Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

An LLM reads a dataset's statistics and writes a complete object-detection network architecture from scratch, outperforming nano-scale YOLO baselines on several benchmarks.

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 ↗

An LLM with retrieval-augmented generation synthesizes object-detection architectures directly from dataset meta-features; the resulting models are competitive with YOLOv11n but at 5.6–6.7M parameters, not the 1.9M stated in the abstract.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection The one-shot LLM-from-data synthesis idea is genuinely new, but the abstract's 1.9M parameter claim contradicts the tables, and the nano-scale comparisons are confounded, so the empirical case as written doesn't hold. the 5 major comments →

arxiv 2512.12281 v2 pith:6CHHU4XB submitted 2025-12-13 cs.CV

Cognitive-YOLO: LLM-Driven Architecture Synthesis from First Principles of Data for Object Detection

Jiahao Zhao This is my paper
classification cs.CV
keywords data-driven architecture synthesisLLMobject detectionYOLOneural architecture description languagedataset profilingretrieval-augmented generationautomated model design
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.

The reading

The paper tries to show that a modern object-detection network can be designed by an LLM in one shot, using only a small set of quantitative facts about the target dataset — object scale histogram, scene density, class imbalance, brightness/contrast — plus a library of known high-performance modules. The claim is that the LLM's ability to reason about why and where to place those modules matters more than the quality of the modules themselves. If correct, this replaces expensive neural architecture search with minutes of inference, and makes custom architectures for vertical domains (rail inspection, agriculture, fire safety) practical for non-experts. Across five datasets, the synthesized models at 5.6–6.7M parameters match or exceed nano YOLO baselines at 2.0–2.7M parameters on mAP@0.5:0.95, which the paper interprets as disproportionate returns, evidence of structural superiority rather than just extra parameters.

Core claim

The central discovery claim is that architecture synthesis from dataset 'first principles' works: the Data-Driven Architect Agent maps meta-features to retrieval queries, the LLM (instantiated with Gemini 2.5 Pro) turns the results into a Neural Architecture Description Language (NADL) blueprint — a JSON description of modules, connections, and parameters — and a hybrid compiler instantiates it into a trainable YOLO-family model. The paper argues that the LLM's synthesis stage is the primary driver of performance: in ablations on the Rail Surface Defect dataset, removing the dataset profile drops accuracy from 74.3% to 71.8% mAP@0.5:0.95, while removing RAG-module retrieval drops it only to

What carries the argument

The load-bearing machinery is the three-stage Analyze–Synthesize–Compile pipeline. Analyze: a Dataset Profiler extracts meta-features (object scale histogram, per-image density, class imbalance, brightness/contrast) and a ReAct agent retrieves candidate modules from a curated SOTA knowledge base. Synthesize: the LLM reasons over these inputs to generate the Neural Architecture Description Language (NADL), a structured JSON blueprint specifying modules, connections, and parameters. Compile: a rule-based transpiler turns the NADL into Ultralytics YAML configs, with a validation layer that checks topology and channel compatibility before training. The NADL is what lets the LLM act as a holistic

Load-bearing premise

The claim stands or falls on whether the dataset meta-features and the LLM's reasoning — not the added parameters or chance — are what produce the accuracy gains; the paper's ablations do not hold parameter count fixed.

What would settle it

Train the best synthesized architecture (Rail Surface Defect, 6.7M) with its parameter count trimmed to YOLOv12n's 2.5M, or train a YOLOv12n scaled up to 6.7M, and compare mAP@0.5:0.95. If the synthesized model's advantage disappears at matched budget, the 'data-first structural superiority' claim collapses.

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

If this is right

  • If the synthesis claim holds, vertical-scene detector customization becomes a one-shot, minutes-long process: profile the data, let the LLM write the blueprint, compile, train — no NAS search loop.
  • The ablation result implies that dataset understanding is worth more than a curated module library; teams with modest module collections can still get strong architectures if their reasoning is data-grounded.
  • The cost model shifts from N search iterations times training to a single training run, making architecture design practical for small labs and edge deployment.
  • If the same reasoning transfers, the approach extends to instance segmentation and pose estimation, where the same meta-features should guide head and neck choices.

Where Pith is reading between the lines

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

  • A natural test the paper does not run: hold parameter count constant and compare the synthesized architecture against a YOLOv12n resized to 6.7M parameters; this would separate topology effects from budget effects. If the gain persists at matched budget, the 'structurally superior' claim is strongly supported; if not, the result may be partially a size effect.
  • If the first-principles mapping is causal, then perturbing a single meta-feature (e.g., inflating the small-object fraction) should produce a measurable, predictable architectural change (e.g., a higher-resolution feature map or a transformer decoder). A reader could verify this by feeding the LLM two synthetic dataset profiles that differ only in scale distribution.
  • The paper's qualitative heatmap evidence suggests a broader interpretability claim: LLM-driven architectures produce cleaner feature maps because the reasoning step explicitly targets background suppression. This is testable on any dataset with high clutter.
Share X Bluesky LinkedIn Reddit HN

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

5 major / 5 minor

Summary. This paper presents Cognitive-YOLO, a system that uses an LLM (Gemini 2.5 Pro) operating under a ReAct-paradigm agent to synthesize object-detection architectures from a profile of dataset meta-features (object scale histogram, scene density, class imbalance, brightness/contrast). The pipeline has three stages: a Dataset Profiler extracts meta-features; a Data-Driven Architect Agent retrieves candidate modules from a curated SOTA module knowledge base via RAG; and the LLM 'holistic architect' reasons over these inputs to emit a structured Neural Architecture Description Language (NADL) blueprint; a hybrid compiler instantiates the blueprint as Ultralytics YAML or PyTorch code, with a validation layer and CI/CD training loop. The authors evaluate the approach on five datasets (rail surface defects, rice disease, fire detection, drone detection, student behavior), comparing against YOLOv5n-YOLOv12n with models of 5.6-6.7M parameters, and report an ablation holding parameters fixed across variants with and without the dataset profile and RAG. They claim the synthesized models show superior parameter efficiency, that the gains are a 'disproportionate performance return,' and that data-driven 'first principles' reasoning—not the module library—is the primary driver of performance, with a case study on fire detection and qualitative feature-map visualizations as supporting evidence.

Significance. If the central claims were established, the paper would make a useful contribution: replacing multi-iteration LLM-guided NAS with a one-shot, data-conditioned synthesis is an appealing paradigm, and the decoupled NADL/compiler design is a clean interface between LLM output and executable code. The constant-parameter ablation design (Table 2) is the right kind of control, and the honesty of the section 4.1 caveat about parameter-scale mismatch is appreciated. However, the manuscript's own evidence does not support its claims: the abstract's 'down to 1.9M' contradicts Tables 1-2, the comparisons are capacity-confounded with no same-scale manual baseline, the measured gains at 2-3x parameters are modest (0.2-2.7 mAP), and the causal attribution to 'first principles' reasoning is not tested against any same-parameter standard architecture or variance across seeds. The strengths (clear pipeline description, five-domain evaluation, structured ablation, transparent case study) cannot offset the unsubstantiated central empirical claim.

major comments (5)
  1. [Abstract; §4.1; Table 1; Figure 3] The abstract's headline claim that synthesized models 'markedly compress the parameter count (down to 1.9M)' contradicts the paper's own results: every Cognitive-YOLO row in Tables 1 and 2 reports 5.6-6.7M parameters, and §4.1 explicitly states the models 'belong to different magnitude classes in scale' from the 1.8-2.7M nano baselines. The architecture in Figure 3 (six C2f stages with widths up to 1024 plus an RTDETRDecoder) cannot correspond to a 1.9M model. Since parameter efficiency is a central selling point, this internal inconsistency directly undermines the abstract's claim; the abstract, tables, and §4.1 must be made consistent, and the 'down to 1.9M' claim either substantiated with the corresponding model or removed.
  2. [§4.1; Table 1; §5] Every claimed advantage is measured against nano-scale baselines (1.8-2.7M) using models of 5.6-6.7M parameters, and no same-parameter manual architecture is evaluated. The 'disproportionate performance return' argument is not established: on Rail Surface Defect, 2.7x parameters (2.5M->6.7M) buys +2.7 mAP, but on Drone Detection 2.1x parameters buys +0.9 mAP, on Fire Detection 2.5x parameters buys +0.2 mAP, and on Student Behavior 2.4x parameters buys +0.5 mAP. These small absolute gains are consistent with ordinary capacity scaling; without width-matched YOLO variants (e.g., a ~6M YOLOv12s or width-scaled YOLOv12n), the claims of 'structurally superior and task-specific' design and 'SOTA-level performance in its parameter class' (§5) have no baseline against which to be tested.
  3. [§4.2; Table 2] The internal ablation correctly holds parameters fixed (6.7/6.7/6.7 M on Rail Surface Defect) and shows the dataset profile contributes about +2.5 mAP (71.8->74.3), which is a genuine, well-designed control. However, the interpretation overreaches in two ways. First, the claim that 'Without RAG' (72.5) beating YOLOv12n (71.6) 'demonstrates that the LLM's synthesis capability is the primary driver' is confounded by the 2.7x parameter difference. Second, all results are single runs with no seeds or variance; deltas of 0.3-2.5 mAP across conditions could reflect noise. The sentence 'This proves our core hypothesis' is therefore too strong; repeated runs and a same-parameter standard-architecture baseline are needed.
  4. [Abstract; §3.1; §5] The abstract claims a 'fully automated Analyze-Synthesize-Compile pipeline,' but §3.1 states that modules that cannot be automatically parsed 'are flagged for manual implementation and verification by experts,' and §5 concedes 'the current knowledge base construction still relies on some expert intervention.' This is an internal inconsistency between the stated contribution and the acknowledged limitation. In addition, the CI/CD 'closed-loop' described in §3.3 implies iterative refinement toward performance, yet no iteration results are reported; all numbers in Tables 1-2 appear to come from a single synthesis pass, so the loop's role is not evidenced.
  5. [§4.1; §4.5] The experimental section omits information required to check any result in Tables 1-2: the five datasets are named but never described (size, class distribution, train/test splits); no training hyperparameters (epochs, batch size, image resolution, optimizer, learning rate, augmentation) are given; the contents and versions of the SOTA module library are unspecified; and no example of a complete NADL blueprint or the LLM prompt is provided, despite NADL being central to the method. The cost claims in §4.5 (a 1-minute dataset analysis, a <5-minute NADL generation) are likewise unsupported. Without these, the central empirical claims cannot be reproduced or independently verified.
minor comments (5)
  1. [References] The reference list contains an editorial artifact: ref [16] includes the text 'As stated in the paper, placeholder year might be 2025,' and several entries have empty '[Online; accessed 2025-08-25]' fields. The bibliography should be cleaned and verified before submission.
  2. [§4.5; general] Typos and grammatical slips: 'less than 5 minute'; 'Our cost is primarily twofold'; 'but machines can only solve this task after the rise of deep learning' (§2.1) is an overstatement. Also, Table 3's quoted strings have visible line-break artifacts ('sparse scenes and numerous' / 'negative samples').
  3. [§3.2; §3.3] The paper defines NADL and describes the prompting strategy but never shows a complete NADL JSON example or any prompt template, even in an appendix. Since NADL is the core interface of the framework, one full example would substantially aid reproducibility and clarity.
  4. [§4.4; Figure 4] A single qualitative feature-map example is reasonable as illustration, but the caption's claim that the visualization 'explains its superior mAP' overreaches; the comparison between a C3k2-layer activation and a C2f-layer activation is not normalized or quantified, so the interpretation should be softened.
  5. [§3.3] The compiler section mentions generation for PyTorch and torch.js, but no experiments validate these non-Ultralytics paths; either present a minimal demonstration or clearly scope the claim to the Ultralytics path used in Table 1.

Circularity Check

0 steps flagged

No circularity: empirical LLM-guided architecture pipeline with external validation; confounded baselines are a correctness issue, not a circular one.

full rationale

This paper contains no equation-level derivation in which an output is defined from the quantity it supposedly predicts. The pipeline is empirical: dataset meta-features are profiled, an LLM produces a NADL blueprint from those features plus retrieved modules, a compiler instantiates the model, and it is trained and evaluated on held-out test data. The reported mAP values are independent outcomes, not re-statements of the inputs. There are no self-citations; the references are external LLM/NAS/detector works, so no load-bearing uniqueness theorem or ansatz is imported from the authors' own prior work. The Table 2 ablation holds parameter count constant across variants (e.g., 6.7M for all Rail Surface Defect rows), so 'Without Dataset Profile' vs full is a genuine input manipulation, however confounded by the absence of a same-scale fixed-architecture baseline. The serious weaknesses are experimental, not circular: Table 1 compares 5.6-6.7M models against 1.8-3.2M nano baselines, and the paper itself concedes this in Section 4.1: 'We note that the models automatically generated by Cognitive-YOLO (approx. 5.6M-6.7M parameters) and these nano-scale baselines (approx. 1.8M-3.2M parameters) belong to different magnitude classes in scale. Therefore, the purpose of this experiment is not to claim a direct SOTA replacement at the nano-scale budget.' The abstract's claim of models 'down to 1.9M' parameters is inconsistent with Table 1. These are correctness and evidence problems - an unsupported causal attribution and an internal inconsistency - not circularity, because no result is true by construction of its own definition. The 'first principles' chain in Section 4.3 is a post-hoc narrative rather than a derivation; that is an overclaim, but it does not reduce the result to its inputs. No circular step can be exhibited, so the score is 0.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 1 invented entities

The central claim rests on several unverified design choices: the profiling features, the curated module library, the proprietary LLM's reasoning, and the compiler's rule set. None of these are justified by first principles or by controlled comparison; they are free parameters of the system. The paper provides no code or data to independently audit these choices.

free parameters (3)
  • Dataset profiler meta-feature set = not specified
    The choice of which quantitative features to extract (object scale histogram bins, per-image density mean/max, class imbalance ratios, brightness/contrast) is selected by the authors and is not derived from theory or experiments. Different profiling choices could lead to different architectures.
  • SOTA module library composition = curated via web crawler and expert review
    The knowledge base (C2f, hgnetv2_b0, CSWin tiny, RepC3, AIFI DyT, BiFusion, RTDETRDecoder, etc.) is an expert-curated set; the inclusion/exclusion of modules affects the search space and results.
  • LLM prompt templates and ReAct agent iteration count = not specified
    The prompting strategy and the termination condition of the agentic retrieval loop are not disclosed; these are critical free choices that influence the generated NADL and hence performance.
axioms (4)
  • domain assumption Ultralytics YAML instantiation yields models that can be trained to a fair comparison with the baselines.
    The compiler transpiles NADL to YAML and the authors assume this is equivalent to coding the model directly; they further assume the same training/validation protocol is used, but no training hyperparameters are provided (Section 3.3, 4.1).
  • ad hoc to paper The LLM (Gemini 2.5 Pro) can reliably reason about architecture design from natural-language meta-features.
    The entire synthesis relies on the proprietary LLM's capability to map high-level dataset descriptions to module choices. This capability is not independently established; it is assumed to be present (Section 3.2, 4.2).
  • domain assumption The dataset meta-features capture the 'first principles' needed for architecture design.
    Object scale distribution, density, imbalance, brightness, and contrast are assumed to be the key drivers of architecture performance. The paper does not compare against a comprehensive feature set or a theory of what matters (Section 3.1).
  • domain assumption The reported mAP values are computed with a consistent, unbiased evaluation protocol across all models.
    No dataset splits, I/O thresholds, or calibration details are given; the authors assume standard Ultralytics metrics are applied equally (Section 4.1).
invented entities (1)
  • Neural Architecture Description Language (NADL) no independent evidence
    purpose: A structured JSON representation of network architectures used as the intermediate output of the LLM and input to the compiler.
    NADL is a new software artifact introduced by the paper. It has no falsifiable external handle; it is purely a descriptive format, so it does not function as a scientific entity.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Cognitive-YOLO: LLM-Driven Architecture Synthesis from First Principles of Data for Object Detection." pith.science (2026). https://pith.science/paper/6CHHU4XB

@misc{pith2026251212281,
  author       = {Pith},
  title        = {Pith review of: Cognitive-YOLO: LLM-Driven Architecture Synthesis from First Principles of Data for Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6CHHU4XB}},
  note         = {Machine review of arXiv:2512.12281}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

General-purpose object detection algorithms face a domain-shift bottleneck when deployed in vertical scenarios, and manually customizing a network architecture relies heavily on expert experience with high trial-and-error cost. To address these problems, we present \textbf{Cognitive-YOLO}, an adaptive object-detection model synthesis system driven by a large language model (LLM) collaborating with an autonomous agent. Cognitive-YOLO builds a fully automated \emph{Analyze--Synthesize--Compile} pipeline. First, a data-profiling module quantitatively extracts dataset features, guiding a ReAct-paradigm agent to autonomously retrieve matching network modules from a curated component library. Second, the LLM performs logical reasoning and topological assembly conditioned on the data features, producing a specification-compliant neural architecture description. Finally, a backend hybrid compiler dynamically instantiates the model, closing the loop with end-to-end training and evaluation scheduling. Cross-validation on datasets including rail-surface defects and rice diseases shows that the lightweight models synthesized by Cognitive-YOLO markedly compress the parameter count (down to 1.9M) while matching or surpassing mainstream baselines such as YOLOv11n in mAP@0.5:0.95, effectively balancing model compactness and feature-representation capability.

Figures

Figures reproduced from arXiv: 2512.12281 by Jiahao Zhao.

Figure 1
Figure 1. Figure 1: Comparison between past and present model design [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The Cognitive-YOLO Workflow. Our framework is a three-stage ”analyze-synthesize-compile” pipeline. (1) Analyze: The Dataset Profiler extracts meta-features, and an Agentic Search queries the SOTA Module Database. (2) Synthesize: The LLM Architect (Gemini 2.5 Pro) reasons on these inputs to generate the NADL blueprint. (3) Compile: The hybrid Compiler instantiates the NADL into deployable code for multiple … view at source ↗
Figure 3
Figure 3. Figure 3: Taking fire detection as an example, compare the differ [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 3
Figure 3. Figure 3: This reasoning chain—linking specific data prop [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of internal feature maps (Heatmaps) on the ’Fire Detection’ dataset. This figure provides visual proof of our ’first principles’ approach. (b) The feature map from the base￾line YOLOv12n is extremely noisy, with high-activation hotspots (red areas) scattered diffusely across background elements. (c) In stark contrast, the activation in our Cognitive-YOLO—which benefits from the Transformer En… view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. A Hybrid Vision-Language Architecture for Automated Defect Reasoning and Report Generation in Industrial Inspection

    cs.CV 2026-05 unverdicted novelty 6.0

    A decoupled pipeline with YOLO detection, deterministic prompt encoding, and QLoRA-adapted 1.5B LLM achieves superior structured report generation compared to monolithic VLMs on synthetic maintenance data.

Reference graph

Works this paper leans on

32 extracted references · 3 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Super-gradients, 2021

    Shay Aharon, Louis-Dupont, Ofri Masad, Kate Yurkova, Lotem Fridman, Lkdci, Eugene Khvedchenya, Ran Rubin, Natan Bagrov, Borys Tymchenko, Tomer Keren, Alexander Zhilko, and Eran-Deci. Super-gradients, 2021. 2

    2021

  2. [2]

    Qwen2.5- vl technical report, 2025

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5- vl technical report, 2025. 2

    2025

  3. [3]

    Yolov4: Optimal Speed and Accuracy of Object Detec- tion, 2020

    Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal Speed and Accuracy of Object Detec- tion, 2020. [Online; accessed 2025-08-25]. 2

    2020

  4. [4]

    Language models are few- shot learners.Advances in Neural Information Processing Systems, 33:1877–1901

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Sub- biah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakan- tan, Pranav Shyam, Girish Sastry, Amanda Askell, Sand- hini Agarwal, Ariel Herbert-V oss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jef- frey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin,...

    1901

  5. [5]

    End-to- end object detection with transformers, 2020

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to- end object detection with transformers, 2020. 2

    2020

  6. [6]

    Yolo-world: Real-time open- vocabulary object detection

    Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xing- gang Wang, and Ying Shan. Yolo-world: Real-time open- vocabulary object detection. In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 16901–16911. IEEE, 2024. 2

    2024

  7. [7]

    Zhang, Han Bao, Hanwei Xu, Haocheng Wang, Haowei Zhang, Honghui Ding, Huajian Xin, Huazuo Gao, Hui Li, Hui Qu, J

    DeepSeek-AI, Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Daya Guo, Dejian Yang, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, Fuli Luo, Guangbo Hao, Guanting Chen, Guowei Li, H. Zhang, Han Bao, Hanwei Xu, Haocheng Wang, Haowei Zhang, Honghui Ding, Huaj...

  8. [8]

    DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junx- iao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai...

  9. [9]

    Bert: Pre-training of deep bidirectional transform- ers for language understanding, 2018

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transform- ers for language understanding, 2018. 2

    2018

  10. [10]

    LPZero: Language model zero-cost proxy search from zero

    Peijie Dong, Lujun Li, Xiang Liu, Zhenheng Tang, Xuebo Liu, Qiang Wang, and Xiaowen Chu. LPZero: Language model zero-cost proxy search from zero. InFindings of the Association for Computational Linguistics: EMNLP 2024,

    2024

  11. [11]

    Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation

    Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Ma- lik. Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation. InProceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition (CVPR),

  12. [12]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Ab- hinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Ko- renev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aure- lien Rodriguez, Austen Gregerson,...

    2024

  13. [13]

    Rae, Oriol Vinyals, and Laurent Sifre

    Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, Tom Hennigan, Eric Noland, Katie Millican, George van den Driessche, Bogdan Damoc, Aurelia Guy, Simon Osindero, Karen Simonyan, Erich Elsen, Jack W. Rae, Oriol Vinyals, and Laurent Sifre...

    2022

  14. [14]

    Deim: Detr with improved match- ing for fast convergence

    Shihua Huang, Zhichao Lu, Xiaodong Cun, Yongjun Yu, Xiao Zhou, and Xi Shen. Deim: Detr with improved match- ing for fast convergence. In2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 15162–15171. IEEE, 2025. 2

    2025

  15. [15]

    Ganesh Jawahar, Muhammad Abdul-Mageed, Laks V . S. Lak- shmanan, and Dujian Ding. LLM performance predictors are good initializers for architecture search.arXiv preprint arXiv:2310.16712, 2023. Version 2, 7 Aug 2024. 3

    Pith/arXiv arXiv 2023

  16. [16]

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

    Zipeng Ji, Guanghui Zhu, Chunfeng Yuan, and Yihua Huang. RZ-NAS: Enhancing LLM-guided neural architecture search via reflective zero-cost strategy. InProceedings of the 42nd International Conference on Machine Learning (ICML), 2025. As stated in the paper, placeholder year might be 2025. 3

    2025

  17. [17]

    Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models, 2020. 1

    2020

  18. [18]

    Retrieval-augmented generation for knowledge- intensive nlp tasks

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K¨uttler, Mike Lewis, Wen-tau Yih, Tim Rockt¨aschel, Sebastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge- intensive nlp tasks. InProceedings of the 34th International Conference on Neural Information Processing Systems, Red H...

    2020

  19. [19]

    Zen-nas: A zero-shot nas for high-performance image recognition

    Ming Lin, Pichao Wang, Zhenhong Sun, Hesen Chen, Xiuyu Sun, Qi Qian, Hao Li, and Rong Jin. Zen-nas: A zero-shot nas for high-performance image recognition. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 347–356, 2021. 3

    2021

  20. [20]

    M. U. Nasir, S. Earle, J. Togelius, S. James, and C. Cleghorn. LLMatic: Neural architecture search via large language mod- els and quality-diversity optimization. InProceedings of the Genetic and Evolutionary Computation Conference (GECCO ’24), pages 1110–1118. Association for Computing Machin- ery, 2024. 3

    2024

  21. [21]

    ChatGPT can now see, hear, and speak, 2023

    OpenAI. ChatGPT can now see, hear, and speak, 2023. Ac- cessed: 2025-08-25. 1

    2023

  22. [22]

    GPT-4 Technical Report

    OpenAI. GPT-4 Technical Report. Technical Report arXiv:2303.08774, OpenAI, 2023

    Pith/arXiv arXiv 2023

  23. [23]

    GPT-4V(ision) System Card

    OpenAI. GPT-4V(ision) System Card. Technical report, OpenAI, 2023. Accessed: 2025-08-25. 1

    2023

  24. [24]

    Language models are unsuper- vised multitask learners

    Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsuper- vised multitask learners. 2019. 2

    2019

  25. [25]

    You only look once: Unified, real-time object detec- tion, 2015

    Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detec- tion, 2015. 1

    2015

  26. [26]

    Agentic retrieval-augmented generation: A survey on agentic rag, 2025

    Aditi Singh, Abul Ehtesham, Saket Kumar, and Tala Talaei Khoei. Agentic retrieval-augmented generation: A survey on agentic rag, 2025. 3 11

    2025

  27. [27]

    Llama: Open and Efficient Foundation Language Models, 2023

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Mar- tinet, Marie-Anne Lachaux, Timoth ´ee Lacroix, Baptiste Rozi`ere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and Efficient Foundation Language Models, 2023. [Online; accessed 2025-08-25]. 2

    2023

  28. [28]

    Github - ultralytics/yolov5: Yolov5 in PyTorch > ONNX > CoreML > TFLite,

    ultralytics. Github - ultralytics/yolov5: Yolov5 in PyTorch > ONNX > CoreML > TFLite, . [Online; accessed 2025-08- 25]. 2

    2025

  29. [29]

    Github - ultralytics/ultralytics: Ultralytics YOLO,

    ultralytics. Github - ultralytics/ultralytics: Ultralytics YOLO, . [Online; accessed 2025-08-25]. 2

    2025

  30. [30]

    Chain-of-thought prompting elicits reasoning in large lan- guage models, 2022

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large lan- guage models, 2022. 3

    2022

  31. [31]

    React: Synergizing reasoning and acting in language models, 2023

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models, 2023. 3 12

    2023

  32. [2024]

    arXiv preprint arXiv:2410.04808. 3

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.