically closer to the target. The surrogate is trained on the same openEMS data, and no correlation between surrogate score and true simulated |S11| error is reported. Thus the primary quantitative evidence for 'test-time compute improves accuracy' reduces to best-of-N minimization of the fitted scoring function.

full rationale

The paper's inverse-design pipeline is otherwise self-contained: the target curves are user-specified (Eqs. 2-3), the two VAEs are trained on an openEMS dataset (Sections 4-5), and the final validation in Figure 4 uses full-wave openEMS simulations external to the fitted models. The one concrete circular step is confined to Section 6.2 / Figure 3: the test-time selection rule minimizes the Surrogate Scorer, and the plotted quantity is the minimum of that same scorer over expanding candidate pools. Best-of-N selection of any scorer makes the minimum non-increasing in pool size by construction, so Figure 3's scaling curves cannot, by themselves, establish that increased test-time compute yields designs with better true electromagnetic response. The independent openEMS check covers only three targets and, as the paper concedes, the 5.0 GHz design fails the depth requirement even after the 10x20 search, consistent with the surrogate over-ranking that this self-referential metric cannot detect. No self-citation chain, imported uniqueness theorem, or ansatz-via-citation was found. Overall, the central framework claim is not fully circular because an external simulator is used for final illustration, but the primary quantitative scaling evidence is partly self-referential, giving a score of 5.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The free parameters are architecture and hyperparameter choices, most importantly the unspecified frequency mask used in every objective. The assumptions that carry the most weight are dataset representativeness, the fidelity of openEMS labels, and the reliability of the surrogate scorer as a selection criterion.

free parameters (5)
  • VAE beta (KLD weight) = 0.016
    Chosen experimentally to balance reconstruction and generative quality; annealed over the first 100 epochs.
  • CVAE disentanglement weight eta = 0.1
    Weights the adversarial predictor that removes response information from the geometry latent code; set by hand.
  • Surrogate beta-NLL beta = 0.5
    Hyperparameter for the heteroscedastic Gaussian likelihood used by the surrogate scorer.
  • Frequency mask = not specified
    Regions outside target frequency bands are masked in the latent search objective and in both scorers, but the exact mask construction is not defined. This choice shapes every quantitative result.
  • Latent dimensions = zy=64, zx=16
    Architecture choices made by hand; they affect the capacity of both stages but are not treated as fitted parameters.
assumptions (4)
  • domain assumption openEMS FDTD simulation of S11 for these geometries is accurate enough to serve as ground truth
    All labels and the Oracle Scorer are based on openEMS; no measurement or fabrication validation is provided.
  • domain assumption The 1292 simulated designs span the practically relevant rectangular patch antenna design space
    Bounds L in [7.5, 52.5], W/L in [0.8, 2], and p in (-L/2, 0] are chosen without a coverage or convergence analysis.
  • ad hoc to paper A product of Lorentzian notches is an adequate representation of target antenna responses
    Targets are defined by Equation 2; real antennas have higher harmonics, which the framework treats as masked regions rather than modeling explicitly.
  • domain assumption Adversarial disentanglement removes y information from zx sufficiently for controlled generation
    Decorrelation is stated as achieved, but no quantitative measure of mutual information or conditioning strength is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving Generative Inverse Design of Rectangular Patch Antennas with Test Time Optimization." pith.science (2026). https://pith.science/paper/FH4LBPKV

@misc{pith2026250518188,
  author       = {Pith},
  title        = {Pith review of: Improving Generative Inverse Design of Rectangular Patch Antennas with Test Time Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FH4LBPKV}},
  note         = {Machine review of arXiv:2505.18188}
}
read the original abstract

We propose a two-stage deep learning framework for the inverse design of rectangular patch antennas. Our approach leverages generative modeling to learn a latent representation of antenna frequency response curves and conditions a subsequent generative model on these responses to produce feasible antenna geometries. We further demonstrate that leveraging search and optimization techniques at test-time improves the accuracy of the generated designs and enables consideration of auxiliary objectives such as manufacturability. Our approach generalizes naturally to different design criteria, and can be easily adapted to more complex geometric design spaces.

Figures

Figures reproduced from arXiv: 2505.18188 by the authors.

Figure 1
Figure 1. Configuration of a Rectangular Patch Antenna fed via coaxial line through the ground plane. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Overview of our two-stage generative inverse design framework. Stage 1 learns a latent representation of [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Scaling performance as the number of curves (left) and the number of designs per curve (right) is increased. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Comparison of the idealised target S11 curve y ∗ (black dashed), the dominant-mode analytic resonance fr,TM10 (dotted vertical), and the simulated S11 of designs x˜ generated with two test-time compute budgets. Blue = 1 curve × 1 design, red = 10 curves × 20 designs. N…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 14 canonical work pages

  1. [1]

    Antenna theory

    Constantine A Balanis. Antenna theory. Wiley-Blackwell, Hoboken, NJ, 4 edition, January 2016

    work page 2016

  2. [2]

    Searching latent program spaces, 2024

    Clément Bonnet and Matthew V Macfarlane. Searching latent program spaces, 2024. URL https://arxiv. org/abs/2411.08706

    arXiv 2024

  3. [3]

    Elsherbeni, and Veysel Demir

    Yiming Chen, Atef Z. Elsherbeni, and Veysel Demir. Machine learning for microstrip patch antenna design: Observations and recommendations. In 2022 United States National Committee of URSI National Radio Science Meeting (USNC-URSI NRSM), pages 256–257, 2022. doi: 10.23919/USNC-URSINRSM57467.2022.9881476

    work page arXiv 2022

  4. [4]

    Elsherbeni, and Veysel Demir

    Yiming Chen, Atef Z. Elsherbeni, and Veysel Demir. Machine learning design of printed patch antenna. In 2022 IEEE International Symposium on Antennas and Propagation and USNC-URSI Radio Science Meeting (AP-S/URSI), pages 201–202, 2022. doi: 10.1109/AP-S/USNC-URSI47032.2022.9887043

    arXiv 2022

  5. [5]

    On the measure of intelligence, 2019

    François Chollet. On the measure of intelligence, 2019. URL https://arxiv.org/abs/1911.01547

    arXiv 2019

  6. [6]

    A review on the design and optimization of antennas using machine learning algorithms and techniques

    Hilal El Misilmani, Tarek Naous, and Salwa Al Khatib. A review on the design and optimization of antennas using machine learning algorithms and techniques. International Journal of RF and Microwave Computer-Aided Engineering, 2020, 07 2020. doi: 10.1002/mmce.22356

    doi:10.1002/mmce.22356 2020

  7. [7]

    Gansynth: Adversarial neural audio synthesis

    Jesse Engel, Kumar Krishna Agrawal, Shuo Chen, Ishaan Gulrajani, and Adam Roberts. Gansynth: Adversarial neural audio synthesis. In ICLR, 2019

    work page 2019

  8. [8]

    Wei, David Duvenaud, José Miguel Hernández-Lobato, Benjamín Sánchez- Lengeling, Dennis Sheberla, Jorge Aguilera-Iparraguirre, Timothy D

    Rafael Gómez-Bombarelli, Jennifer N. Wei, David Duvenaud, José Miguel Hernández-Lobato, Benjamín Sánchez- Lengeling, Dennis Sheberla, Jorge Aguilera-Iparraguirre, Timothy D. Hirzel, Ryan P. Adams, and Alán Aspuru- Guzik. Automatic chemical design using a data-driven continuous representation of molecules. ACS Central Science, 4(2):268–276, Feb 2018. ISSN ...

    doi:10.1021/acscentsci.7b00572 2018

Show all 25 references
  1. [9]

    Equations for microstrip circuit design

    Erik Hammerstad. Equations for microstrip circuit design. 1975 5th European Microwave Conference, pages 268–272, 1975. URL https://api.semanticscholar.org/CorpusID:23699478. 9 A PREPRINT - MAY 28, 2025

    1975

  2. [10]

    In ICLR, 2017

    Irina Higgins, Loic Matthey, Arka Pal, et al.β-vae: Learning basic visual concepts with a constrained variational framework. In ICLR, 2017

    2017

  3. [11]

    The Finite Element Method in Electromagnetics

    Jianming Jin. The Finite Element Method in Electromagnetics . Wiley - IEEE. John Wiley & Sons, Nashville, TN, 3 edition, March 2014

    2014

  4. [12]

    John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, Alex Bridgland, Clemens Meyer, Simon A. A. Kohl, Andrew J. Ballard, Andrew Cowie, Bernardino Romera-Paredes, Stani...

    2021

  5. [13]

    Auto-encoding variational bayes

    Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014

    2014

  6. [14]

    openems - open electromagnetic field solver, accessed 2024

    Thorsten Liebig. openems - open electromagnetic field solver, accessed 2024. URL https://www.openEMS.de

    2024

  7. [15]

    Adversarial autoencoders

    Alireza Makhzani, Jonathon Shlens, Navdeep Jaitly, Ian Goodfellow, and Brendan Frey. Adversarial autoencoders. arXiv:1511.05644, 2015

    2015 arXiv

  8. [16]

    Schoenholz, Muratahan Aykol, Gowoon Cheon, and Ekin Dogus Cubuk

    Amil Merchant, Simon Batzner, Samuel S. Schoenholz, Muratahan Aykol, Gowoon Cheon, and Ekin Dogus Cubuk. Scaling deep learning for materials discovery. Nature, 624(7990):80–85, November 2023. ISSN 1476-4687. doi: 10.1038/s41586-023-06735-9. URL http://dx.doi.org/10.1038/s41586...

    2023 doi

  9. [17]

    Plug & play generative networks: Conditional iterative generation of images in latent space, 2017

    Anh Nguyen, Jeff Clune, Yoshua Bengio, Alexey Dosovitskiy, and Jason Yosinski. Plug & play generative networks: Conditional iterative generation of images in latent space, 2017. URL https://arxiv.org/abs/ 1612.00005

    2017 arXiv

  10. [18]

    Mathematical discoveries from program search with large language models

    Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M Pawan Kumar, Emilien Dupont, Francisco J R Ruiz, Jordan S Ellenberg, Pengming Wang, Omar Fawzi, Pushmeet Kohli, and Alhussein Fawzi. Mathematical discoveries from program search with large la...

    2024

  11. [19]

    On the pitfalls of heteroscedastic uncertainty estimation with probabilistic neural networks, 2022

    Maximilian Seitzer, Arash Tavakoli, Dimitrije Antic, and Georg Martius. On the pitfalls of heteroscedastic uncertainty estimation with probabilistic neural networks, 2022. URL https://arxiv.org/abs/2203.09168

    2022 arXiv

  12. [20]

    Machine learning techniques for optimizing design of double t-shaped monopole antenna

    Yashika Sharma, Hao Helen Zhang, and Hao Xin. Machine learning techniques for optimizing design of double t-shaped monopole antenna. IEEE Transactions on Antennas and Propagation , 68(7):5658–5663, 2020. doi: 10.1109/TAP.2020.2966051

    2020

  13. [21]

    Mastering the game of go with deep neural networks and tree search

    David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George van den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, Sander Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy Lillicrap, Madelei...

    2016

  14. [22]

    Learning structured output representation using deep conditional generative models

    Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. In C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing Systems , volume 28. Curran Associa...

    2015

  15. [23]

    Computational electrodynamics

    Allen Taflove and Susan Hagness. Computational electrodynamics. Artech House antennas and propagation library. Artech House, Norwood, MA, 3 edition, May 2005

    2005

  16. [24]

    Solving olympiad geometry without human demonstrations

    Trieu H Trinh, Yuhuai Wu, Quoc V Le, He He, and Thang Luong. Solving olympiad geometry without human demonstrations. Nature, 625(7995):476–482, January 2024

    2024

  17. [25]

    Automated antenna design via domain knowledge-informed reinforcement learning and imitation learning

    Zhaohui Wei, Zhao Zhou, Peng Wang, Jian Ren, Yingzeng Yin, Gert Frølund Pedersen, and Ming Shen. Automated antenna design via domain knowledge-informed reinforcement learning and imitation learning. IEEE Transactions on Antennas and Propagation, 71(7):5549–5557, 2023. doi: 10....

    2023

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.