pith. sign in

arxiv: 2606.10294 · v1 · pith:VRQXUVN3new · submitted 2026-06-09 · 💻 cs.LG · cs.AI· cs.AR· cs.NE· physics.comp-ph

LLM-Guided Neural Architecture Search for Robust Co-Design of Physical Neural Networks

Tyler King , Timothee Leleu This is my paper

Pith reviewed 2026-06-27 13:44 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.ARcs.NEphysics.comp-ph
keywords neural architecture searchLLM-guided optimizationunconventional hardwareoptical neural networkshardware co-designrobustness to non-idealitiesenergy efficiencyMZI hardware
0
0 comments X

The pith

UH-NAS uses language models as evolutionary operators in a hardware-agnostic search to co-optimize neural network accuracy and energy on physical platforms.

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

The paper introduces UH-NAS, a framework that treats different hardware platforms as swappable backends during neural architecture search, with large language models serving as the operators that evolve candidate networks. It seeks architectures that jointly maximize task performance while respecting platform-specific energy costs, physical non-idealities, and precision limits. A sympathetic reader would care because most existing NAS methods are locked to one hardware family, making cross-platform evaluation and adaptation to new devices difficult. By keeping the search algorithm fixed and only exchanging the backend models and simulators, the method supports direct system-level comparisons. Experiments on optical MZI hardware indicate that the resulting architectures are more diverse and maintain better robustness under non-idealities than both standard NAS baselines and prior LLM-assisted approaches.

Core claim

UH-NAS is a hardware-agnostic NAS framework that integrates language models as evolutionary operators to co-optimize accuracy and inference energy. Hardware is exposed as a swappable backend equipped with per-platform energy models, physical constraints, and non-ideality simulators, so the same search procedure works across backends. When tested on optical MZI hardware, UH-NAS produces architectures that are more diverse and robust to non-idealities than conventional baselines while also outperforming existing LLM-to-NAS methods.

What carries the argument

Unconventional Hardware Neural Architecture Search (UH-NAS) framework, in which LLMs act as evolutionary operators and hardware is handled through interchangeable energy models and non-ideality simulators.

If this is right

  • The identical search procedure can be applied to multiple hardware platforms by swapping only the backend models, enabling fair cross-platform comparisons.
  • Discovered architectures show improved robustness when non-idealities such as noise or precision loss are present.
  • Performance exceeds both conventional NAS methods and earlier LLM-guided NAS techniques on the evaluated optical platform.
  • Co-design of architecture with hardware-specific constraints becomes feasible without rewriting the search algorithm for each new platform.

Where Pith is reading between the lines

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

  • The approach could shorten the time needed to adapt neural networks to newly emerging hardware technologies by reusing the same search logic.
  • If simulator fidelity improves, the method might produce designs that transfer to physical devices with little additional fine-tuning.
  • Similar LLM-driven evolutionary search might be applied to related co-design problems such as circuit layout or sensor placement.

Load-bearing premise

The per-platform energy models and non-ideality simulators must accurately capture real hardware behavior so that architectures found in simulation remain robust when transferred to physical devices.

What would settle it

Deploying the architectures discovered by UH-NAS on actual optical MZI hardware and checking whether they exhibit measurably higher robustness to real non-idealities and greater architectural diversity than the baselines.

Figures

Figures reproduced from arXiv: 2606.10294 by Timothee Leleu, Tyler King.

Figure 1
Figure 1. Figure 1: Pareto fronts of MNIST validation error (%) versus energy per inference (nJ) obtained with UH-NAS across multiple hardware backends, including GPU (Blackwell B200, FP16), CPU (Xeon 8380, FP32), and MZI mesh–based optical systems under both realistic and worst-case non-idealities. Optical hardware op￾erates at substantially lower energy but with higher error rates due to device non-idealities; UH-NAS identi… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the UH-NAS framework. Each generation begins with hardware-specific inputs (sub-operation energy costs, seed architectures, and non-ideality parameters). A selection operator (NSGA-II with MAP-Elites diversity) chooses parents, which the LLM mutates or recombines via hardware-specific prompts. Candidates are trained and evaluated for accuracy, energy, and robustness under non-idealities. Result… view at source ↗
Figure 3
Figure 3. Figure 3: Ablation comparing UH-NAS (with LLM) and UH-NAS without LLM on the MZI optical backend under worst-case non￾idealities. Without LLM-guided generation, the search degenerates to minor seed architecture modifications, producing only 26 unique designs out of 250 evaluations and a restricted Pareto front. With the LLM, UH-NAS explores 203 unique architectures, yielding broader coverage and lower validation err… view at source ↗
Figure 4
Figure 4. Figure 4: Robustness of UH-NAS–discovered architectures across non-ideality regimes on the MZI optical backend. The Pareto front optimized under worst-case noise (left) and realistic noise (right) is re-evaluated at four non-ideality levels: none, realistic, medium, and worst-case, with the shaded gray region highlighting the robustness gap of the other training environment (worst-case comparison on the left, realis… view at source ↗
Figure 5
Figure 5. Figure 5: Spearman rank correlation between 13 zero-cost (ZC) proxies and validation accuracy on 200 randomly drawn NAS-Bench￾201/CIFAR-10 architectures, evaluated under clean and worst-case MZI ONN non-idealities (σϕ = 0.05, σγ = 0.005, crosstalk= 0.15). Under clean conditions, most proxies show positive correlations consistent with prior literature. Under worst-case non-idealities, nearly all proxies exhibit negat… view at source ↗
Figure 6
Figure 6. Figure 6: Best observed MNIST accuracy (worst-case non￾idealities) over 30 NAS generations, averaged over 3 runs, for three LLM backbones. GPT-4.1 and Gemini 3.1 Flash-Lite con￾verge to approximately 97% accuracy within 10–15 generations, while GPT-4.1-nano shows limited improvement due to its reduced capacity for generating effective system prompts and diverse archi￾tectures. sensitivity arises due to phase error (… view at source ↗
Figure 7
Figure 7. Figure 7: Log-log correlation between analytical energy estimates and real RAPL energy validation on an AWS c6i.metal in￾stance with Xeon Platinum 8375C (Ice Lake-SP), 2.90 GHz, single core. Only models along the Pareto frontier were tested, giving a log-space correlation of r 2 = 0.828. Averaged over 50,000 forward passes. Xeon silicon AWS c6i.metal instances, utilizing Intel’s RAPL MSR PKG ENERGY STATUS counter av… view at source ↗
Figure 8
Figure 8. Figure 8: Example synthesized system prompt from the meta-learning phase (GPT-4.1-nano). The underlined sentence contains a factual error: namely that optical MACs at∼20 fJ are cheap and should be maximized, not minimized due to their low computational cost. 14 [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Example synthesized system prompt from the meta-learning phase (GPT-4.1). Unlike the GPT-4.1-nano prompt ( [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Example synthesized system prompt from the meta-learning phase (Gemini 3.1 Flash-Lite). This prompt is notably more concise than the GPT-4.1 variant ( [PITH_FULL_IMAGE:figures/full_fig_p016_10.png] view at source ↗
read the original abstract

Deploying neural networks on unconventional hardware demands architectures that co-optimize task accuracy and platform-specific constraints such as energy cost, physical non-idealities, and numerical precision. Existing neural architecture search (NAS) methods are typically tailored to a single hardware family, limiting cross-platform comparison and generalization. We introduce Unconventional Hardware Neural Architecture Search (UH-NAS), a hardware-agnostic, LLM-guided NAS framework that integrates language models as evolutionary operators to co-optimize accuracy and inference energy. By exposing hardware as a swappable backend with per-platform energy models, physical constraints, and non-ideality simulators, UH-NAS enables fair system-level comparisons across various backends without modifying the search algorithm. Tested on optical MZI hardware, UH-NAS discovers more diverse, robust architectures than conventional baselines while outperforming existing LLM-to-NAS approaches. Additional ablations on architecture robustness under non-idealities and the role of system prompts highlight the importance of architecture-hardware co-design for emerging computing platforms.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 1 minor

Summary. The paper introduces Unconventional Hardware Neural Architecture Search (UH-NAS), a hardware-agnostic LLM-guided NAS framework that uses language models as evolutionary operators to co-optimize task accuracy and inference energy across swappable hardware backends. Each backend supplies per-platform energy models, physical constraints, and non-ideality simulators, enabling cross-platform comparisons without altering the search algorithm. Evaluated on optical MZI hardware, the method is claimed to yield more diverse and robust architectures than conventional NAS baselines and prior LLM-to-NAS approaches, with supporting ablations on non-ideality robustness and system-prompt effects.

Significance. If the empirical claims are substantiated by quantitative results, the work could advance hardware-aware NAS for emerging physical platforms by demonstrating a unified search procedure that incorporates platform-specific non-idealities while remaining algorithmically hardware-agnostic.

minor comments (1)
  1. [Abstract] The abstract reports outperformance and robustness ablations but supplies no quantitative results, error bars, dataset sizes, or statistical tests.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their summary of the manuscript and for noting the potential significance of UH-NAS for hardware-aware NAS on unconventional platforms. The recommendation is listed as uncertain, yet the report contains no specific major comments to address. We remain available to provide additional quantitative details, ablations, or clarifications on the empirical claims if requested by the editor or referee.

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper introduces UH-NAS as an LLM-guided NAS framework with swappable hardware backends for co-optimizing accuracy and energy. All central claims rest on empirical comparisons of discovered architectures against baselines on optical MZI models, with no equations, fitted parameters, or derivations presented that reduce to self-definition, renamed inputs, or self-citation chains. The method treats energy models and simulators as external oracles and reports search outcomes under that assumption; the argument does not contain any load-bearing step that is equivalent to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are stated. Full manuscript would be needed to audit energy models, LLM prompt templates, or simulator assumptions.

pith-pipeline@v0.9.1-grok · 5712 in / 1101 out tokens · 22025 ms · 2026-06-27T13:44:52.082989+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

81 extracted references · 6 canonical work pages · 1 internal anchor

  1. [1]

    Neural architecture search: A survey.Journal of Machine Learning Research, 20(55):1–21, 2019

    Thomas Elsken, Jan Hendrik Metzen, and Frank Hut- ter. Neural architecture search: A survey.Journal of Machine Learning Research, 20(55):1–21, 2019

    2019

  2. [2]

    Mnasnet: Platform-aware neural architecture search for mobile

    Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Va- sudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. InProceedings of the IEEE/CVF conference on computer vision and pattern recogni- tion, pages 2820–2828, 2019

    2019

  3. [3]

    Hw-nas-bench: Hardware-aware neural architecture search benchmark

    Chaojian Li, Zhongzhi Yu, Yonggan Fu, Yongan Zhang, Yang Zhao, Haoran You, Qixuan Yu, Yue Wang, and Yingyan Celine Lin. Hw-nas-bench: Hardware-aware neural architecture search benchmark. arXiv preprint arXiv:2103.10584, 2021

    work page arXiv 2021

  4. [4]

    Flash: F ast neura l a rchitecture s earch with h ardware optimization.ACM Transactions on Embedded Computing Systems (TECS), 20(5s):1– 26, 2021

    Guihong Li, Sumit K Mandal, Umit Y Ogras, and Radu Marculescu. Flash: F ast neura l a rchitecture s earch with h ardware optimization.ACM Transactions on Embedded Computing Systems (TECS), 20(5s):1– 26, 2021

    2021

  5. [5]

    Device- circuit-architecture co-exploration for computing-in- memory neural accelerators.IEEE Transactions on Computers, 70(4):595–605, 2020

    Weiwen Jiang, Qiuwen Lou, Zheyu Yan, Lei Yang, Jingtong Hu, Xiaobo Sharon Hu, and Yiyu Shi. Device- circuit-architecture co-exploration for computing-in- memory neural accelerators.IEEE Transactions on Computers, 70(4):595–605, 2020

    2020

  6. [6]

    Nas4rram: neural network architec- ture search for inference on rram-based accelerators

    Zhihang Yuan, Jingze Liu, Xingchen Li, Longhao Yan, Haoxiang Chen, Bingzhe Wu, Yuchao Yang, and Guangyu Sun. Nas4rram: neural network architec- ture search for inference on rram-based accelerators. Science China Information Sciences, 64(6):160407, 2021

    2021

  7. [7]

    Analognets: Ml-hw co-design of noise-robust tinyml models and always-on analog compute-in-memory ac- celerator.arXiv preprint arXiv:2111.06503, 2021

    Chuteng Zhou, Fernando Garcia Redondo, Ju- lian B ¨uchel, Irem Boybat, Xavier Timoneda Co- mas, SR Nandakumar, Shidhartha Das, Abu Sebas- tian, Manuel Le Gallo, and Paul N Whatmough. Analognets: Ml-hw co-design of noise-robust tinyml models and always-on analog compute-in-memory ac- celerator.arXiv preprint arXiv:2111.06503, 2021

    work page arXiv 2021

  8. [8]

    Nax: neural architecture and mem- ristive xbar based accelerator co-design

    Shubham Negi, Indranil Chakraborty, Aayush Ankit, and Kaushik Roy. Nax: neural architecture and mem- ristive xbar based accelerator co-design. InProceed- ings of the 59th ACM/IEEE Design Automation Con- ference, pages 451–456, 2022

    2022

  9. [9]

    Analognas-bench: A nas benchmark for analog in- memory computing.arXiv preprint arXiv:2506.18495, 2025

    Aniss Bessalah, Hatem Mohamed Abdelmoumen, Karima Benatchba, and Hadjer Benmeziane. Analognas-bench: A nas benchmark for analog in- memory computing.arXiv preprint arXiv:2506.18495, 2025

    work page arXiv 2025

  10. [10]

    Analognas: A neural network design framework for accurate inference with analog in-memory computing

    Hadjer Benmeziane, Corey Lammie, Irem Boybat, Malte Rasch, Manuel Le Gallo, Hsinyu Tsai, Ra- machandran Muralidhar, Smail Niar, Ouarnoughi Hamza, Vijay Narayanan, et al. Analognas: A neural network design framework for accurate inference with analog in-memory computing. In2023 IEEE Interna- tional Conference on Edge Computing and Communi- cations (EDGE), ...

    2023

  11. [11]

    Uncertainty modeling of emerging device based computing-in-memory neural accelerators with application to neural architecture search

    Zheyu Yan, Da-Cheng Juan, Xiaobo Sharon Hu, and Yiyu Shi. Uncertainty modeling of emerging device based computing-in-memory neural accelerators with application to neural architecture search. InProceed- ings of the 26th Asia and South Pacific Design Automa- tion Conference, pages 859–864, 2021

    2021

  12. [12]

    Xplorenas: Explore adversar- ially robust and hardware-efficient neural architectures for non-ideal xbars.ACM Transactions on Embedded Computing Systems, 22(4):1–17, 2023

    Abhiroop Bhattacharjee, Abhishek Moitra, and Priyadarshini Panda. Xplorenas: Explore adversar- ially robust and hardware-efficient neural architectures for non-ideal xbars.ACM Transactions on Embedded Computing Systems, 22(4):1–17, 2023

    2023

  13. [13]

    Quan- tumnas: Noise-adaptive search for robust quantum circuits

    Hanrui Wang, Yongshan Ding, Jiaqi Gu, Yujun Lin, David Z Pan, Frederic T Chong, and Song Han. Quan- tumnas: Noise-adaptive search for robust quantum circuits. In2022 IEEE International Symposium on High-Performance Computer Architecture (HPCA), pages 692–708. IEEE, 2022

    2022

  14. [14]

    Adept: Automatic differentiable design of photonic tensor cores

    Jiaqi Gu, Hanqing Zhu, Chenghao Feng, Zixuan Jiang, Mingjie Liu, Shuhan Zhang, Ray T Chen, and David Z Pan. Adept: Automatic differentiable design of photonic tensor cores. InProceedings of the 59th ACM/IEEE Design Automation Conference, pages 937– 942, 2022

    2022

  15. [15]

    Harris, Scott Skirlo, Mihika Prabhu, Tom Baehr-Jones, Michael Hochberg, Xin Sun, Shijie Zhao, Hugo Larochelle, Dirk Englund, and Marin Soljaˇci´c

    Yichen Shen, Nicholas C. Harris, Scott Skirlo, Mihika Prabhu, Tom Baehr-Jones, Michael Hochberg, Xin Sun, Shijie Zhao, Hugo Larochelle, Dirk Englund, and Marin Soljaˇci´c. Deep learning with coherent nanopho- tonic circuits.Nature Photonics, 11(7):441–446, 2017

    2017

  16. [16]

    Large-scale optical neural networks based on photoelectric multiplication

    Ryan Hamerly, Liane Bernstein, Alexander Sludds, Marin Soljaˇci´c, and Dirk Englund. Large-scale optical neural networks based on photoelectric multiplication. Physical Review X, 9(2), 2019

    2019

  17. [17]

    Photonics for artificial intelligence and neuromorphic computing

    Bhavin J Shastri, Alexander N Tait, Thomas Fer- reira de Lima, Wolfram HP Pernice, Harish Bhaskaran, C David Wright, and Paul R Prucnal. Photonics for artificial intelligence and neuromorphic computing. Nature Photonics, 15(2):102–114, 2021

    2021

  18. [18]

    Evo- prompting: Language models for code-level neural architecture search.Advances in neural information processing systems, 36:7787–7817, 2023

    Angelica Chen, David Dohan, and David So. Evo- prompting: Language models for code-level neural architecture search.Advances in neural information processing systems, 36:7787–7817, 2023. 9 Unconventional Hardware Neural Architecture Search

    2023

  19. [19]

    Can gpt-4 perform neural architecture search?arXiv preprint arXiv:2304.10970, 2023

    Mingkai Zheng, Xiu Su, Shan You, Fei Wang, Chen Qian, Chang Xu, and Samuel Albanie. Can gpt-4 perform neural architecture search?arXiv preprint arXiv:2304.10970, 2023

    work page arXiv 2023

  20. [20]

    Llmatic: Neural architecture search via large language models and quality diversity optimization

    Muhammad Umair Nasir, Sam Earle, Julian Togelius, Steven James, and Christopher Cleghorn. Llmatic: Neural architecture search via large language models and quality diversity optimization. InProceedings of the Genetic and Evolutionary Computation Confer- ence, GECCO ’24, page 1110–1118. ACM, 2024

    2024

  21. [21]

    RZ-NAS: Enhancing LLM-guided neural architecture search via reflective zero-cost strategy

    Zipeng Ji, Guanghui Zhu, Chunfeng Yuan, and Yi- hua Huang. RZ-NAS: Enhancing LLM-guided neural architecture search via reflective zero-cost strategy. In Aarti Singh, Maryam Fazel, Daniel Hsu, Simon Lacoste-Julien, Felix Berkenkamp, Tegan Maharaj, Kiri Wagstaff, and Jerry Zhu, editors,Proceedings of the 42nd International Conference on Machine Learn- ing, ...

    2025

  22. [22]

    Llm- nas: Llm-driven hardware-aware neural architecture search, 2025

    Hengyi Zhu, Grace Li Zhang, and Shaoyi Huang. Llm- nas: Llm-driven hardware-aware neural architecture search, 2025

    2025

  23. [23]

    Fouda, Hadjer Benmeziane, Kaoutar El Maghraoui, Abu Sebastian, Wei D

    Olga Krestinskaya, Mohammed E. Fouda, Hadjer Benmeziane, Kaoutar El Maghraoui, Abu Sebastian, Wei D. Lu, Mario Lanza, Hai Li, Fadi Kurdahi, Suhaib A. Fahmy, Ahmed Eltawil, and Khaled N. Salama. Neural architecture search for in-memory computing-based deep learning accelerators.Nature Reviews Electrical Engineering, 1(6):374–390, May 2024

    2024

  24. [24]

    Luxnas: A coherent photonic neural network powered by neural architecture search

    Amin Shafiee, Febin Sunny, Sudeep Pasricha, and Mahdi Nikdast. Luxnas: A coherent photonic neural network powered by neural architecture search. In CLEO 2025. Optica Publishing Group, 2025

    2025

  25. [25]

    QuantumDARTS: Differentiable quantum architecture search for variational quantum algorithms

    Wenjie Wu, Ge Yan, Xudong Lu, Kaisen Pan, and Junchi Yan. QuantumDARTS: Differentiable quantum architecture search for variational quantum algorithms. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scar- lett, editors,Proceedings of the 40th International Conference on Machine Learning, volume 202 ofPro- c...

    2023

  26. [26]

    Quantum architecture search: A survey

    Darya Martyniuk, Johannes Jung, and Adrian Paschke. Quantum architecture search: A survey. In2024 IEEE International Conference on Quantum Computing and Engineering (QCE), page 1695–1706. IEEE, 2024

    2024

  27. [27]

    AutoSNN: Towards energy-efficient spiking neural networks

    Byunggook Na, Jisoo Mok, Seongsik Park, Dongjin Lee, Hyeokjun Choe, and Sungroh Yoon. AutoSNN: Towards energy-efficient spiking neural networks. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sabato, edi- tors,Proceedings of the 39th International Conference on Machine Learning, volume 162 ofProceedings of Machine ...

    2022

  28. [28]

    Efficient spiking neural network design via neural architecture search

    Jiaqi Yan, Qianhui Liu, Malu Zhang, Lang Feng, De Ma, Haizhou Li, and Gang Pan. Efficient spiking neural network design via neural architecture search. Neural Networks, 173:106172, May 2024

    2024

  29. [29]

    Mathematical discoveries from pro- gram search with large language models.Nature, 625(7995):468–475, 2024

    Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M Pawan Kumar, Emilien Dupont, Francisco JR Ruiz, Jordan S Ellenberg, Pengming Wang, Omar Fawzi, et al. Mathematical discoveries from pro- gram search with large language models.Nature, 625(7995):468–475, 2024

    2024

  30. [30]

    AlphaEvolve: A coding agent for scientific and algorithmic discovery

    Alexander Novikov, Ngˆan V ˜u, Marvin Eisenberger, Emilien Dupont, Po-Sen Huang, Adam Zsolt Wagner, Sergey Shirobokov, Borislav Kozlovskii, Francisco JR Ruiz, Abbas Mehrabian, et al. Alphaevolve: A coding agent for scientific and algorithmic discovery.arXiv preprint arXiv:2506.13131, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  31. [31]

    Ruiz, Julian Schrittwieser, Grzegorz Swirszcz, et al

    Alhussein Fawzi, Matej Balog, Aja Huang, Thomas Hubert, Bernardino Romera-Paredes, Moham- madamin Barekatain, Alexander Novikov, Francisco J R. Ruiz, Julian Schrittwieser, Grzegorz Swirszcz, et al. Discovering faster matrix multiplication algorithms with reinforcement learning.Nature, 610(7930):47–53, 2022

    2022

  32. [32]

    Evolution of heuristics: Towards efficient au- tomatic algorithm design using large language model

    Fei Liu, Xialiang Tong, Mingxuan Yuan, Xi Lin, Fu Luo, Zhenkun Wang, Zhichao Lu, and Qingfu Zhang. Evolution of heuristics: Towards efficient au- tomatic algorithm design using large language model. arXiv preprint arXiv:2401.02051, 2024

    work page arXiv 2024

  33. [33]

    Reevo: Large language models as hyper-heuristics with reflective evolution.Advances in neural information processing systems, 37:43571– 43608, 2024

    Haoran Ye, Jiarui Wang, Zhiguang Cao, Federico Berto, Chuanbo Hua, Haeyeon Kim, Jinkyoo Park, and Guojie Song. Reevo: Large language models as hyper-heuristics with reflective evolution.Advances in neural information processing systems, 37:43571– 43608, 2024

    2024

  34. [34]

    K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan. A fast and elitist multiobjective genetic algorithm: Nsga- ii.IEEE Transactions on Evolutionary Computation, 6(2):182–197, April 2002. 10 Unconventional Hardware Neural Architecture Search

    2002

  35. [35]

    Gpt-nas: Neural architecture search meets generative pre-trained transformer model.Big Data Mining and Analytics, 8(1):45–64, February 2025

    Caiyang Yu, Xianggen Liu, Yifan Wang, Yun Liu, Wentao Feng, Xiong Deng, Chenwei Tang, and Jiancheng Lv. Gpt-nas: Neural architecture search meets generative pre-trained transformer model.Big Data Mining and Analytics, 8(1):45–64, February 2025

    2025

  36. [36]

    NAS-bench- suite-zero: Accelerating research on zero cost prox- ies

    Arjun Krishnakumar, Colin White, Arber Zela, Renbo Tu, Mahmoud Safari, and Frank Hutter. NAS-bench- suite-zero: Accelerating research on zero cost prox- ies. InThirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022

    2022

  37. [37]

    Lecun, L

    Y . Lecun, L. Bottou, Y . Bengio, and P. Haffner. Gradient-based learning applied to document recog- nition.Proceedings of the IEEE, 86(11):2278–2324, 1998

    1998

  38. [38]

    Reliability anal- ysis of optical neural networks with non-ideal signal transmission.Optical Fiber Technology, 87:103928, October 2024

    Ye Su, Pengju Fu, Yichen Ye, Junxiong Chai, Xiao Jiang, Hongyu Yang, and Yiyuan Xie. Reliability anal- ysis of optical neural networks with non-ideal signal transmission.Optical Fiber Technology, 87:103928, October 2024

    2024

  39. [39]

    Hyperband: A novel bandit-based approach to hyperparameter opti- mization, 2018

    Lisha Li, Kevin Jamieson, Giulia DeSalvo, Afshin Rostamizadeh, and Ameet Talwalkar. Hyperband: A novel bandit-based approach to hyperparameter opti- mization, 2018

    2018

  40. [40]

    NVIDIA Blackwell Architec- ture Technical Brief

    NVIDIA Corporation. NVIDIA Blackwell Architec- ture Technical Brief. NVIDIA, 2024

    2024

  41. [41]

    Efficient processing of deep neural networks: A tutorial and survey, 2017

    Vivienne Sze, Yu-Hsin Chen, Tien-Ju Yang, and Joel Emer. Efficient processing of deep neural networks: A tutorial and survey, 2017

    2017

  42. [42]

    Intel xeon platinum 8380 processor (60m cache, 2.30 ghz) — product specifications

    Intel Corporation. Intel xeon platinum 8380 processor (60m cache, 2.30 ghz) — product specifications. In- tel ARK (Product Information), SKU 212287, 2021. Accessed: 2026-04-24

    2021

  43. [43]

    Nahmias, Thomas Ferreira de Lima, Alexander N

    Mitchell A. Nahmias, Thomas Ferreira de Lima, Alexander N. Tait, Hsuan-Tung Peng, Bhavin J. Shas- tri, and Paul R. Prucnal. Photonic multiply-accumulate operations for neural networks.IEEE Journal of Se- lected Topics in Quantum Electronics, 26(1):1–18, Jan- uary 2020

    2020

  44. [44]

    Asymptotically fault-tolerant programmable photonics.Nature Communications, 13(1), November 2022

    Ryan Hamerly, Saumil Bandyopadhyay, and Dirk En- glund. Asymptotically fault-tolerant programmable photonics.Nature Communications, 13(1), November 2022

    2022

  45. [45]

    A low power fully analog mac operation using standard cmos technology for neural network inference

    Mohamed Bouchakour, Emmanuel Bergeret, Gilles Sicard, Kamel Abdelouahab, and Franc ¸ois Berry. A low power fully analog mac operation using standard cmos technology for neural network inference. In 2025 23rd IEEE Interregional NEWCAS Conference (NEWCAS), page 10–14. IEEE, 2025

    2025

  46. [46]

    ADC Performance Survey 1997–

    Boris Murmann. ADC Performance Survey 1997–

    1997

  47. [47]

    Accessed: 2026-04-24

    https://github.com/bmurmann/ ADC-survey, 2026. Accessed: 2026-04-24

    2026

  48. [48]

    Ian A. D. Williamson, Tyler W. Hughes, Momchil Minkov, Ben Bartlett, Sunil Pai, and Shanhui Fan. Re- programmable electro-optic nonlinear activation func- tions for optical neural networks.IEEE Journal of Selected Topics in Quantum Electronics, 26(1):1–12, January 2020

    2020

  49. [49]

    Chen, and David Z

    Jiaqi Gu, Hanqing Zhu, Chenghao Feng, Zixuan Jiang, Ray T. Chen, and David Z. Pan. L2ight: Enabling on-chip learning for optical neural networks via ef- ficient in-situ subspace optimization. InConference on Neural Information Processing Systems (NeurIPS), 2021

    2021

  50. [50]

    Harris, Ayon Basumallik, Vijay Janapa Reddi, Ajay Joshi, and Darius Bunandar

    Cansu Demirkiran, Furkan Eris, Gongyu Wang, Jonathan Elmhurst, Nick Moore, Nicholas C. Harris, Ayon Basumallik, Vijay Janapa Reddi, Ajay Joshi, and Darius Bunandar. An electro-photonic system for accelerating deep neural networks.ACM Jour- nal on Emerging Technologies in Computing Systems, 19(4):1–31, 2023

    2023

  51. [51]

    Transferable learning on analog hardware

    Sri Krishna Vadlamani, Dirk Englund, and Ryan Hamerly. Transferable learning on analog hardware. Science Advances, 9(28), 2023

    2023

  52. [52]

    Hughes, Tae- won Park, Ben Bartlett, Ian A

    Sunil Pai, Zhanghao Sun, Tyler W. Hughes, Tae- won Park, Ben Bartlett, Ian A. D. Williamson, Mom- chil Minkov, Maziyar Milanizadeh, Nathnael Abebe, Francesco Morichetti, Andrea Melloni, Shanhui Fan, Olav Solgaard, and David A. B. Miller. Experi- mentally realized in situ backpropagation for deep learning in photonic neural networks.Science, 380(6643):398–...

    2023

  53. [53]

    Canceling thermal cross-talk effects in photonic integrated circuits.Jour- nal of Lightwave Technology, 37(4):1325–1332, Febru- ary 2019

    Maziyar Milanizadeh, Douglas Aguiar, Andrea Mel- loni, and Francesco Morichetti. Canceling thermal cross-talk effects in photonic integrated circuits.Jour- nal of Lightwave Technology, 37(4):1325–1332, Febru- ary 2019

    2019

  54. [54]

    Prioritize minimizing op- tical MAC counts and electronic activation energy

    Jiaqi Gu, Zheng Zhao, Chenghao Feng, Hanqing Zhu, Ray T. Chen, and David Z. Pan. Roq: A noise-aware quantization scheme towards robust optical neural net- works with low-bit controls. In2020 Design, Automa- tion; Test in Europe Conference; Exhibition (DATE), page 1586–1589. IEEE, March 2020. 11 Unconventional Hardware Neural Architecture Search A. Hardwar...

    2020

  55. [55]

    Design linear blocks to be as shallow as possible and avoid unnecessary ReLU activations within the optical domain to reduce electronic energy and phase noise sensitivity

  56. [56]

    Prioritize optical linear and convolutional operations over electronic ones, as they consume significantly less energy (∼20fJ/MAC), and ensure they are arranged after the transition point (Flatten or AdaptPool)

  57. [57]

    Minimize the number of MZI components by avoiding overly wide layers, since MZI count scales quadratically; choose layer widths that balance expressivity with hardware complexity

  58. [58]

    Reduce crosstalk impact by limiting miniblock output coupling—prefer architectures with fewer adjacent outputs within each miniblock or include dropout to improve crosstalk tolerance

  59. [59]

    Leverage convolutional layers before the transition point to extract spatial features efficiently, as they do not incur ADC/DAC penalties, and keep the depth of the network shallow to mitigate phase noise accumulation

  60. [60]

    Incorporate BN layers within the optical domain to enhance phase-noise robustness without incurring electronic activation costs, but avoid excessive BN or Dropout that could increase complexity and crosstalk sensitivity

  61. [61]

    Reduce the total number of parameters and layer widths while maintaining sufficient representational capacity to accommodate the non-idealities, thus controlling crosstalk and phase noise effects

  62. [62]

    Use 1×1 convolutions selectively for channel mixing to keep the model compact and reduce the number of optical elements, especially in deeper layers

  63. [63]

    Figure 8.Example synthesized system prompt from the meta-learning phase (GPT-4.1-nano)

    Design the network with a single transition point followed by MZI-based linear blocks to limit electronic boundary crossings, thereby conserving energy and reducing crosstalk pathways. Figure 8.Example synthesized system prompt from the meta-learning phase (GPT-4.1-nano). The underlined sentence contains a factual error: namely that optical MACs at∼20fJ a...

  64. [64]

    Prioritize batch normalization (BN) immediately after optical MZI layers to mitigate phase noise and offset drift, as shown in photonic NN literature

  65. [65]

    Introduce dropout after MZI layers to improve tolerance to crosstalk and random phase errors—dropout acts as a regularizer against structured analog noise

  66. [66]

    Minimize the number of electronic ReLU activations, since their per-element energy cost far exceeds that of optical MACs; use them only where nonlinearity is crucial for expressivity

  67. [67]

    Exploit the efficiency of Conv2d blocks for early spatial feature extraction, as they are native to the hardware and do not incur extra ADC/DAC energy penalty

  68. [68]

    Be aware that crosstalk in MZI meshes couples adjacent outputs within each miniblock—avoid excessive width increases that exacerbate cumulative crosstalk

  69. [69]

    Constrain layer widths and output channels to multiples of 8 (except the classifier head) to align with hardware-friendly quantization and mesh partitioning

  70. [70]

    Use only a single transition point (Flatten or AdaptPool) between Conv and MZI blocks, respecting the mesh’s lack of native support for skip/residual connections

  71. [71]

    Favor architectures that separate spatial processing (Conv2d) and dense classification (MZI Linear), leveraging each domain’s strengths and minimizing unnecessary domain crossings

  72. [72]

    Remember that parameter count scales quadratically with MZI layer width—balance model capacity with physical mesh resource limits

  73. [73]

    Figure 9.Example synthesized system prompt from the meta-learning phase (GPT-4.1)

    Consider the compound effect of quantization noise (8b weights/inputs) and analog noise sources when stacking multiple linear/MZI layers; normalization and dropout help preserve signal integrity. Figure 9.Example synthesized system prompt from the meta-learning phase (GPT-4.1). Unlike the GPT-4.1-nano prompt (Figure 8), this prompt correctly identifies el...

  74. [74]

    Prioritize BN (Batch Normalization) in every MZI layer to normalize signal variance, which is critical for mitigating the impact of phase noise (std=0.05) on weight precision

  75. [75]

    Minimize the use of ReLU activations; since they cost ∼10 pJ/element vs ∼20 fJ/MAC for optical operations, use them only where non-linearity is strictly required for convergence

  76. [76]

    Exploit the miniblock (k=4) structure by grouping highly correlated features together, as crosstalk is localized within these miniblocks

  77. [77]

    Use MZILinearDrop to combat the 0.15 crosstalk factor; dropout acts as a structural regularizer that forces the network to learn redundant representations less sensitive to local hardware fluctuations

  78. [78]

    Favor ConvK1BNReLU for channel mixing over massive linear layers to keep the MZI count ( O(N 2)) manageable and reduce the accumulation of phase noise

  79. [79]

    Balance the depth of the Conv front-end against the MZI back-end; deeper architectures increase the total MZI count, thereby multiplying the cumulative phase noise impact

  80. [80]

    When non-ideal accuracy drops significantly, prefer widening the layer slightly over increasing depth, as depth leads to faster signal degradation due to the lack of skip connections

Showing first 80 references.