Scaling Closed-Loop Feature Channel Configuration with LLMs

Dmitry Ignatov; Radu Timofte; Tolgay Atinc Uzun

REVIEW 3 major objections 3 minor 24 references

Scaling Closed-Loop Feature Channel Configuration with LLMs

T0 review · 3 major / 3 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read This paper claims that scaling a closed-loop LLM channel-search loop from sparse to dense sampling (250 candidates per cycle) preserves and strengthens the search signal: best one-epoch CIFAR-100 accuracy rises from 0.3144 to 0.3676, later

desk verdict Honest, careful scaling study of an LLM channel-search loop; the data accounting and confound audit are strong, but the headline frontier gain rests on a single run measured only on an unvalidated one-epoch proxy. read the letter →

arxiv 2607.20516 v1 pith:A3MBYVKY submitted 2026-07-07 cs.LG cs.AI

Tolgay Atinc Uzun , Radu Timofte , Dmitry Ignatov This is my paper

classification cs.LGcs.AI

keywords neuralarchitecturesearchchannelconfigurationlargelanguagemodelsclosed-loopCIFAR-100proxyevaluationparameterefficiencycodegeneration

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 the accuracy signal seen in a small closed-loop LLM-based channel search is real and becomes stronger with more sampling, not a lucky draw. Scaling the same loop to 250 candidates per cycle and strictly filtering 2000 generated programs down to 462 verified CIFAR-100 evaluations, the authors find positive per-cycle accuracy trends and a frontier that improves from 0.3144 to 0.3676. Later cycles also shift the frontier toward far smaller networks, matching or beating the early best with an order-of-magnitude fewer parameters. The denser sample makes architectural patterns measurable — about 42% of architectures use non-power-of-two widths, and top models consistently place more capacity in middle and lower-resolution blocks — which is why a careful reader should care: LLM-generated channel priors become a distributional, testable phenomenon rather than anecdotal discoveries.

What carries the argument

The load-bearing mechanism is the closed-loop code-generation protocol: a language model receives a baseline network and a target improvement, emits an executable network program, the program is trained for exactly one epoch on CIFAR-100 under a fixed protocol, and successfully evaluated candidates are fed back into the model via parameter-efficient fine-tuning for the next cycle. Scaling this loop to 250 candidates per cycle, combined with a strict candidate-accounting rule (only candidates with a valid epoch log and CIFAR-100 metadata count), is what produces the trend and frontier claims. The architectural claims are carried by two additional instruments: static extraction of layer widths

What would settle it

Train the top candidates from early and late cycles (A0/B134, A6/B190, A5/B99) under the full training schedule and check whether the one-epoch ranking matches the final ranking; if final accuracies re-order the candidates enough to erase the 0.3676 vs 0.3144 gap, the central claim is a proxy artifact. A complementary check: run a second independent 8-cycle experiment with the same protocol and budget; if the best accuracy stays at or below A0's 0.3144 (the expected maximum of a 250-sample cycle is roughly 0.31), the frontier trend was sampling noise on a single realization.

Watch

Extended reading notes

Core claim

Central claim: the channel-search behavior found in a sparse earlier study transfers to denser sampling. Across eight cycles of 250 candidates each, 462 of 2000 generated programs pass strict filtering as verified CIFAR-100 runs. Per-cycle mean accuracy trends up (slope 9.87e-4, p=0.043); the best accuracy rises from 0.3144 to 0.3676; and the later frontier is far more parameter-efficient (11.8M vs 166.5M parameters for the early incumbent). Denser sampling exposes regularities: 41.8% of verified networks use non-power-of-two widths, and top models concentrate capacity in middle/low-resolution blocks. A confound audit says the channel signal survives controls for size, depth, MACs, and non-c

Load-bearing premise

The load-bearing premise is that one-epoch CIFAR-100 accuracy ranks generated architectures in the same order that full training would; the paper states in Section 10 that no candidate was trained for more than one epoch and that a full-training run is needed to check this correlation.

Editorial extensions

If this is right

If the claim holds, larger candidate budgets for LLM-based architecture search yield genuine frontier improvements, so the best architecture found should continue to rise with scale rather than saturating at an early lucky draw.
Later cycles dominate the early frontier across parameter budgets: within matched parameter quartiles, later-cycle maxima are consistently higher, so search evaluation should compare accuracy-parameter Pareto trade-offs, not a single best accuracy.
Among generated networks, spending a larger fraction of primary-convolution parameters at lower spatial resolutions is associated with faster early learning, and this association is not explained by total parameter count, depth, or non-channel code choices.
Non-power-of-two channel widths are common in the LLM's channel prior (41.8% of architectures) and appear in the best model, so restricting channel search to conventional powers of two would exclude the discovered optimum.
Reporting the full generation ledger — successes, failures, and non-target evaluations — matters: strict filtering removes the majority of generated programs from the analysis and prevents false-positive accuracy trajectories.

Reading between the lines

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

A natural testable extension the paper leaves implicit is a causal intervention: hold depth, MACs, and parameter budget fixed and move channels from early/high-resolution layers to later/low-resolution layers; the quartile-matching analysis predicts a controlled accuracy gain if the association is causal rather than correlational.
The one-epoch proxy is the hinge: if a full-training run re-ranks the top candidates materially, the frontier and architectural regularities would need to be re-benchmarked; the paper itself flags this uncertainty in its limitations.
Because later cycles depend on fine-tuning from earlier generations, the samples are not independent; a second independent run with the same budget would be needed to separate a true distributional shift from the expected maximum of a noisy order statistic.
The non-power-of-two widths may reflect the model's prior rather than learned search; comparing channel distributions from base and fine-tuned models across cycles could reveal how much of the pattern is actually induced by accuracy feedback.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Colleague, quick take: this is a solid empirical extension of the authors' own closed-loop LLM channel-search line, and it deserves a serious referee, but the central claim—that scaling improves the frontier—is more fragile than the abstract suggests.

What's genuinely new: the denser sampling (462 verified CIFAR-100 evaluations from 8 cycles of 250 candidates), the Pareto-efficiency analysis, the resolution-wise channel allocation findings, and the confound audit. The accounting in Table 1 is meticulous: failed code, no-output, non-target tasks all separated, and the analysis script regenerates all tables from raw folders. That is reproducible in spirit. The confound audit (controls for stride, downsampling, hidden-code flags, etc.) is a real attempt to show the signal lives in channels, not in training-code drift. Credit where due.

Where it gets soft: the headline improvement from 0.3144 to 0.3676 is a single order statistic—one run, one candidate (A6/B190). The paper itself notes a second run would produce a different frontier. The trend significance is computed on eight non-i.i.d. cycle-level points with wide overlapping bootstrap CIs; the top-5/top-10 trends are more stable but still cycle-level. And the deeper issue is the one the stress-test flagged: every claim is measured on one-epoch CIFAR-100 accuracy, and the paper provides no evidence that one-epoch rankings correlate with full-training rankings. Section 10 admits this explicitly. If the proxy doesn't rank-preserve, the specific 'best model' and Pareto-frontier claims may not transfer to real trained performance. That said, the distributional observations—non-power-of-two widths in 41.8% of candidates, resolution-wise capacity correlations—are descriptive facts about the generated population under the proxy; they don't collapse entirely if the proxy is imperfect, but their usefulness for architecture design does depend on proxy validity.

The citation pattern is reasonable: heavy self-citation of [16] is appropriate since this is a direct continuation, and the LEMUR/NNGPT ecosystem citations are context. No red flags there.

Who's this for: people working on LLM-based NAS and proxy evaluation. The paper is honest and gives a good template for accounting in generated-code search.

Bottom line: deserves peer review, with a request for a replication run or proper variance estimate and validation of the proxy against full training. As is, treat its headline as conditional.

Referee Report

3 major / 3 minor

Summary. The paper scales a closed-loop LLM-based channel-configuration search on CIFAR-100 from a sparse setting to 250 generated candidates per cycle across eight complete cycles (2000 candidates total, 462 strict CIFAR-100 evaluations). It reports a positive linear trend in per-cycle mean accuracy (slope 9.87e-4, p=0.043), stronger improvement in the high-performing frontier (best accuracy rising from 0.3144 to 0.3676, with positive top-5 and top-10 per-cycle mean trends), improved parameter efficiency of later-cycle top models, and architectural regularities (41.8% of architectures contain non-power-of-two widths; higher low-resolution parameter share associates with higher one-epoch accuracy). The paper includes a meticulous candidate-accounting protocol (Table 1), a confound audit for hidden code changes (Section 7.4), and a self-contained analysis script that rebuilds all tables and figures from raw folders.

Significance. If the central claims hold, the paper provides a useful scaled empirical analysis of LLM-driven channel search, with a transparent reporting protocol for generated-code NAS experiments. The strengths are the strict data accounting, the explicit confound audit, the leave-cycle-out predictive analysis, and the candid limitations. The main significance is conditional: the accuracy, efficiency, and architectural conclusions all rest on one-epoch CIFAR-100 accuracy as a proxy, and the headline frontier improvement is a single order statistic from one run. The paper acknowledges these threats but does not yet resolve them.

major comments (3)

[Section 10 (Limitations)] The sole dependent variable throughout the paper is one-epoch CIFAR-100 accuracy. Section 10 states: 'No candidate was trained for more than one epoch, and a full-training run on the top candidates would clarify whether one-epoch rankings correlate with final rankings.' This is not a peripheral caveat: the accuracy trends (Table 2, Fig. 2), the Pareto-frontier claims (Fig. 3), and the channel-pattern associations (Figs. 6–8) are all measured against this proxy. If one-epoch accuracy does not preserve full-training rankings, the claimed frontier improvement (0.3144 → 0.3676), the parameter-efficiency comparison (A6/B102 vs A0/B134), and the architectural regularities could be artifacts of early-learning speed. The paper provides no internal validation (e.g., a subsample of candidates trained for a full schedule) and no citation establishing that one-epoch CIFAR-100 rankings track final ra
[Section 5 (Accuracy Trends Under Scaling)] The central scaling signal is the improvement of the best observed accuracy from 0.3144 (A0/B134) to 0.3676 (A6/B190). This is a single order statistic from a single run. Table 2 and the text report that bootstrap 95% confidence intervals on per-cycle means and maxima are wide and overlap across cycles. Section 10 itself notes that 'a second run with the same prompt, LLM, and protocol would produce a different frontier and would help separate a true accuracy shift from chance.' The top-5/top-10 mean trends are more stable (p < 0.02), but they are computed from only eight cycle-level points and are also from one run. I recommend either (a) additional independent runs (at least three) to estimate the distribution of the high-water mark, or (b) an internal control comparing the observed frontier improvement to a permutation/baseline distribution where cycle labels are shuffled, reporting th
[Section 7.3 (Dynamic PyTorch Evidence)] The Spearman correlations and p-values in Fig. 6 (e.g., ρ=0.477, p<10^-26; ρ=0.516, p<10^-8) treat the 462 strict evaluations as independent samples. They are not independent: candidates are generated sequentially by an LLM fine-tuned on earlier successful candidates (Section 3), so later cycles share lineage. The leave-cycle-out random-forest analysis (Fig. 7B) partially addresses this for prediction, but the correlation and matched-quartile tests in Figs. 6 and 7A do not. Similarly, the confound audit in Fig. 8 reports partial correlations with p-values computed under an i.i.d. assumption. I suggest a clustered permutation test (clustering by cycle) or a mixed-effects model with cycle as a random effect to confirm that the channel-feature associations and their p-values survive. Also, multiple features are tested (resolution shares, stage-wise ratios, widths), so some correction for mu

minor comments (3)

[References [3] and Section 3] The paper describes the skeleton as 'AirNet-like residual convolutional skeleton' and cites [3] (Chee et al., AirNet: Self-supervised affine registration for 3D medical images). That AirNet is a medical image registration network, not a residual CNN for classification. Either the citation is wrong or the architecture name is misleading. Please correct or clarify.
[Figure 2] The caption mentions a '95% confidence band' for per-cycle means, but the text does not describe how the band was computed (e.g., normal approximation, bootstrap, or t-interval). Please add a sentence describing the construction.
[Section 4] The strict-success rule references '1.json' as the epoch file. For readers unfamiliar with the pipeline, clarify whether '1.json' is the training result for epoch 1 and how it differs from eval_info.json.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the central claims are empirical measurements from newly collected data, with the acknowledged one-epoch proxy as a validity limitation rather than a circular derivation.

full rationale

The paper is an empirical scaling study, not a derivation. The claimed frontier improvement (0.3144 to 0.3676), per-cycle trends, parameter-efficiency comparison, and channel-pattern regularities are all computed from 462 newly verified CIFAR-100 evaluations generated in the described 8-cycle run, with a strict accounting protocol (Section 4) and a self-contained analysis script. The random-forest prediction uses leave-cycle-out cross-validation (Section 7.3), so it is not fitted on the test cycles. The one-epoch accuracy proxy is acknowledged in Section 10 as unvalidated against full training ("No candidate was trained for more than one epoch, and a full-training run on the top candidates would clarify whether one-epoch rankings correlate with final rankings"); that is a limitation on external validity, not a circular reduction, because the proxy is a fixed evaluation protocol rather than a quantity defined in terms of the paper's conclusions. Reliance on the authors' prior pipeline [16] and LEMUR/NNGPT infrastructure is methodological self-citation, but no load-bearing mathematical or statistical claim is justified solely by that citation. The paper also explicitly discloses the order-statistic caveat and the need for a second run, further indicating the claims are reported as observations rather than forced by construction.

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

The central claims rest on the one-epoch proxy validity, the statistical treatment of non-i.i.d. cycle data, and the strict filtering protocol. No free parameters are fit to data to produce the claims; LoRA hyperparameters are standard choices. No new physical or conceptual entities are introduced.

assumptions (4)

domain assumption One-epoch CIFAR-100 accuracy is a meaningful proxy for early learning and architecture quality.
Invoked throughout Sections 5-7; the paper notes in Section 10 that full-training correlation is not established.
domain assumption The eight complete cycles (A0-A7) can be treated as units for linear-trend regression with reported p-values.
Used for all slope p-values in Section 5; Section 10 discloses that generated samples are non-i.i.d. because LoRA fine-tuning depends on earlier cycles, but the p-values are not corrected for this.
domain assumption Strict filtering (1.json + eval_info.json confirming cifar-100) correctly identifies target-valid evaluations.
Defined in Section 4 and used for all accuracy and architecture claims; reasonable but a modeling choice.
domain assumption The AirNet-like residual skeleton and OlympicCoder-7B base model are fixed and representative.
Stated in Section 3; the paper acknowledges other LLMs/datasets may yield different channel priors.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Scaling Closed-Loop Feature Channel Configuration with LLMs." pith.science (2026). https://pith.science/paper/A3MBYVKY

@misc{pith2026260720516,
  author       = {Pith},
  title        = {Pith review of: Scaling Closed-Loop Feature Channel Configuration with LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A3MBYVKY}},
  note         = {Machine review of arXiv:2607.20516}
}

read the original abstract

Promising initial results in closed-loop large-language-model-based channel-configuration search demonstrated that neural-network widths can be optimized directly through executable code generation and accuracy feedback. However, those results were obtained from a relatively sparse set of valid evaluations, leaving open whether the observed optimization behavior transfers to a denser sampling regime and whether additional architectural regularities emerge when more generated networks are evaluated. To test this, the same search setting is scaled to 250 candidate networks per fine-tuning cycle. The analysis covers 2000 generated candidates from 8 complete cycles, yielding 462 verified CIFAR-100 evaluations after task and metadata filtering. Per-cycle mean accuracy exhibits a positive linear trend with slope 9.87e-4 (p=0.043), while the high-performing frontier improves more strongly: the best observed accuracy increases from 0.3144 to 0.3676, and both the top-5 and top-10 cycle-level means exhibit positive trends. The scaled run also reveals improved parameter efficiency. The best model reaches 0.3676 with 11.8M parameters, compared with an early high-performing model at 0.3144 with 166.5M parameters. Beyond accuracy, the larger sample exposes architectural regularities that were difficult to assess from sparse observations. Non-power-of-two channel widths occur in 41.8% of verified candidates, and the strongest models share structured channel-allocation patterns characterized by moderate early widths and expanded middle or later blocks. These findings indicate that the channel-search signal observed in the initial study transfers

Figures

Figures reproduced from arXiv: 2607.20516 by the authors.

**Figure 2.** Accuracy trends and frontier dynamics. Left: per-cycle mean accuracy for all strict CIFAR-100 candidates with a 95% confidence band, interquartile shading, median line, and top-5 mean. Right: the per-cycle best and cumulative high-water mark; the dashed reference is the best accuracy reported in the preceding study [16]. The trend p-values are computed on eight cycle-level points; bootstrap 95% confidence intervals … view at source ↗

**Figure 3.** Accuracy versus parameter count. Points are strict CIFAR-100 candidates, colored by generation cycle. The red curve is the empirical Pareto frontier. Late cycles include candidates that improve over the early high-parameter incumbent while using far fewer parameters. dominated by A6/B190 (lower accuracy at higher parameter counts), and the per-cycle maximum at matched parameter budgets is consistently higher in cyc… view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Channel profiles faceted by network depth. Within each depth group, lines are colored by accuracy (thin red = low, thick green = high). The pattern of wider middle blocks for better models holds across depth values. reaches 0.3276 and the best architecture with at leas…

**Figure 5.** Figure 5: Layer-local channel statistics. Left: non-power-of-two width rate by convolutional layer index. Right: Spearman correlation between layer width and validation accuracy. Later layers show stronger positive association with accuracy than the input stem. and pooling laye…

**Figure 6.** Figure 6: Resolution-aware dynamic analysis. Left: primary-convolution parameter allocation by output feature-map resolution. Top-decile and efficient high-accuracy candidates allocate more capacity to lower-resolution stages than bottom-decile candidates. Right: resolution-sp…

**Figure 7.** Figure 7: Matched-budget and predictive analyses. A: within each parameter-count quartile, candidates with higher low-resolution primary-convolution parameter share (green) outperform the lower half (gray), so the resolution effect is not a size artifact. B: predicted vs. obse…

**Figure 8.** Figure 8: Confound audit for the channel-configuration claims. A: partial Spearman ρ of four channel features under progressively stronger control sets. B: leave-cycle-out rank prediction for non-channel controls, controls plus channel features, and channel features only. C: top…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 16 linked inside Pith

[16]

In: Proceedings of the International Conference on Pattern Recognition (ICPR) (2026),https://arxiv.org/abs/2601.08517, to appear

Uzun, T.A., Ignatov, D., Timofte, R.: Closed-loop LLM discovery of non-standard channel priors in vision models. In: Proceedings of the International Conference on Pattern Recognition (ICPR) (2026),https://arxiv.org/abs/2601.08517, to appear

arXiv 2026
[1]

arXiv preprint arXiv:2101.08134 (2021)

Abdelfattah, M.S., Mehrotra, A., Dudziak, Ł., Lane, N.D.: Zero-cost proxies for lightweight NAS. arXiv preprint arXiv:2101.08134 (2021)

arXiv 2021
[2]

arXiv preprint arXiv:2502.20422 (2025)

Cai, Z., Tang, Y., Lai, Y., Wang, H., Chen, Z., Chen, H.: SEKI: Self-evolution and knowledge inspiration based neural architecture search via large language models. arXiv preprint arXiv:2502.20422 (2025)

arXiv 2025
[3]

arXiv preprint arXiv:1810.02583 (2018)

Chee, E., Wu, Z.: Airnet: Self-supervised affine registration for 3d medical images using neural networks. arXiv preprint arXiv:1810.02583 (2018)

arXiv 2018
[4]

In: Advances in Neural Information Processing Systems (2023)

Chen, A., Dohan, D.M., So, D.R.: EvoPrompting: Language models for code-level neural architecture search. In: Advances in Neural Information Processing Systems (2023)

2023
[5]

In: Advances in Neural Information Processing Systems (2023)

Dettmers, T., Pagnoni, A., Holtzman, A., Zettlemoyer, L.: QLoRA: Efficient fine- tuning of quantized LLMs. In: Advances in Neural Information Processing Systems (2023)

2023
[6]

arXiv preprint arXiv:2504.10552 (2025),https://arxiv.org/abs/2504.10552

Goodarzi, A.T., Kochnev, R., Khalid, W., Goudarzi, H.T., Qin, F., Uzun, T.A., Dhameliya, Y.S., Kathiriya, Y.K., Bentyn, Z.A., Ignatov, D., Timofte, R.: LEMUR neural network dataset: Towards seamless AutoML. arXiv preprint arXiv:2504.10552 (2025),https://arxiv.org/abs/2504.10552

arXiv 2025
[7]

In: Proceedings of the European Conference on Computer Vision (ECCV) (2018)

He, Y., Lin, J., Liu, Z., Wang, H., Li, L.J., Han, S.: AMC: AutoML for model compression and acceleration on mobile devices. In: Proceedings of the European Conference on Computer Vision (ECCV) (2018)

2018

Show all 24 references

[8]

arXiv preprint arXiv:2106.09685 (2021)

Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)

2021 arXiv
[9]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops

Kochnev, R., Khalid, W., Uzun, T.A., Zhang, X., Dhameliya, Y.S., Qin, F., Vysyaraju, C., Duvvuri, R., Goyal, A., Ignatov, D., Timofte, R.: NNGPT: Re- thinking AutoML with large language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogniti...

2026
[10]

arXiv preprint arXiv:2509.26037 (2025)

Li, Z., Lin, Z., Wang, Y.: CoLLM-NAS: Collaborative large language mod- els for efficient knowledge-guided neural architecture search. arXiv preprint arXiv:2509.26037 (2025)

2025 arXiv
[11]

arXiv preprint arXiv:1806.09055 (2018)

Liu, H., Simonyan, K., Yang, Y.: DARTS: Differentiable architecture search. arXiv preprint arXiv:1806.09055 (2018)

2018 arXiv
[12]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2019)

Liu, Z., Mu, H., Zhang, X., Guo, Z., Yang, X., Cheng, K.T., Sun, J.: MetaPruning: Meta learning for automatic neural network channel pruning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2019)

2019
[13]

In: Proceedings of the IEEE Interna- tional Conference on Computer Vision (2017)

Liu, Z., Li, J., Shen, Z., Huang, G., Yan, S., Zhang, C.: Learning efficient convolu- tional networks through network slimming. In: Proceedings of the IEEE Interna- tional Conference on Computer Vision (2017)

2017
[14]

arXiv preprint arXiv:2306.01102 (2023)

Nasir, M.U., Earle, S., Cleghorn, C.W., James, S., Togelius, J.: LLMatic: Neural architecture search via large language models and quality diversity optimization. arXiv preprint arXiv:2306.01102 (2023)

2023 arXiv
[15]

Penedo, G., Lozhkov, A., Kydlíček, H., Allal, L.B., Beeching, E., Lajarín, A.P., Gallouédec, Q., Habib, N., Tunstall, L., von Werra, L.: Olympiccoder.https: //huggingface.co/open-r1/OlympicCoder-7B(2025)

2025
[17]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops

Uzun, T.A., Khalid, W., Din, S.U., Mulukuledu, S.R., Singh, A., Vysyaraju, C., Duvvuri, R., Goyal, A., Lukhi, Y.R., A Hussain, M., Jesani, K., Shrestha, U., Mittal, Y., Kochnev, R., Kadam, P., Ikram, M., Moradiya, H., Arslanian, A., Ignatov, D., Timofte, R.: LEMUR 2: Unlocking...

2026
[18]

arXiv preprint arXiv:2305.05351 (2023)

Yu,C.,Liu,X.,Wang,Y.,Liu,Y.,Feng,W.,Xiong,D.,Tang,C.,Lv,J.:GPT-NAS: Evolutionary neural architecture search with the generative pre-trained model. arXiv preprint arXiv:2305.05351 (2023)

2023 arXiv
[19]

arXiv preprint arXiv:1903.11728 (2019)

Yu, J., Huang, T.S.: AutoSlim: Towards one-shot architecture search for channel numbers. arXiv preprint arXiv:1903.11728 (2019)

1903 arXiv
[20]

arXiv preprint arXiv:1902.08142 (2019)

Yu, K., Sciuto, C., Jaggi, M., Musat, C., Salzmann, M.: Evaluating the search phase of neural architecture search. arXiv preprint arXiv:1902.08142 (2019)

1902 arXiv
[21]

Zheng, M., Su, X., You, S., Wang, F., Qian, C., Xu, C., Albanie, S.: Can GPT-4 perform neural architecture search? arXiv preprint arXiv:2304.10970 (2023)

2023 arXiv
[22]

arXiv preprint arXiv:2001.01233 (2020)

Zhou, D., Zhou, X., Zhang, W., Loy, C.C., Yi, S., Zhang, X., Ouyang, W.: EcoNAS: Finding proxies for economical neural architecture search. arXiv preprint arXiv:2001.01233 (2020)

2001 arXiv
[23]

arXiv preprint arXiv:2408.11330 (2024)

Zhou, X., Wu, X., Feng, L., Lu, Z., Tan, K.C.: Design principle transfer in neural architecture search via large language models. arXiv preprint arXiv:2408.11330 (2024)

2024 arXiv
[24]

arXiv preprint arXiv:1611.01578 (2016)

Zoph, B., Le, Q.V.: Neural architecture search with reinforcement learning. arXiv preprint arXiv:1611.01578 (2016)

2016 arXiv

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [16]

In: Proceedings of the International Conference on Pattern Recognition (ICPR) (2026),https://arxiv.org/abs/2601.08517, to appear

Uzun, T.A., Ignatov, D., Timofte, R.: Closed-loop LLM discovery of non-standard channel priors in vision models. In: Proceedings of the International Conference on Pattern Recognition (ICPR) (2026),https://arxiv.org/abs/2601.08517, to appear

arXiv 2026

[2] [1]

arXiv preprint arXiv:2101.08134 (2021)

Abdelfattah, M.S., Mehrotra, A., Dudziak, Ł., Lane, N.D.: Zero-cost proxies for lightweight NAS. arXiv preprint arXiv:2101.08134 (2021)

arXiv 2021

[3] [2]

arXiv preprint arXiv:2502.20422 (2025)

Cai, Z., Tang, Y., Lai, Y., Wang, H., Chen, Z., Chen, H.: SEKI: Self-evolution and knowledge inspiration based neural architecture search via large language models. arXiv preprint arXiv:2502.20422 (2025)

arXiv 2025

[4] [3]

arXiv preprint arXiv:1810.02583 (2018)

Chee, E., Wu, Z.: Airnet: Self-supervised affine registration for 3d medical images using neural networks. arXiv preprint arXiv:1810.02583 (2018)

arXiv 2018

[5] [4]

In: Advances in Neural Information Processing Systems (2023)

Chen, A., Dohan, D.M., So, D.R.: EvoPrompting: Language models for code-level neural architecture search. In: Advances in Neural Information Processing Systems (2023)

2023

[6] [5]

In: Advances in Neural Information Processing Systems (2023)

Dettmers, T., Pagnoni, A., Holtzman, A., Zettlemoyer, L.: QLoRA: Efficient fine- tuning of quantized LLMs. In: Advances in Neural Information Processing Systems (2023)

2023

[7] [6]

arXiv preprint arXiv:2504.10552 (2025),https://arxiv.org/abs/2504.10552

Goodarzi, A.T., Kochnev, R., Khalid, W., Goudarzi, H.T., Qin, F., Uzun, T.A., Dhameliya, Y.S., Kathiriya, Y.K., Bentyn, Z.A., Ignatov, D., Timofte, R.: LEMUR neural network dataset: Towards seamless AutoML. arXiv preprint arXiv:2504.10552 (2025),https://arxiv.org/abs/2504.10552

arXiv 2025

[8] [7]

In: Proceedings of the European Conference on Computer Vision (ECCV) (2018)

He, Y., Lin, J., Liu, Z., Wang, H., Li, L.J., Han, S.: AMC: AutoML for model compression and acceleration on mobile devices. In: Proceedings of the European Conference on Computer Vision (ECCV) (2018)

2018