Pith. sign in

REVIEW 3 major objections 52 references

LEMUR 2 unifies heterogeneous architecture generators, multi-task evaluation, and real-device deployment into a corpus of over 14,000 neural nets and 750,000 training records for data-driven design.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-10 20:14 UTC pith:WK7CND7Z

load-bearing objection Solid multi-pipeline corpus and edge/VR metadata; the “LLM-AutoML foundation” claim outruns the short-budget evidence they themselves flag. the 3 major comments →

arxiv 2607.06839 v1 pith:WK7CND7Z submitted 2026-07-07 cs.LG cs.CV

LEMUR 2: Unlocking Neural Network Diversity for AI

Tolgay Atinc Uzun , Waleed Khalid , Saif U Din , Sai Revanth Mulukuledu , Akashdeep Singh , Chandini Vysyaraju , Raghuvir Duvvuri , Avi Goyal
show 12 more authors
Yashkumar Rajeshbhai Lukhi Muhammad A. Hussain Krunal Jesani Usha Shrestha Yash Mittal Roman Kochnev Pritam Kadam Mohsin Ikram Harsh R. Moradiya Alice Arslanian Dmitry Ignatov Radu Timofte
This is my paper
classification cs.LG cs.CV
keywords neural architecture searchAutoMLneural network diversityLLM-driven architecture generationedge deploymentmultimodal benchmarksretrieval-augmented generationarchitecture transferability
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Existing neural architecture benchmarks mostly enumerate a single narrow cell space on one or two vision tasks and rarely record how models behave on real devices. This paper presents LEMUR 2, an extensible framework that generates architectures by several independent routes—source-code mutation, genetic and reinforcement-learning search, fractal recursion, few-shot language-model prompting, and retrieval of reusable modules from public code—then trains them under a shared protocol across image classification, captioning, text-to-image, language modeling, and mixture-of-experts setups. The same models are converted and timed automatically on Android and Unity VR runtimes, so accuracy, hyperparameters, and measured latency sit in one structured store. The resulting corpus of more than 14,000 distinct architectures and 750,000 training records is offered as the empirical foundation for language-model-guided AutoML and for studying which design motifs transfer across modalities and hardware.

Core claim

A single open framework can couple many different architecture origins—AST channel mutation, genetic evolution, reinforcement-learning skeleton completion, fractal multi-column construction, few-shot LLM prompting, and retrieval-augmented extraction of over 900 validated PyTorch modules—with multi-domain training logs and automated mobile and VR latency measurements, producing a reusable corpus large and diverse enough to support LLM fine-tuning and cross-domain architectural analysis beyond the scope of fixed-cell NAS benchmarks.

What carries the argument

The unified generative–evaluative–deployment pipeline (including NN-RAG module mining and the NN-Lite / NN-VR conversion-and-benchmark loops) that forces every candidate, regardless of origin, into a common training schema and real-device measurement record.

Load-bearing premise

That rankings and diversity signals obtained from short, fixed training budgets remain informative for fully trained models, for transfer across tasks, and for fine-tuning architecture-generating language models.

What would settle it

Train the top- and bottom-ranked architectures from each generator family to full convergence under identical long schedules; if the short-budget orderings reverse, or if Android-emulator latency rankings reverse on physical devices, the claimed foundation for data-driven design fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • The corpus supplies supervised examples for fine-tuning language models that propose new architectures.
  • Accuracy–latency trade-offs can be studied directly from measured mobile and VR numbers rather than from surrogate costs alone.
  • Architectural transferability can be quantified across classification, captioning, text-to-image, and language-modeling tasks under matched compute.
  • Different generation methods (mutation, evolution, fractal, retrieval, prompting) become comparable under one logging schema.
  • Reusable, dependency-closed neural primitives extracted from public code become an executable substrate for automated pipelines.

Where Pith is reading between the lines

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

  • If short-budget rankings prove stable under full training, the same store could serve as a standard pretraining substrate for architecture-design models the way large text corpora serve language models.
  • The multi-generator plus multi-hardware pattern can be extended to audio, graph, or reinforcement-learning search spaces without redesigning the storage layer.
  • Observed context-overflow failures under longer few-shot prompts imply that retrieval-augmented block libraries may scale more reliably than pure prompting for complex design spaces.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 0 minor

Summary. LEMUR 2 presents a large-scale, extensible neural-architecture corpus and pipeline that unifies heterogeneous generators (AST channel mutation, genetic algorithms, RL/GRPO masked completion, fractal recursive construction, few-shot LLM prompting, and NN-RAG module extraction from public PyTorch codebases) with multi-task evaluation (image classification, captioning, text-to-image, language modeling, MoE) and real-device deployment metadata via NN-Lite (Android/TFLite) and NN-VR (Unity/Barracuda). The authors report >14,000 distinct architectures and >750,000 structured training records, plus concrete pipeline statistics (e.g., ~7.5k Android conversions, 941 validated RAG blocks, success rates and best-per-run metrics in Table 2). The central claim is that this multi-origin, multi-task, deployment-aware resource supplies a data foundation for LLM-driven AutoML and architectural generalization beyond narrow cell-based NAS benchmarks.

Significance. If the corpus is released as described and the short-budget rankings prove sufficiently stable for the intended downstream uses, LEMUR 2 would be a useful systems contribution: it expands NAS-style resources beyond fixed cell spaces, adds real-device latency/memory metadata that most benchmarks lack, and couples generation methods (including retrieval-augmented mining of public modules) with multimodal tasks. Strengths include concrete counts and success rates (Table 2), open-code generation rather than closed search spaces, and explicit deployment pipelines. The work is primarily a dataset/systems release rather than a theoretical result; its lasting value depends on public accessibility of the data interface, model sources, and logs, and on whether the short-horizon signals actually improve LLM fine-tuning or transfer—points the manuscript asserts via concurrent arXiv work but does not fully demonstrate here.

major comments (3)
  1. §3 and Table 2 (including footnotes a/b): The paper itself states that evaluations use bounded schedules so scores are compute-budgeted rather than uniformly converged, and that comparisons require matched horizons. Yet best-per-run accuracy under those short schedules (e.g., GA 0.8004, fractal 0.8018 at 5 epochs, few-shot 0.3874, AST 0.3110) and group medians (Fig. 9) are presented as primary evidence of generator quality and of the corpus as a foundation for LLM fine-tuning and cross-domain generalization. Without stability checks (rank correlation under longer training, or an ablation showing that short-budget features improve LLM-driven synthesis), the leap from “large multi-pipeline corpus” to the abstract’s AutoML/generalization claim is under-supported.
  2. Abstract / §1.2 / Conclusion vs. Table 1 and §4: Cross-domain transferability and architectural generalization are repeatedly claimed, but the reported results remain largely task-siloed (CIFAR classification for most generators; separate captioning BLEU-4, CLIP scores, MoE on CIFAR-10). There is no quantitative transfer experiment (e.g., architecture motifs or rankings that improve a held-out modality) inside this manuscript. The multi-task coverage is real; the transferability claim needs either evidence or a clearer scope limitation.
  3. §4 / Table 2 and the NNGPT citations: Empirical “proof of value” for LLM-driven AutoML is largely deferred to concurrent arXiv papers by the same group. For a self-contained journal contribution, either include a minimal, reproducible demonstration that LEMUR 2 records improve an LLM architecture generator (or a simple meta-learner) relative to a non-LEMUR baseline, or rephrase the central claim to what is actually shown: a multi-origin corpus with deployment metadata.

Circularity Check

1 steps flagged

Dataset/systems release is not circular by construction; moderate self-citation load-bearing only for the AutoML-utility claim.

specific steps
  1. self citation load bearing [§1.2 Our Contribution (final paragraph); also Abstract closing claim]
    "The empirical value of the LEMUR dataset, as demonstrated within the NNGPT project, is proved by its successful application to LLM-guided generative synthesis, exploration of complex architectural structures, and improved generalization across multiple computer vision tasks [13, 17, 21, 25, 33, 45, 47, 48]."

    The paper’s central utility claim—that LEMUR 2 supplies a data foundation for LLM fine-tuning and LLM-driven AutoML—is ‘proved’ almost entirely by a chain of concurrent/prior arXiv works with overlapping authors (NNGPT, NN-RAG, fractal/LLM architecture papers). Those citations are not machine-checked uniqueness results or external independent replications; they are the same research program. The corpus construction itself is independent, but the load-bearing argument that the corpus advances LLM-AutoML reduces to self-citation rather than external evidence.

full rationale

LEMUR 2 is primarily a corpus and pipeline paper: it enumerates generative methods (AST mutation, GA, RL, fractal, few-shot LLM, NN-RAG), multi-task training records, and deployment pipelines (NN-Lite, NN-VR), then reports measured counts and short-budget metrics. Those empirical constructions do not reduce by definition to their inputs, nor do they rename a known closed-form result as a prediction. The only circularity-adjacent pattern is that the load-bearing claim that the corpus is a proven foundation for LLM-driven AutoML and architectural generalization is justified mainly by concurrent/prior papers from the same group (NNGPT, NN-RAG, fractal/LLM-NAS, etc.) rather than independent external benchmarks. That is self-citation load-bearing for the utility narrative, not for the existence or contents of the dataset itself. Central content (architecture generation, multi-task logs, on-device latency metadata) remains independently described and falsifiable. Score 3 reflects proportionate weight: some self-citation on the value claim, with the core systems contribution still independent. No self-definitional equations, fitted-as-prediction, uniqueness theorems, or ansatz-smuggling via citation were found.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 5 invented entities

As a systems/dataset paper, load-bearing content is operational assumptions and design choices rather than physical axioms. The central claim rests on (i) that heterogeneous code-level generators plus short-budget training logs constitute useful diversity for AutoML/LLMs, (ii) reward and search hyperparameters of the generators, and (iii) invented pipeline names that package known conversion/search ideas. No formal theorem is claimed; free parameters are search/reward/training knobs that shape which architectures enter the corpus and how they are ranked.

free parameters (7)
  • RL composite reward weights = −1.0 / +0.2 / +1.0×Δacc
    §2.2.2 sets −1.0 fail, +0.2 per stage, +1.0×Δaccuracy; these hand-chosen weights steer GRPO and thus which completions enter the corpus.
  • Few-shot exemplar count n = n∈{1,3,6,...}; best cited n=3
    §2.2.6–4.2 treats number of exemplars as tunable; reported best balanced mean at n=3, collapse at n=6—directly affects which LLM architectures are kept.
  • Bounded training schedules (epochs/mini-epochs) = task-specific short budgets (often 1–5 epochs)
    §3 and Table 2: short fixed budgets (e.g., fractal 5 epochs, single-epoch transform search, RL one mini-epoch) define all comparative accuracies used to market generator quality.
  • Genetic algorithm fitness horizon = small number of epochs (unspecified exact count in main text)
    §2.2.3 fitness from few-epoch validation accuracy on CIFAR-10; horizon choice biases which AlexNet-style variants survive.
  • Fractal depth N and column width = sampled N and width (exact grid not fully tabulated in main text)
    §2.2.4 topology parameters that define the fractal search space and thus the 1,200+ architectures.
  • Data-transform fixed training recipe = 1 epoch, bs=64, lr=0.01, mom=0.9, dropout=0.2
    §2.2.7 ranks 6,000 pipelines by one-epoch ResNet accuracy with fixed lr/momentum/dropout; recipe choice defines ‘best’ augmentation.
  • VR frame budget threshold = 11 ms @ 90 Hz
    §2.3.1 evaluates latency against 11 ms/frame (90 Hz); threshold defines success metadata for NN-VR.
axioms (5)
  • domain assumption Heterogeneous code-level generation (mutation, evolution, LLM, RAG) plus shared logging yields architectural diversity that is useful for discovering general design principles and for LLM fine-tuning.
    Stated throughout abstract, §1, and conclusion; not independently proven—assumed as the motivation for corpus scale.
  • domain assumption Short compute-matched training runs are sufficient to rank architectures and generators for corpus inclusion and comparative claims.
    §3 explicitly adopts bounded schedules; all Table 2 generator comparisons rest on this.
  • domain assumption AST/channel mutations and masked LLM completions that pass compile+dummy forward are valid members of the design space for AutoML study.
    §2.2.1–2.2.2 validation criteria; shape-valid code is treated as meaningful architecture diversity.
  • domain assumption On-device TFLite/Barracuda latency under AVD/Unity settings is a relevant proxy for deployment-aware architecture quality.
    §2.3 NN-Lite/NN-VR; emulator and engine-specific numbers are presented as real-device performance metadata.
  • standard math Standard deep-learning training and NAS evaluation practice (CIFAR/COCO metrics, perplexity/BLEU/CLIP scores) is an adequate yardstick for multi-domain transfer claims.
    Background evaluation conventions used in §2.1 and §4 without re-derivation.
invented entities (5)
  • LEMUR 2 corpus/framework independent evidence
    purpose: Unified store of >14k architectures, 750k training records, generators, and deployment metadata as AutoML substrate.
    Named extension of LEMUR; the paper’s primary product. Independent evidence is the claimed public repo and reported counts, not external physics-style prediction.
  • NN-RAG independent evidence
    purpose: Mine and validate self-contained torch.nn.Module blocks from public repos as reusable generation primitives.
    Introduced as retrieval-augmented architectural motif library (~900 modules). Falsifiable via the extraction pipeline and pass rate (73%), but mainly an engineering construct.
  • NN-Lite independent evidence
    purpose: Automated PyTorch→TFLite Android deployment and latency logging at scale.
    Pipeline entity packaging conversion + AVD benchmarking; evidence is reported 7,512 models processed.
  • NN-VR independent evidence
    purpose: ONNX→Unity/Barracuda validation against VR frame-time and memory criteria.
    Pipeline entity for immersive deployment metadata; evidence is reported large-scale conversion success rate.
  • NN-Caption no independent evidence
    purpose: LLM-driven generation of encoder–decoder captioners under LEMURNet interface.
    Task-specific generator contributing 357 architectures; engineering wrapper rather than a new scientific object.

pith-pipeline@v1.1.0-grok45 · 18015 in / 4696 out tokens · 58435 ms · 2026-07-10T20:14:34.581266+00:00 · methodology

0 comments
read the original abstract

Existing NAS benchmarks (e.g., NAS-Bench, NATS-Bench) cover only narrow, task-specific regions of the architectural design space and lack cross-domain or deployment-aware evaluation. LEMUR 2 introduces a large-scale, extensible framework unifying generative, evaluative, and deployment pipelines to unlock neural-network diversity. It comprises over 14,000 distinct architectures and more than 750,000 structured training records documenting model performance, hyperparameters, and task outcomes. These models were produced through AST-based code mutation, genetic and reinforcement-learning evolution, generation of fractal architectures, and synthesis guided by a Large Language Model (LLM). This includes deep models generated with the retrieval-augmented system NN-RAG, which derived and used architectural motifs from over 900 PyTorch modules extracted from public repositories. LEMUR 2 further employs NN-VR and NN-Lite pipelines for automated deployment and latency benchmarking on heterogeneous mobile and Unity-based VR platforms, providing real-device performance metadata. It spans multimodal tasks, image captioning, text-to-image synthesis, and language modeling, supporting cross-domain analysis of architectural transferability. By linking diverse architectures, tasks, and deployment data, LEMUR 2 provides the data foundation for LLM fine-tuning and coupling diverse architectural origins with large-scale, cross-platform empirical validation. This dataset defines a new basis for reproducible and data-driven AI design, advancing the emerging paradigm of LLM-driven AutoML and architectural generalization across modalities and hardware.

Figures

Figures reproduced from arXiv: 2607.06839 by Akashdeep Singh, Alice Arslanian, Avi Goyal, Chandini Vysyaraju, Dmitry Ignatov, Harsh R. Moradiya, Krunal Jesani, Mohsin Ikram, Muhammad A. Hussain, Pritam Kadam, Radu Timofte, Raghuvir Duvvuri, Roman Kochnev, Saif U Din, Sai Revanth Mulukuledu, Tolgay Atinc Uzun, Usha Shrestha, Waleed Khalid, Yashkumar Rajeshbhai Lukhi, Yash Mittal.

Figure 1
Figure 1. Figure 1: Taxonomy of NAS benchmarks organized by their pri [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Reinforcement Learning pipeline overview. Using the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: NN-RAG pipeline overview. a new architecture consistent with the LEMUR interface; the number of exemplars is treated as a tunable parameter. Each generated architecture is normalized into a canonical form and assigned a whitespace-stripped MD5 hash, which serves as a database key for fast detection and removal of exact duplicates prior to training. Non-duplicate architec￾tures are trained under the standar… view at source ↗
Figure 5
Figure 5. Figure 5: Overview of the brute-force generation of augmentation [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Best per-run image captioning BLEU-4 distributions. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Top 10 image classification models across the dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Median accuracy per model group for image classifica [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗

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

52 extracted references · 52 canonical work pages · 15 internal anchors

  1. [1]

    AugmentGest: Can Random Data Cropping Augmentation Boost Gesture Recognition Performance?

    Nada Aboudeshish, Dmitry Ignatov, and Radu Timofte. Augmentgest: Can random data cropping augmentation boost gesture recognition performance?arXiv preprint arXiv:2506.07216, 2025. 5

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  2. [2]

    Jahs-bench-201: A foundation for research on joint architecture and hyperparameter search

    Archit Bansal, Danny Stoll, Maciej Janowski, Arber Zela, and Frank Hutter. Jahs-bench-201: A foundation for research on joint architecture and hyperparameter search. InPro- ceedings of the 36th Conference on Neural Information Pro- cessing Systems (NeurIPS 2022) — Datasets & Benchmarks Track, 2022. Available at GitHub:https://github. com/automl/jahs_bench...

    work page 2022

  3. [3]

    Approximating CNNs with Bag-of-local-Features models works surprisingly well on ImageNet

    Wieland Brendel and Matthias Bethge. Approximating cnns with bag-of-local-features models works surprisingly well on imagenet.arXiv preprint arXiv:1904.00760, 2019. 4

    work page internal anchor Pith review Pith/arXiv arXiv 1904

  4. [4]

    AIRNet: Self-Supervised Affine Registration for 3D Medical Images using Neural Networks

    Evelyn Chee and Zhenzhou Wu. Airnet: Self-supervised affine registration for 3d medical images using neural net- works.arXiv preprint arXiv:1810.02583, 2018. 4

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  5. [5]

    Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling

    Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling.arXiv preprint arXiv:1412.3555, 2014. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  6. [6]

    Ai on the edge: An automated pipeline for pytorch-to-android deployment and benchmarking.Preprints, 2025

    Saif U Din, Muhammad Ahsan Hussain, Mohsin Ikram, Dmitry Ignatov, and Radu Timofte. Ai on the edge: An automated pipeline for pytorch-to-android deployment and benchmarking.Preprints, 2025. 2, 6

    work page 2025

  7. [7]

    NAS-Bench-201: Extending the scope of reproducible neural architecture search

    Xuanyi Dong and Yi Yang. NAS-Bench-201: Extending the scope of reproducible neural architecture search. InInter- national Conference on Learning Representations (ICLR),

    work page

  8. [8]

    NATS-Bench: Benchmarking nas algorithms for ar- chitecture topology and size.IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2021

    Xuanyi Dong, Lu Liu, Katarzyna Musial, and Bogdan Gabrys. NATS-Bench: Benchmarking nas algorithms for ar- chitecture topology and size.IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2021. doi:10. 1109/TPAMI.2021.3054824. 2, 3

    work page arXiv 2021

  9. [9]

    Transnas-bench-101: Improving transferability and generalizability of cross-task neural architecture search

    Yawen Duan, Xin Chen, Hang Xu, Zewei Chen, Xiaodan Liang, Tong Zhang, and Zhenguo Li. Transnas-bench-101: Improving transferability and generalizability of cross-task neural architecture search. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5251–5260, 2021. 2, 3

    work page 2021

  10. [10]

    Jeffrey L. Elman. Finding structure in time.Cognitive Sci- ence, 14(2):179–211, 1990. 3

    work page 1990

  11. [11]

    NNGPT: Rethinking AutoML with Large Language Models

    Roman Kochnev et al. NNGPT: Rethinking AutoML with large language models.arXiv preprint arXiv:2511.20333,

    work page internal anchor Pith review Pith/arXiv arXiv

  12. [12]

    A Retrieval-Augmented Generation Approach to Extracting Algorithmic Logic from Neural Networks

    Waleed Khalid et al. A retrieval-augmented generation ap- proach to extracting algorithmic logic from neural networks. arXiv preprint arXiv:2512.04329, 2025. 5

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  13. [13]

    From Memorization to Creativity: LLM as a Designer of Novel Neural Architectures

    Waleed Khalid et al. From memorization to creativity: Llm as a designer of novel neural-architectures.arXiv preprint, arXiv:2601.02997, 2026. 2, 5

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  14. [14]

    VIST-GPT: Ushering in the Era of Visual Storytelling with LLMs?

    Mohamed Gado, Towhid Taliee, Muhammad Danish Memon, Dmitry Ignatov, and Radu Timofte. Vist-gpt: Ush- ering in the era of visual storytelling with llms?arXiv preprint arXiv:2504.19267, 2025. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  15. [15]

    LEMUR Neural Network Dataset: Towards Seamless AutoML, 2025

    Arash Torabi Goodarzi, Roman Kochnev, Waleed Khalid, Furui Qin, Tolgay Atinc Uzun, Yashkumar Sanjaybhai Dhameliya, Yash Kanubhai Kathiriya, Zofia Antonina Ben- tyn, Dmitry Ignatov, and Radu Timofte. LEMUR Neural Network Dataset: Towards Seamless AutoML, 2025. 1, 2, 3

    work page 2025

  16. [16]

    Tensorflow hub.https://www.tensorflow

    Google. Tensorflow hub.https://www.tensorflow. org/hub, 2018. 2

    work page 2018

  17. [17]

    Resource- efficient iterative LLM-based NAS with feedback memory

    Xiaojie Gu, Dmitry Ignatov, and Radu Timofte. Resource- efficient iterative llm-based nas with feedback memory. arXiv preprint, arXiv:2603.12091, 2026. 2

    work page arXiv 2026

  18. [18]

    Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020. 3

    work page 2020

  19. [19]

    Long short-term memory.Neural Computation, 9(8):1735–1780, 1997

    Sepp Hochreiter and J ¨urgen Schmidhuber. Long short-term memory.Neural Computation, 9(8):1735–1780, 1997. 3

    work page 1997

  20. [20]

    Densely connected convolutional net- works

    Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kil- ian Q Weinberger. Densely connected convolutional net- works. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 4700–4708, 2017. 4

    work page 2017

  21. [21]

    Llm as a neural architect: Controlled generation of image cap- tioning models under strict api contracts.arXiv preprint, arXiv:2512.14706, 2025

    Krunal Jesani, Dmitry Ignatov, and Radu Timofte. Llm as a neural architect: Controlled generation of image cap- tioning models under strict api contracts.arXiv preprint, arXiv:2512.14706, 2025. 2

    work page arXiv 2025

  22. [22]

    Auto-keras: An efficient neural architecture search system

    Haifeng Jin, Qingquan Song, and Xia Hu. Auto-keras: An efficient neural architecture search system. InProceedings of the 25th ACM SIGKDD international conference on knowl- edge discovery & data mining, pages 1946–1956, 2019. 2

    work page 1946

  23. [23]

    Auto-Encoding Variational Bayes

    Diederik P Kingma and Max Welling. Auto-encoding varia- tional bayes.arXiv preprint arXiv:1312.6114, 2013. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  24. [24]

    Nas-bench-nlp: neural architecture search benchmark for natural language processing.IEEE Access, 10:45736–45747, 2022

    Nikita Klyuchnikov, Ilya Trofimov, Ekaterina Artemova, Mikhail Salnikov, Maxim Fedorov, Alexander Filippov, and Evgeny Burnaev. Nas-bench-nlp: neural architecture search benchmark for natural language processing.IEEE Access, 10:45736–45747, 2022. 2, 3

    work page 2022

  25. [25]

    Roman Kochnev, Arash Torabi Goodarzi, Zofia Antonina Bentyn, Dmitry Ignatov, and Radu Timofte. Optuna vs Code Llama: Are LLMs a New Paradigm for Hyperparame- ter Tuning? InProceedings of the IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 5664–5674, 2025. 2

    work page 2025

  26. [26]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 4

    work page 2009

  27. [27]

    Imagenet classification with deep convolutional neural net- works

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural net- works. InAdvances in Neural Information Processing Sys- tems. Curran Associates, Inc., 2012. 4

    work page 2012

  28. [28]

    FractalNet: Ultra-Deep Neural Networks without Residuals

    Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Fractalnet: Ultra-deep neural networks without residuals.arXiv preprint arXiv:1605.07648, 2016. 4

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  29. [29]

    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 Lin. Hw-nas-bench: Hardware-aware neural archi- tecture search benchmark.arXiv preprint arXiv:2103.10584,

    work page internal anchor Pith review Pith/arXiv arXiv

  30. [30]

    LibCST: A concrete syntax tree parser and serializer library for python.https://github

    LibCST developers. LibCST: A concrete syntax tree parser and serializer library for python.https://github. com/Instagram/LibCST, 2025. Version 1.8.6, ac- cessed November 14, 2025. 5

    work page 2025

  31. [31]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740–755. Springer, 2014. 2, 3

    work page 2014

  32. [32]

    Nas-bench-asr: Reproducible neural architecture search for speech recogni- tion

    Abhinav Mehrotra, Lukasz Dudziak, et al. Nas-bench-asr: Reproducible neural architecture search for speech recogni- tion. InICLR, 2021. 3

    work page 2021

  33. [33]

    Preparation of Fractal-Inspired Computational Architectures for Advanced Large Language Model Analysis

    Yash Mittal, Dmitry Ignatov, and Radu Timofte. Prepa- ration of Fractal-Inspired Computational Architectures for Advanced Large Language Model Analysis.arXiv preprint arXiv:2511.07329, 2025. 2, 4

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  34. [34]

    Olson and Jason H

    Randal S. Olson and Jason H. Moore. Tpot: A tree-based pipeline optimization tool for automating machine learn- ing. InProceedings of the Workshop on Automatic Machine Learning, pages 66–74, New York, New York, USA, 2016. PMLR. 2

    work page 2016

  35. [35]

    Bleu: a method for automatic evaluation of machine translation

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. InProceedings of the 40th Annual Meeting on Association for Computational Linguistics, page 311–318, USA, 2002. Association for Computational Linguistics. 3

    work page 2002

  36. [36]

    Pytorch: An im- perative style, high-performance deep learning library, 2019

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas K ¨opf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An im- perative style, high-perf...

    work page 2019

  37. [37]

    Pytorch hub.https://pytorch

    PyTorch Foundation. Pytorch hub.https://pytorch. org/hub, 2019. 2

    work page 2019

  38. [38]

    Nas-bench-graph: Benchmarking graph neural architecture search.Advances in neural information process- ing systems, 35:54–69, 2022

    Yijian Qin, Ziwei Zhang, Xin Wang, Zeyang Zhang, and Wenwu Zhu. Nas-bench-graph: Benchmarking graph neural architecture search.Advances in neural information process- ing systems, 35:54–69, 2022. 3

    work page 2022

  39. [39]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021. 3

    work page 2021

  40. [40]

    James Reed, Zachary DeVito, Horace He, Ansley Ussery, and Jason Ansel. torch. fx: Practical program capture and transformation for deep learning in python.Proceedings of Machine Learning and Systems, 4:638–651, 2022. 4

    work page 2022

  41. [41]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 3

    work page 2022

  42. [42]

    U- net: Convolutional networks for biomedical image segmen- tation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- net: Convolutional networks for biomedical image segmen- tation. InInternational Conference on Medical image com- puting and computer-assisted intervention, pages 234–241. Springer, 2015. 3

    work page 2015

  43. [43]

    Ex- ploring the collaboration between vision models and llms for enhanced image classification.Dimensions, 27(1), 2025

    Bhavya Rupani, Dmitry Ignatov, and Radu Timofte. Ex- ploring the collaboration between vision models and llms for enhanced image classification.Dimensions, 27(1), 2025. doi:10.13140/RG.2.2.14615.69284. 2

    work page doi:10.13140/rg.2.2.14615.69284 2025

  44. [44]

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of math- ematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024. 4

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  45. [45]

    From brute force to semantic insight: Performance-guided data transformation design with llms.arXiv preprint, arXiv:2601.03808, 2026

    Usha Shrestha, Dmitry Ignatov, and Radu Timofte. From brute force to semantic insight: Performance-guided data transformation design with llms.arXiv preprint, arXiv:2601.03808, 2026. 2, 5

    work page arXiv 2026

  46. [46]

    Nas-bench-301 and the case for surrogate benchmarks for neural architecture search

    Julien Siems, Lucas Zimmer, Arber Zela, Jovita Lukasik, Margret Keuper, and Frank Hutter. Nas-bench-301 and the case for surrogate benchmarks for neural architecture search. InNeurIPS, 2020. 3

    work page 2020

  47. [47]

    Closed-Loop LLM Discovery of Non-Standard Channel Priors in Vision Models

    Tolgay Atinc Uzun, Dmitry Ignatov, and Radu Timofte. Closed-loop llm discovery of non-standard channel priors in vision models.arXiv preprint, arXiv:2601.08517, 2026. 2, 4

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  48. [48]

    Enhancing LLM-Based Neural Network Generation: Few-Shot Prompting and Efficient Validation for Automated Architecture Design

    Chandini Vysyaraju, Raghuvir Duvvuri, Avi Goyal, Dmitry Ignatov, and Radu Timofte. Enhancing llm-based neural network generation: Few-shot prompting and efficient val- idation for automated architecture design.arXiv preprint, arXiv:2512.24120, 2025. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  49. [49]

    P.J. Werbos. Backpropagation through time: what it does and how to do it.Proceedings of the IEEE, 78(10):1550–1560,

    work page

  50. [50]

    Nas-bench-101: Towards reproducible neural architecture search

    Chris Ying, Aaron Klein, Eric Christiansen, Esteban Real, Kevin Murphy, and Frank Hutter. Nas-bench-101: Towards reproducible neural architecture search. InInternational conference on machine learning, pages 7105–7114. PMLR,

    work page

  51. [51]

    Nas-bench- 1shot1: Benchmarking and dissecting one-shot neural archi- tecture search

    Arber Zela, Julien Siems, and Frank Hutter. Nas-bench- 1shot1: Benchmarking and dissecting one-shot neural archi- tecture search. InICLR, 2020. 3

    work page 2020

  52. [52]

    Self-attention generative adversarial networks

    Han Zhang, Ian Goodfellow, Dimitris Metaxas, and Augus- tus Odena. Self-attention generative adversarial networks. In International conference on machine learning, pages 7354–

    work page