LLM-Driven Neural Network Generation with Same-Family Architecture Guidance: Disentangling Transfer and Adaptation

# LLM-Driven Neural Network Generation with Same-Family Architecture Guidance: Disentangling Transfer and Adaptation
Source: [https://arxiv.org/html/2607.05704](https://arxiv.org/html/2607.05704)
Kabir Dev Paul Baghel Radu Timofte Dmitry Ignatov Computer Vision Lab, CAIDAS, University of Würzburg, Germany

###### Abstract

Large language models \(LLMs\) can generate neural\-network modifications, but unrestricted generation is often invalid or harmful\. This paper studies a narrower setting: improving a weak target model using a stronger same\-family source model from a neural\-network database\. We propose a source\-guided candidate\-generation protocol with non\-source controls, source\-conditioned candidates, and a no\-LLMhp\_copyablation under equal evaluation budgets\. The protocol reports validity separately from accuracy and selects the best valid candidate only when it improves the target\. On CIFAR\-10, the strongest source\-guided candidate reaches 0\.5049 accuracy versus 0\.2398 for the best non\-source candidate, a \+0\.2651 advantage, while improving a weak target originally at 0\.1254; a five\-epoch check preserves the gain at 0\.7686 versus 0\.4839\. On SVHN AlexNet with DeepSeek\-Coder\-6\.7B, source\-guided transfer reaches 0\.7880 versus 0\.2254, a \+0\.5626 advantage; a fresh repeat reaches 0\.8069 versus 0\.2509, a \+0\.5560 advantage\. Direct source\-recipe copy produces 0\.1959 on SVHN AlexNet, matching the original target, whilehp\_transferreaches 0\.7880, showing that the LLM adapts rather than copies the source recipe\. Family\-level analysis shows the clearest positive signals for AlexNet, with 6/8 wins across SVHN, Imagenette, and CelebA\-Gender, andalt\_nn1, with 8/10 wins on CIFAR\-10\.

## 1Introduction

LLM\-based neural\-network generation is attractive because it searches directly over executable model code and training recipes\. Instead of designing a hand\-written search space, one can ask a model to propose candidate modifications and then train the result\. Prior work in the ABrain/LEMUR ecosystem has explored LLMs for architecture generation, prompt design, hyperparameter tuning, and iterative NAS with feedback memory\[[11](https://arxiv.org/html/2607.05704#bib.bib1),[5](https://arxiv.org/html/2607.05704#bib.bib14),[10](https://arxiv.org/html/2607.05704#bib.bib4),[8](https://arxiv.org/html/2607.05704#bib.bib16)\]\. The practical failure mode is clear: unrestricted generation often produces invalid or low\-quality code\.

This paper asks the following scientific question:

> Given a weak target model, does a stronger same\-family source model help an LLM generate better valid candidates than a non\-source baseline under the same evaluation budget?

This question is useful because it separates two different effects\. A weak model may improve simply because the LLM changes the learning rate, batch size, or transform\. That is a non\-source baseline effect\. The source\-guided effect only exists if access to a strong source model improves the outcome beyond the non\-source controls\.

The method is treated as a controlled candidate\-generation algorithm\. Validity, best valid accuracy, and mean valid accuracy are reported separately so that code\-generation reliability and model quality are not conflated\.

#### Contributions\.

This paper makes the following contributions:

- •It defines a source\-guided weak\-target candidate\-generation protocol with non\-source controls, source\-conditioned candidates, and a no\-LLMhp\_copyablation baseline\.
- •It provides evidence for two regimes, recipe\-transfer and recipe\-adaptation, characterized by thehp\_copydecomposition\.
- •It reports architecture\-family\-level win rates showing where source guidance works reliably and where it does not\.

## 2Related Work

This work sits at the intersection of neural architecture search, hyperparameter optimization, LLM code generation, and retrieval\-conditioned generation\.

#### Neural architecture search and AutoML\.

Classical neural architecture search \(NAS\) studies how to automate network design through a search space, a search strategy, and an evaluation strategy\[[6](https://arxiv.org/html/2607.05704#bib.bib19)\]\. Early work used reinforcement learning controllers to generate architectures\[[26](https://arxiv.org/html/2607.05704#bib.bib20)\], while later methods used evolutionary search\[[19](https://arxiv.org/html/2607.05704#bib.bib21)\], differentiable relaxations\[[16](https://arxiv.org/html/2607.05704#bib.bib22)\], and compound model scaling\[[22](https://arxiv.org/html/2607.05704#bib.bib23)\]\. AutoML\-Zero pushes the idea further by evolving learning algorithms from primitive operations rather than only choosing architectures\[[20](https://arxiv.org/html/2607.05704#bib.bib24)\]\. NAS methodology papers also emphasize reproducible search protocols and controlled comparisons\[[15](https://arxiv.org/html/2607.05704#bib.bib25)\]\. The present work lets an LLM propose candidate code/actions, but keeps the comparison scientific by fixing the candidate budget, validation rules, and target/source pairs\.

#### Hyperparameter optimization\.

Hyperparameter optimization is another close baseline because many improvements in this study come from training recipes rather than architecture changes\. Bayesian and sequential approaches to hyperparameter search were studied by Bergstra et al\.\[[2](https://arxiv.org/html/2607.05704#bib.bib26)\], Hyperband frames configuration selection as adaptive resource allocation\[[13](https://arxiv.org/html/2607.05704#bib.bib27)\], and Optuna provides a practical define\-by\-run optimization framework\[[1](https://arxiv.org/html/2607.05704#bib.bib28)\]\. Thehp\_defaultandhp\_transferarms in this paper can be read as LLM\-generated hyperparameter/transform candidates evaluated under the same spirit as black\-box HPO\.

#### LLMs for code generation and repair\.

Large language models trained on code, such as Codex, established that LLMs can synthesize executable Python from natural\-language specifications\[[4](https://arxiv.org/html/2607.05704#bib.bib29)\]\. AlphaCode showed that large\-scale sampling plus filtering can solve competitive\-programming tasks\[[14](https://arxiv.org/html/2607.05704#bib.bib30)\]\. Self\-refinement and agentic feedback methods such as Self\-Refine and Reflexion show that generation quality can improve when model outputs are evaluated and fed back into later attempts\[[17](https://arxiv.org/html/2607.05704#bib.bib31),[21](https://arxiv.org/html/2607.05704#bib.bib32)\]\. This paper uses the same broad generate–validate idea, but the validation target is neural\-network training/evaluation rather than unit tests or programming contest outputs\.

#### Retrieval, prompting, and source conditioning\.

Few\-shot prompting showed that examples in context can steer LLM behavior without parameter updates\[[3](https://arxiv.org/html/2607.05704#bib.bib33)\]\. Retrieval\-augmented generation adds non\-parametric evidence to the prompt so that generation can depend on external documents or memories\[[12](https://arxiv.org/html/2607.05704#bib.bib34)\]\. ReAct combines language\-model reasoning with actions against an external environment\[[25](https://arxiv.org/html/2607.05704#bib.bib35)\]\. The source\-guided setting here is a specialized retrieval problem where the retrieved item is a stronger neural network with code, transforms, hyperparameters, and measured accuracy\.

#### LLMs as optimizers and LLM\-based NAS\.

Recent work treats LLMs as components of optimization loops\. OPRO uses LLMs to propose candidate solutions from a prompt containing previous candidates and scores\[[24](https://arxiv.org/html/2607.05704#bib.bib36)\]\. EvoPrompt connects LLMs with evolutionary algorithms for discrete prompt optimization\[[9](https://arxiv.org/html/2607.05704#bib.bib37)\]\. LLMatic combines LLM code generation with quality\-diversity optimization for NAS\[[18](https://arxiv.org/html/2607.05704#bib.bib38)\]\. These systems are closest in spirit to the present work\. The difference is that this paper studies a narrower weak\-target transfer question: given a weak model and a stronger same\-family source model, does source\-conditioned candidate generation produce different outcomes from non\-source controls?

#### ABrain/LEMUR context\.

The implementation builds directly on LLM\-based neural\-network synthesis in NNGPT\[[11](https://arxiv.org/html/2607.05704#bib.bib1)\], prompt\-based architecture generation\[[5](https://arxiv.org/html/2607.05704#bib.bib14)\], LLM\-based hyperparameter tuning\[[10](https://arxiv.org/html/2607.05704#bib.bib4)\], and resource\-efficient iterative NAS with feedback memory\[[8](https://arxiv.org/html/2607.05704#bib.bib16)\]\. It also uses LEMUR neural\-network datasets and metadata as the source of existing architectures, training recipes, and measured accuracies\[[7](https://arxiv.org/html/2607.05704#bib.bib5),[23](https://arxiv.org/html/2607.05704#bib.bib6)\]\. Compared with full\-code generation, this work restricts generation to a controlled contract: hyperparameters, transform code, structured JSON edits, or source\-guided hyperparameter/transform transfer\. This makes validity, repair policy, and source\-guided advantage easier to measure\.

## 3Method

### 3\.1Problem Setup

For a datasetDD, letTTbe a weak target neural network with measured accuracya​\(T\)a\(T\)and letSSbe a stronger same\-family source model with accuracya​\(S\)\>a​\(T\)a\(S\)\>a\(T\)\. Each model includes Python architecture code, a transform, and hyperparameters\. The objective is to generate one or more candidatesT1′,…,Tn′T^\{\\prime\}\_\{1\},\\ldots,T^\{\\prime\}\_\{n\}and select

T∗=arg⁡maxTi′∈𝒱⁡a​\(Ti′\),T^\{\*\}=\\arg\\max\_\{T^\{\\prime\}\_\{i\}\\in\\mathcal\{V\}\}a\(T^\{\\prime\}\_\{i\}\),\(1\)where𝒱\\mathcal\{V\}is the set of candidates that parse, satisfy the expected API, train/evaluate successfully, and produce a valid accuracy\. The target is improved only ifa​\(T∗\)\>a​\(T\)a\(T^\{\*\}\)\>a\(T\)\.

The target/source pairing is intentionally weak\-to\-strong\. The experiments test whether a high\-performing same\-family model contains transferable evidence that helps repair or improve a weak member of the same family\.

### 3\.2Retrieval

The retrieval rule used in the current experiments is simple and deterministic: for a weak target family, select a stronger same\-family model from the LEMUR database\. This produces pairs such as weak AlexNet to strong AlexNet, weak DarkNet to strong DarkNet, or weakalt\_nn1to strongalt\_nn1111Thealt\_nn1family used in CIFAR\-10 experiments contains AlexNet\-style custom CNNs \(5 Conv2d layers, 3 Linear layers, SGD training\); thearchbestsource preserves this style while thehighsourcesource may be a structurally different same\-family member such as a residual bottleneck network\.\. This same\-family restriction is important because it makes the source model relevant without requiring the LLM to translate across unrelated architectures\.

### 3\.3Candidate Action Spaces

The umbrella term in this paper is*candidate generation*\. Each LLM output is a candidate action that is trained and evaluated under the same budget\. The reported protocol uses four arms: two non\-source controls and two source\-conditioned arms\. This keeps the algorithm simple enough to reproduce while still testing whether the retrieved source adds value\.

ArmSourceLLMArch\.Recipehp\_defaultNoYesFixedChangedbaseline\_editNoYesChangedOptionalhp\_transferYesYesFixedChangedanalogical\_editYesYesChangedOptional

Table 1:Frozen four\-arm candidate\-generation protocol used in the main experiments\. “Recipe” denotes hyperparameters and input transforms\.Table[1](https://arxiv.org/html/2607.05704#S3.T1)defines the action space used in the main experiments\. The four arms are:

- •hp\_default: non\-source hyperparameter/transform control using only the target\.
- •baseline\_edit: non\-source structured edit using only the target\.
- •hp\_transfer: source\-guided transfer of the source hyperparameters and transform to the target\.
- •analogical\_edit: source\-conditioned structured architecture edit\.

Inhp\_transfer, the target architecture is held fixed\. The prompt exposes the source training recipe and asks the LLM to emit a target candidate recipe using the same output contract ashp\_default\. The fields allowed to change are the learning rate, momentum, batch size, dropout when supported, and input transform\. Validation only applies deterministic compatibility fixes, such as dropping a hyperparameter key unsupported by the trainer or replacing an invalid transform with the standard transform for that dataset\. It never copies the source architecture into the target and never uses measured accuracy to rewrite or select a candidate\.

Earlier screening runs included full\-code generation, delta edits, fixed templates, safe/risky source\-conditioned edits, and wider pooling templates\. Those runs are excluded from the main comparison; the representation screen in Table[2](https://arxiv.org/html/2607.05704#S3.T2)motivates the constrained action format used in Table[1](https://arxiv.org/html/2607.05704#S3.T1)\.

RepresentationValidBest acc\.RoleFull code0/16–Too brittleDelta patch14/160\.5557Evaluable but weakerStructured edit16/160\.6282Frozen protocol basis

Table 2:Historical CIFAR\-10 representation screening\. These runs are not averaged into the main results; they justify moving from unconstrained full\-code generation to constrained edit/action spaces\.
### 3\.4Validation and Minimal Repair

The main results use minimal deterministic validation\. Both non\-source and source\-guided arms receive the same cleanup: schema parsing, removal of unsupported hyperparameter keys, and dataset\-compatible transform guards\. Broader semantic repairs, fallback candidates, and accuracy\-based rewriting are excluded from the reported main comparison\.

### 3\.5Algorithm

The algorithm can be stated as follows:

1. 1\.Select a weak target modelTTon datasetDD\.
2. 2\.Retrieve the strongest or a high\-performing same\-family source modelSS\.
3. 3\.Generate non\-source baseline candidates without exposingSSto the LLM\.
4. 4\.Generate source\-guided candidates while exposingSSand its training recipe to the LLM\.
5. 5\.Train and evaluate all candidates under the same budget\.
6. 6\.Report valid generation/evaluation rate, best valid accuracy, gain over the original target, and source\-guided advantage over the best non\-source candidate\.
7. 7\.When choosing a final model, accept the best candidate only when it beats the original target; otherwise retainTT\.

## 4Experimental Setup

### 4\.1Metrics

The evaluation reports complementary metrics rather than collapsing every run into a single score:

- •valid evaluation count: how many candidate arms actually trained/evaluated;
- •best valid accuracy: the strongest evaluated candidate found by a group;
- •mean valid accuracy: average quality over evaluated candidates;
- •gain over original: best valid accuracy minus target accuracy;
- •source\-guided advantage: best source\-guided accuracy minus best non\-source accuracy, reported together with mean differences when available;
- •failure mode: whether failures come from no output, extraction failure, invalid architecture, shape mismatch, or training timeout\.

### 4\.2Statistical Analysis

The statistical unit for accuracy comparisons is a dataset–LLM–target pair\. For example, SVHN with DeepSeek\-Coder\-6\.7B on an AlexNet target is one pair\. A pair is called comparable only when both sides of the comparison have at least one valid evaluated candidate: one or more non\-source candidates and one or more source\-guided candidates\. This prevents invalid generations from being silently treated as low\-accuracy models while still reporting validity separately\.

For valid generation and valid evaluation rates, we report 95% Wilson confidence intervals\. This is used because validity is a binary outcome: each generated candidate either evaluates successfully or it does not\. Wilson intervals are more stable than the simple normal approximation when the number of samples is modest or the rate is close to 0 or 1\.

For source\-guided improvement, we use a paired bootstrap confidence interval over comparable dataset–LLM–target pairs\. For each pair, we compute

Δ=max⁡a​\(Tsource′\)−max⁡a​\(Tnon′\)\.\\Delta=\\max a\(T^\{\\prime\}\_\{\\mathrm\{source\}\}\)\-\\max a\(T^\{\\prime\}\_\{\\mathrm\{non\}\}\)\.\(2\)We then resample the comparable pairs with replacement and recompute the meanΔ\\Deltamany times\. The 2\.5th and 97\.5th percentiles form the bootstrap 95% confidence interval\. Differences with absolute value below 0\.001 are treated as practical ties\. We also report the win/tie/loss count as a simple non\-parametric summary\.

### 4\.3Implementation

Experiments are run in the TuneNNGen pipeline using LEMUR image\-classification models\. The LLM is loaded in the local workspace, generates candidate actions, and each valid candidate neural network is trained for one epoch\. TuneNNGen can support LoRA\-style LLM fine\-tuning, but post\-generation LLM fine\-tuning is disabled in the reported runs\. The results evaluate retrieval\-guided candidate generation rather than adaptation of the LLM itself\.

### 4\.4Reproducibility Protocol

Final experiments are launched from frozen queue scripts and logged with dataset, LLM config, target/source ids, target/source accuracies, action arm, repair mode, candidate budget, and seed offset\. The main CIFAR\-10 and SVHN AlexNet studies useN=32N=32attempted candidates per arm; smaller robustness probes useN=8N=8\. Reported runs use one neural\-network training epoch per candidate, temperature 0\.75, top\-k=50k=50, top\-p=0\.95p=0\.95, and prompt batch size 1\. The full command list, environment variables, software versions, ledger paths, and regeneration commands are stored in the reproducibility README\. Because the logged runs were produced without PyTorch deterministic mode, exact bitwise replay is not claimed; reproducibility is defined at the protocol level, using the same frozen scripts and reporting validity, best valid accuracy, and mean valid accuracy\.

### 4\.5LLMs

The current completed experiments include:

- •OlympicCoder\-7B;
- •DeepSeek\-Coder\-1\.3B\-Instruct;
- \(CompletedN=32N=32main\-table runs for CIFAR\-10 and SVHN AlexNet were not finished under the experimental budget and are therefore excluded from Tables[3](https://arxiv.org/html/2607.05704#S5.T3)and[4](https://arxiv.org/html/2607.05704#S5.T4); ds13b results appear in Table[10](https://arxiv.org/html/2607.05704#S5.T10)and contribute to the family win rates in Table[11](https://arxiv.org/html/2607.05704#S5.T11)\. Table[11](https://arxiv.org/html/2607.05704#S5.T11)pools all comparable logged pairs regardless ofNN; Tables[3](https://arxiv.org/html/2607.05704#S5.T3)and[4](https://arxiv.org/html/2607.05704#S5.T4)report only the frozenN=32N=32controlled comparison\.\)
- •DeepSeek\-Coder\-6\.7B\-Instruct;
- •Qwen2\.5\-Coder\-7B\-Instruct\.

The latter two models were added to test whether the effect is specific to OlympicCoder\-7B or survives with stronger/different code LLMs on a single RTX 4090\.

### 4\.6Datasets

Completed experiments cover CIFAR\-10, CIFAR\-100, SVHN, MNIST, Imagenette, CelebA\-Gender, and Places365\. The strongest evidence is concentrated on CIFAR\-10 and SVHN\. CIFAR\-100 and Imagenette are harder stress tests overall\. For non\-AlexNet architectures, source\-guided advantages are small or unstable on these datasets\. AlexNet is an exception: logged AlexNet runs on Imagenette show consistent positive results across all three LLMs tested, discussed in Section 5\.3\. MNIST is often too easy, CelebA\-Gender saturates near 0\.96 accuracy after dataset\-compatible transforms are enforced, and Places365 is too slow or brittle under the one\-epoch image\-classification evaluator\.

## 5Results

### 5\.1Main Source\-Guided Advantage Results

The main evidence usesN=32N=32attempted candidates*per arm*\. Non\-source candidates combinehp\_defaultandbaseline\_edit, giving 64 attempts per row\. Source\-guided candidates combinehp\_transferandanalogical\_edit, also giving 64 attempts per row\. CIFAR\-10 uses a weak target with original accuracy 0\.1254 and compares two source\-selection rules:highsource, the strongest same\-dataset source, andarchbest, the strongest source satisfying a closer architecture\-similarity constraint\. SVHN uses a weak AlexNet target with original accuracy 0\.1959 and a strong same\-family source with accuracy 0\.9524\.

LLMSourceNon v\.Non bestSrc v\.Src bestBest adv\.Mean adv\.DeepSeek6\.7Barchbest60/640\.239864/640\.5049\+0\.2651\+0\.2693DeepSeek6\.7Bhighsource62/640\.235463/640\.4213\+0\.1859\+0\.1582Qwen2\.5\-7Barchbest30/640\.215636/640\.4757\+0\.2601\+0\.2300Qwen2\.5\-7Bhighsource26/640\.238645/640\.4814\+0\.2428\+0\.1701Olympic7Barchbest49/640\.497158/640\.4806\-0\.0165\+0\.2536Olympic7Bhighsource42/640\.487656/640\.4033\-0\.0843\+0\.1699

Table 3:CIFAR\-10 mainN=32N=32per\-arm results\. Each group aggregates two arms, so each validity count is out of 64\. Source\-guided candidates win best valid accuracy in 4/6 settings and mean valid accuracy in 6/6\.LLMNon v\.Non bestSrc v\.Src bestBest adv\.Mean adv\.DeepSeek6\.7B64/640\.225463/640\.7880\+0\.5626\+0\.0170Qwen2\.5\-7B25/640\.195935/640\.19590\.00000\.0000Olympic7B45/640\.209344/640\.1965\-0\.0128\-0\.0022

Table 4:SVHN AlexNet mainN=32N=32per\-arm results\. DeepSeek\-Coder\-6\.7B gives the strongest cross\-dataset positive result; Qwen2\.5\-Coder\-7B ties and OlympicCoder\-7B slightly favors the non\-source group\.Tables[3](https://arxiv.org/html/2607.05704#S5.T3)and[4](https://arxiv.org/html/2607.05704#S5.T4)are the central results\. Across the nine main dataset–LLM–source settings, source\-guided candidates win best valid accuracy in 5 settings, tie in 1, and lose in 3\. Source\-guided arms also show somewhat higher valid\-evaluation rates in Figure[1](https://arxiv.org/html/2607.05704#S5.F1)\(83\.9% vs\. 80\.6% on CIFAR\-10\), which provides more evaluation attempts and may mechanically inflate best\-valid comparisons; mean valid accuracy comparisons are less affected by this asymmetry\. A budget\-corrected resampling check—capping source\-guided evaluation at the non\-source validity count for each row—confirms that the best\-valid advantage persists in 4/6 CIFAR\-10 settings at 100% of resamples; the two settings where it does not \(OlympicCoder\-7B\) are the same rows that show losses in the uncorrected comparison\. CIFAR\-10 gives the clearest repeated pattern: source\-guided candidates win best valid accuracy in 4/6 settings and win mean valid accuracy in 6/6\. The largest CIFAR\-10 gain is DeepSeek\-Coder\-6\.7B witharchbest: the best candidate improves from the original target accuracy of 0\.1254 to 0\.5049\. SVHN AlexNet with DeepSeek\-Coder\-6\.7B gives the strongest cross\-dataset positive case, with source\-guidedhp\_transferreaching 0\.7880 while the best non\-source candidate reaches 0\.2254\.

SetPairsW/T/LMean adv\.Median95% CIMainN=32N=3295/1/3\+0\.1559\+0\.1859\[\+0\.0363,\+0\.2905\]CIFAR\-1064/0/2\+0\.1422\+0\.2144\[\+0\.0294,\+0\.2465\]SVHN AlexNet31/1/1\+0\.18330\.0000\[\-0\.0128,\+0\.5626\]SVHN robustness91/0/8\-0\.0147\-0\.0078\[\-0\.0282,\-0\.0026\]

Table 5:Best\-valid source\-guided advantage summary\. W/T/L counts wins, practical ties, and losses using a tie threshold of 0\.001\. Confidence intervals are paired bootstrap intervals over dataset–LLM–target pairs\.The bootstrap summary in Table[5](https://arxiv.org/html/2607.05704#S5.T5)supports a narrow reading: the mainN=32N=32settings show a positive average best\-valid source\-guided advantage, and CIFAR\-10 is the strongest repeated positive dataset\. The SVHN AlexNet interval is wide because it contains only three LLM settings\. The additional SVHN robustness pairs are included as a stress test and do not show the same positive pattern\.

![Refer to caption](https://arxiv.org/html/2607.05704v1/figures/fig_main_validity_ci.png)Figure 1:Valid evaluation rates for the mainN=32N=32results\. Bars show observed valid\-evaluation rates; whiskers show Wilson 95% confidence intervals\. Source\-guided candidates have higher validity in the main grouped results, especially on CIFAR\-10\.
### 5\.2Repeats and Longer\-Training Checks

RunNon validNon bestSrc validSrc bestBest adv\.Original64/640\.225463/640\.7880\+0\.5626Fresh repeat64/640\.250964/640\.8069\+0\.5559

Table 6:Independent repeat of the strongest SVHN AlexNet/DeepSeek\-Coder\-6\.7B setting\. Both rows useN=32N=32candidates per arm and the same four\-arm protocol\. The repeated best\-performing arm is source\-guidedhp\_transfer\.Table[6](https://arxiv.org/html/2607.05704#S5.T6)repeats the strongest SVHN setting with a fresh seed offset\. The best source\-guided result remains much higher than the best non\-source result, and validity improves from 63/64 to 64/64 for the source\-guided group\.

The one\-epoch protocol is intended as a fast candidate search test\. To check whether the strongest one\-epoch winners survive longer training, the best valid non\-source and source\-guided candidates from selected settings were re\-evaluated for five epochs without generating new candidates\.

Dataset/LLMSourceNon 5epSrc 5epAdv\.CIFAR\-10/DeepSeek6\.7Barchbest0\.48390\.7686\+0\.2847CIFAR\-10/Qwen2\.5\-7Barchbest0\.49930\.7310\+0\.2317SVHN/DeepSeek6\.7BAlexNet0\.88520\.8985\+0\.0133SVHN/DeepSeek6\.7Brepeat0\.88370\.9047\+0\.0210

Table 7:Five\-epoch re\-evaluation of selected strongest one\-epoch settings\. CIFAR\-10 gains remain large; SVHN gains shrink because non\-source candidates catch up with longer training\.
### 5\.3Robustness and Failure Cases

This section collects stress tests that bound the main positive result\. They are useful because the central claim should survive contact with harder or saturated settings, but they are not weighted the same as the frozenN=32N=32CIFAR\-10 and SVHN AlexNet studies\. First, three additional SVHN weak\-source pairs were run withN=8N=8candidates per arm: BagNet, AirNext, and DPN68\. Table[8](https://arxiv.org/html/2607.05704#S5.T8)summarizes the best\-candidate advantage for each LLM/pair\.

PairLLMBest nonBest srcAdv\.Best src armBagNetDeepSeek6\.7B0\.83770\.8299\-0\.0078hp\_transferBagNetQwen2\.5\-7B0\.83270\.8303\-0\.0024hp\_transferBagNetOlympic7B0\.83900\.8329\-0\.0062hp\_transferAirNextDeepSeek6\.7B0\.76780\.7464\-0\.0215hp\_transferAirNextQwen2\.5\-7B0\.74910\.7473\-0\.0018analogical\_editAirNextOlympic7B0\.74990\.7654\+0\.0155hp\_transferDPN68DeepSeek6\.7B0\.71920\.6884\-0\.0308analogical\_editDPN68Qwen2\.5\-7B0\.72580\.7065\-0\.0193analogical\_editDPN68Olympic7B0\.79450\.7365\-0\.0580analogical\_edit

Table 8:SVHN robustness pairs withN=8N=8candidates per arm\. AirNext/Olympic7B gives a small positive source\-guided result, while BagNet and DPN68 favor non\-source candidates for best valid accuracy\.The SVHN robustness rows show that the large AlexNet gain does not automatically transfer to every SVHN architecture family\. BagNet and DPN68 already allow strong non\-source candidates, and source guidance does not improve the best candidate in those runs\. AirNext has one small positive row, but the overall robustness block is mostly boundary evidence rather than a second strong positive result\.

CIFAR\-100 and Imagenette are harder stress probes\. CIFAR\-100 keeps the small\-image geometry but increases the class count, while Imagenette changes both image scale and visual domain\. Table[9](https://arxiv.org/html/2607.05704#S5.T9)reports these smaller probes as scope evidence: the advantages are small or unstable, and one Imagenette setting produces no comparable valid candidates\.

DatasetLLMN/armBest nonBest srcAdv\.CIFAR\-100DeepSeek6\.7B40\.10910\.1069\-0\.0022CIFAR\-100Qwen2\.5\-7B40\.12800\.1313\+0\.0033CIFAR\-100Olympic7B40\.10780\.1164\+0\.0086ImagenetteDeepSeek6\.7B80\.12840\.1223\-0\.0061ImagenetteQwen2\.5\-7B8–––ImagenetteOlympic7B80\.25960\.2693\+0\.0097

Table 9:Harder\-dataset stress probes \(non\-AlexNet architectures\)\. These runs are smaller than the mainN=32N=32study and are reported as scope evidence: CIFAR\-100 gains are small, and Imagenette is unstable under the current one\-epoch evaluator\.DatasetLLMN/armBest non\-srcBest src\-guidedAdv\.CelebA\-GenderDeepSeek6\.7B80\.96490\.9632\-0\.0017ImagenetteDeepSeek1\.3B40\.29380\.3534\+0\.0596ImagenetteDeepSeek6\.7B40\.23110\.2492\+0\.0181ImagenetteOlympic7B40\.29400\.3656\+0\.0716

Table 10:AlexNet cross\-dataset results from logged runs\. Imagenette gives three consistent wins across all three LLMs tested\. CelebA\-Gender reaches saturation \(∼\\sim0\.96\) on both sides\.Table[10](https://arxiv.org/html/2607.05704#S5.T10)reports the AlexNet results on datasets beyond SVHN\. Imagenette gives three consistent source\-guided wins across all LLMs tested, while CelebA\-Gender saturates near 0\.96 accuracy on both sides\.

FamilyPairsWinsTiesLossesWin ratealt\_nn11080280\.0%AlexNet860275\.0%BagNet1260650\.0%AirNext831437\.5%DarkNet833237\.5%

Table 11:Architecture\-family\-level source\-guided win rates for families with at least five comparable pairs\. Win rate counts best\-valid source\-guided wins over comparable dataset–LLM–target rows\. Single\-row families are excluded from Table[11](https://arxiv.org/html/2607.05704#S5.T11)and treated as preliminary rather than stable evidence\. Table[11](https://arxiv.org/html/2607.05704#S5.T11)includes all comparable logged pairs meeting the validity criterion, including runs withN<32N<32that are excluded from the mainN=32N=32tables\. A small number of earlier exploratory rows predate LLM\-tag standardization and are logged without a specific LLM identifier; among the displayed families, this affects 1 pair in thealt\_nn1family and is included on the basis of the validity criterion alone\.CelebA\-Gender provides a separate saturation check: once dataset\-compatible image geometry is enforced, all main arms reach roughly the same 0\.96 accuracy\.

GroupValidBestMeanNon\-source16/160\.96320\.9405Source\-guided12/160\.96320\.9571Table 12:CelebA\-Gender AlexNet with DeepSeek\-Coder\-6\.7B and minimal geometry validation\. Best accuracy ties, while source\-guided candidates have a higher mean over valid candidates\.The main failure pattern is validity rather than only low accuracy\. Strict no\-repair CIFAR\-10 runs failed because small\-image architectures often collapsed spatial dimensions\. Imagenette candidates often fell below the evaluator accuracy threshold, and Places365 candidates often timed out\. Some LLMs produce fewer valid candidates for the same target, especially Qwen2\.5\-Coder\-7B on several SVHN arms\. These failures are reported separately because validity and model quality answer different questions\.

Failure modeTypical settingSpatial collapseCIFAR\-style small imagesUnsupported HP keysRaw LLM hyperparameter outputExtraction failureStructured edit armsBelow\-threshold accuracyImagenette stress probesTraining timeoutPlaces365 / large\-image runsSaturationMNIST, CelebA\-Gender

Table 13:Main failure and boundary modes observed so far\. Minimal repair addresses schema and geometry validity; stronger semantic repair/fallback is excluded from the main reported comparison\.
### 5\.4Decomposing Source\-Guided Advantage: Recipe Quality vs\. LLM Adaptation

The strongest repeated source\-guided arm in the main experiments ishp\_transfer, which transfers training\-recipe information while keeping the target architecture fixed\. To check whether this arm is simply copying a good source recipe, we add a no\-LLM ablation calledhp\_copy\. This arm directly copies the source hyperparameters and transform into the weak target architecture, applies the same deterministic validation and dataset\-compatible transform guard used by the other arms, and evaluates exactly one candidate\. It therefore separates the value of the retrieved source recipe itself from any LLM\-mediated adaptation of that recipe\.

Table[14](https://arxiv.org/html/2607.05704#S5.T14)shows the resulting decomposition for three key source/target settings\. The columns follow the candidate path from the original weak target, to the best non\-source candidate, to direct source\-recipe copy, and finally to source\-guidedhp\_transfer\. On CIFAR\-10,hp\_copyis already a strong baseline: forarchbest, direct recipe copy reaches 0\.4267, and forhighsource, it reaches 0\.3506\. In both cases,hp\_transferimproves further, reaching 0\.5049 and 0\.4213 respectively\. This suggests a recipe\-transfer regime: much of the improvement comes from the quality of the retrieved training recipe, with the LLM adding a smaller additional gain\.

SettingTargetNon\-src besthp\_copyhp\_transferRecipe gainLLM gainCIFAR\-10archbest0\.12540\.23980\.42670\.5049\+0\.1869\+0\.0782CIFAR\-10highsource0\.12540\.23540\.35060\.4213\+0\.1152\+0\.0707SVHN AlexNet0\.19590\.22540\.19590\.7880\-0\.0295\+0\.5921Imagenette AlexNet0\.23260\.29400\.28920\.3656\-0\.0048\+0\.0764

Table 14:Decomposition of source\-guided gains\.hp\_copydirectly copies the source recipe without LLM generation\. Recipe gain is measured over the best non\-source candidate; LLM gain is measured fromhp\_copytohp\_transfer\. Sincehp\_copyis deterministic and LLM\-independent, its accuracy is fixed per source/target pair; the SVHN AlexNet entry therefore applies across all LLMs tested on that pair\. Imagenette AlexNet is an intermediate case: direct recipe copy improves on the original target \(0\.2892 vs\. 0\.2326\) but does not exceed the best non\-source candidate, whilehp\_transferreaches 0\.3656\.SVHN AlexNet behaves differently\. Directly copying the source recipe gives 0\.1959, matching the original weak target rather than improving it\. The copied source recipe is: lr=6\.747×10−5=6\.747\\times 10^\{\-5\}, momentum=0\.980=0\.980, batch=256=256, transform=norm\_299\_flip=\\texttt\{norm\\\_299\\\_flip\}\. This learning rate is calibrated for a near\-perfect model and is far too small to train a weak target in one epoch\. In this setting,hp\_transferreaches 0\.7880 whilehp\_copyremains at 0\.1959, a gap of \+0\.5921 attributable to LLM\-mediated adaptation of the source recipe\. This is a recipe\-adaptation regime rather than a direct recipe\-transfer regime\.

A fourth setting, Imagenette AlexNet, is intermediate: direct recipe copy improves the target \(hp\_copy= 0\.2892 vs\. original 0\.2326\) but does not reach the best non\-source candidate \(0\.2940\), whilehp\_transferreaches 0\.3656\. The source recipe for Imagenette \(lr = 0\.0189, moderate momentum\) is more suitable for a weak target than the SVHN fine\-tuning recipe, but LLM adaptation still adds meaningful gain\.

These results falsify the simple pass\-through copier explanation forhp\_transfer\. On CIFAR\-10, direct source copying explains a substantial part of the gain, but does not match the best source\-guided result\. On SVHN AlexNet, direct copying fails whilehp\_transfersucceeds\. The evidence therefore supports a narrower mechanism: source recipes can be useful either because they are already transferable, or because the LLM adapts them into a recipe that is more suitable for the weak target\.

## 6Discussion

#### The main positive pattern\.

The clearest pattern is organized by architecture family rather than by dataset alone\. AlexNet gives 6/8 source\-guided best\-valid wins across SVHN, Imagenette, and CelebA\-Gender — specifically, wins on SVHN \(3/4 LLMs\) and Imagenette \(3/3 LLMs\), one narrow loss on SVHN \(OlympicCoder\-7B\), and one saturation boundary case on CelebA\-Gender\.alt\_nn1gives 8/10 wins on CIFAR\-10\. These are the strongest family\-level signals in the current ledger and are more informative than treating all dataset–LLM rows as exchangeable\. Thehp\_copydecomposition gives a mechanistic account of this pattern: in CIFAR\-10, direct source\-recipe copy already improves the weak target andhp\_transferimproves further; in SVHN AlexNet, direct copy fails because the source learning rate is too small for the weak target, whilehp\_transferadapts the recipe and recovers a large gain\. The empirical picture is therefore not one uniform effect, but two regimes: recipe\-transfer when the source recipe is already compatible with the target, and recipe\-adaptation when the retrieved recipe needs to be rewritten before it helps\.

#### Interpreting best and mean accuracy\.

Best valid accuracy measures whether a candidate\-generation arm can find a strong model under a fixed budget\. Mean valid accuracy measures how consistently the arm produces useful evaluated candidates\. Validity explains how much of the generation budget survives to evaluation\. These quantities answer different questions, so the tables report them separately where possible\.

#### Scope of the result\.

The evidence supports source\-guided candidate generation as an additional search arm alongside target\-only search\. The strongest observed mechanism is hyperparameter/transform transfer from the source training recipe\. Source\-conditioned architecture edits are included in the four\-arm protocol, but they are less stable thanhp\_transfer\.

#### Candidate selection\.

The practical algorithm evaluates non\-source and source\-guided candidates under the same budget, then keeps the best valid candidate only if it improves the original target\. This matches the empirical pattern: source guidance produces the largest wins in some regimes, while non\-source baselines win or tie in others\.

## 7Limitations

The current study has several limitations:

- •The strongest positive results still depend on a small number of target/source regimes, especially CIFAR\-10 and SVHN AlexNet\. SVHN AlexNet now has an independent repeat, but the other positive regimes still need repeated runs\.
- •The five\-epoch checks in Table[7](https://arxiv.org/html/2607.05704#S5.T7)cover only the strongest one\-epoch winners\. The rank order among candidates is not validated at longer training horizons, and one\-epoch rankings may not preserve for settings outside the four shown\.
- •For non\-AlexNet architectures, CIFAR\-100 and Imagenette remain harder stress tests where source\-guided advantages are small or unstable\. The AlexNet family is an exception, with consistent Imagenette wins across all three LLMs tested \(Table[10](https://arxiv.org/html/2607.05704#S5.T10)\)\.
- •MNIST is too easy, CelebA\-Gender saturates, and Places365 is too slow or brittle under the current evaluator\.
- •The strongest current result is often hyperparameter/transform transfer, so the paper should avoid overstating architectural discovery\.
- •Thehp\_copydecomposition is currently run on only four settings\. Extending it to more dataset–family pairs would test whether the recipe\-transfer versus recipe\-adaptation split tracks the source–target accuracy gap\.
- •Minimal repair is used in the clean protocol\. It is applied symmetrically, but dataset\-compatible transform guards may still affect measured accuracy and should be ablated further\.
- •Current generation runs do not yet prove that LoRA fine\-tuning improves the LLM; they mainly test source\-guided generation\.

Broader future work should splithp\_transferinto hyperparameter\-only and transform\-only variants, compare strong sources with random or weaker same\-family sources, and repeat the strongest settings under longer training budgets and additional datasets\.

## 8Conclusion

This paper presents a source\-guided candidate\-generation protocol for improving weak neural networks with LLMs\. For some weak same\-family targets, retrieved source evidence improves the best valid candidate over non\-source controls, especially through hyperparameter/transform transfer\. The effect is dataset\-, LLM\-, target\-family\-, and action\-space\-dependent\. The most practical use is to evaluate source\-guided and target\-only arms together, then keep the best valid candidate\.

## References

- \[1\]\(2019\)Optuna: a next\-generation hyperparameter optimization framework\.InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining,pp\. 2623–2631\.Cited by:[§2](https://arxiv.org/html/2607.05704#S2.SS0.SSS0.Px2.p1.1)\.
- \[2\]J\. S\. Bergstra, R\. Bardenet, Y\. Bengio, and B\. Kegl\(2011\)Algorithms for hyper\-parameter optimization\.InAdvances in Neural Information Processing Systems,Vol\.24,pp\. 2546–2554\.Cited by:[§2](https://arxiv.org/html/2607.05704#S2.SS0.SSS0.Px2.p1.1)\.
- \[3\]T\. B\. Brown, B\. Mann, N\. Ryder, M\. Subbiah, J\. Kaplan, P\. Dhariwal, A\. Neelakantan, P\. Shyam, G\. Sastry, A\. Askell, S\. Agarwal, A\. Herbert\-Voss, G\. Krueger, T\. Henighan, R\. Child, A\. Ramesh, D\. M\. Ziegler, J\. Wu, C\. Winter, C\. Hesse, M\. Chen, E\. Sigler, M\. Litwin, S\. Gray, B\. Chess, J\. Clark, C\. Berner, S\. McCandlish, A\. Radford, I\. Sutskever, and D\. Amodei\(2020\)Language models are few\-shot learners\.InAdvances in Neural Information Processing Systems,Vol\.33,pp\. 1877–1901\.Cited by:[§2](https://arxiv.org/html/2607.05704#S2.SS0.SSS0.Px4.p1.1)\.
- \[4\]M\. Chen, J\. Tworek, H\. Jun, Q\. Yuan, H\. P\. d\. O\. Pinto, J\. Kaplan, H\. Edwards, Y\. Burda, N\. Joseph, G\. Brockman, A\. Ray, R\. Puri, G\. Krueger, M\. Petrov, H\. Khlaaf, G\. Sastry, P\. Mishkin, B\. Chan, S\. Gray, N\. Ryder, M\. Pavlov, A\. Power, L\. Kaiser, M\. Bavarian, C\. Winter, P\. Tillet, F\. P\. Such, D\. Cummings, M\. Plappert, F\. Chantzis, E\. Barnes, A\. Herbert\-Voss, W\. H\. Guss, A\. Nichol, A\. Paino, N\. Tezak, J\. Tang, I\. Babuschkin, S\. Balaji, S\. Jain, W\. Saunders, C\. Hesse, A\. N\. Carr, J\. Leike, J\. Achiam, V\. Misra, E\. Morikawa, A\. Radford, M\. Knight, M\. Brundage, M\. Murati, K\. Mayer, P\. Welinder, B\. McGrew, D\. Amodei, S\. McCandlish, I\. Sutskever, and W\. Zaremba\(2021\)Evaluating large language models trained on code\.arXiv preprintarXiv:2107\.03374\.External Links:[Link](https://arxiv.org/abs/2107.03374)Cited by:[§2](https://arxiv.org/html/2607.05704#S2.SS0.SSS0.Px3.p1.1)\.
- \[5\]R\. Duvvuri, C\. Vysyaraju, A\. Goyal, D\. Ignatov, and R\. Timofte\(2026\)Enhancing LLM\-based neural network generation: few\-shot prompting and efficient validation for automated architecture design\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops \(CVPRW\),Note:to appearCited by:[§1](https://arxiv.org/html/2607.05704#S1.p1.1),[§2](https://arxiv.org/html/2607.05704#S2.SS0.SSS0.Px6.p1.1)\.
- \[6\]T\. Elsken, J\. H\. Metzen, and F\. Hutter\(2019\)Neural architecture search: a survey\.Journal of Machine Learning Research20\(55\),pp\. 1–21\.Cited by:[§2](https://arxiv.org/html/2607.05704#S2.SS0.SSS0.Px1.p1.1)\.
- \[7\]A\. T\. Goodarzi, R\. Kochnev, W\. Khalid, F\. Qin, T\. A\. Uzun, Y\. S\. Dhameliya, Y\. K\. Kathiriya, Z\. A\. Bentyn, D\. Ignatov, and R\. Timofte\(2025\)LEMUR neural network dataset: towards seamless AutoML\.arXiv preprintarXiv:2504\.10552\.External Links:[Link](https://arxiv.org/abs/2504.10552)Cited by:[§2](https://arxiv.org/html/2607.05704#S2.SS0.SSS0.Px6.p1.1)\.
- \[8\]X\. Gu, D\. Ignatov, and R\. Timofte\(2026\)Resource\-efficient iterative LLM\-based NAS with feedback memory\.arXiv preprintarXiv:2603\.12091\.External Links:[Link](https://arxiv.org/abs/2603.12091)Cited by:[§1](https://arxiv.org/html/2607.05704#S1.p1.1),[§2](https://arxiv.org/html/2607.05704#S2.SS0.SSS0.Px6.p1.1)\.
- \[9\]Q\. Guo, R\. Wang, J\. Guo, B\. Li, K\. Song, X\. Tan, G\. Liu, J\. Bian, and Y\. Yang\(2023\)Connecting large language models with evolutionary algorithms yields powerful prompt optimizers\.arXiv preprintarXiv:2309\.08532\.External Links:[Link](https://arxiv.org/abs/2309.08532)Cited by:[§2](https://arxiv.org/html/2607.05704#S2.SS0.SSS0.Px5.p1.1)\.
- \[10\]R\. Kochnev, A\. T\. Goodarzi, Z\. A\. Bentyn, D\. Ignatov, and R\. Timofte\(2025\)Optuna vs code llama: are LLMs a new paradigm for hyperparameter tuning?\.InProceedings of the IEEE/CVF International Conference on Computer Vision Workshops \(ICCVW\),pp\. 5664–5674\.External Links:[Link](https://openaccess.thecvf.com/content/ICCV2025W/AIM/papers/Kochnev_Optuna_vs_Code_Llama_Are_LLMs_a_New_Paradigm_for_ICCVW_2025_paper.pdf)Cited by:[§1](https://arxiv.org/html/2607.05704#S1.p1.1),[§2](https://arxiv.org/html/2607.05704#S2.SS0.SSS0.Px6.p1.1)\.
- \[11\]R\. Kochnev, W\. Khalid, T\. A\. Uzun, X\. Zhang, Y\. S\. Dhameliya, F\. Qin, C\. Vysyaraju, R\. Duvvuri, A\. Goyal, D\. Ignatov, and R\. Timofte\(2026\)NNGPT: rethinking AutoML with large language models\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops \(CVPRW\),Note:to appearCited by:[§1](https://arxiv.org/html/2607.05704#S1.p1.1),[§2](https://arxiv.org/html/2607.05704#S2.SS0.SSS0.Px6.p1.1)\.
- \[12\]P\. Lewis, E\. Perez, A\. Piktus, F\. Petroni, V\. Karpukhin, N\. Goyal, H\. Kuttler, M\. Lewis, W\. Yih, T\. Rocktaschel, S\. Riedel, and D\. Kiela\(2020\)Retrieval\-augmented generation for knowledge\-intensive NLP tasks\.InAdvances in Neural Information Processing Systems,Vol\.33,pp\. 9459–9474\.Cited by:[§2](https://arxiv.org/html/2607.05704#S2.SS0.SSS0.Px4.p1.1)\.
- \[13\]L\. Li, K\. Jamieson, G\. DeSalvo, A\. Rostamizadeh, and A\. Talwalkar\(2018\)Hyperband: a novel bandit\-based approach to hyperparameter optimization\.Journal of Machine Learning Research18\(185\),pp\. 1–52\.Cited by:[§2](https://arxiv.org/html/2607.05704#S2.SS0.SSS0.Px2.p1.1)\.
- \[14\]Y\. Li, D\. Choi, J\. Chung, N\. Kushman, J\. Schrittwieser, R\. Leblond, T\. Eccles, J\. Keeling, F\. Gimeno, A\. Dal Lago, T\. Hubert, P\. Choy, C\. de Masson d’Autume, I\. Babuschkin, X\. Chen, P\. Huang, J\. Welbl, S\. Gowal, A\. Cherepanov, J\. Molloy, D\. Mankowitz, E\. S\. Robson, P\. Kohli, N\. de Freitas, K\. Kavukcuoglu, and O\. Vinyals\(2022\)Competition\-level code generation with AlphaCode\.Science378\(6624\),pp\. 1092–1097\.Cited by:[§2](https://arxiv.org/html/2607.05704#S2.SS0.SSS0.Px3.p1.1)\.
- \[15\]M\. Lindauer and F\. Hutter\(2020\)Best practices for scientific research on neural architecture search\.Journal of Machine Learning Research21\(243\),pp\. 1–18\.Cited by:[§2](https://arxiv.org/html/2607.05704#S2.SS0.SSS0.Px1.p1.1)\.
- \[16\]H\. Liu, K\. Simonyan, and Y\. Yang\(2019\)DARTS: differentiable architecture search\.InInternational Conference on Learning Representations,Cited by:[§2](https://arxiv.org/html/2607.05704#S2.SS0.SSS0.Px1.p1.1)\.
- \[17\]A\. Madaan, N\. Tandon, P\. Gupta, S\. Hallinan, L\. Gao, S\. Wiegreffe, U\. Alon, N\. Dziri, S\. Prabhumoye, Y\. Yang, S\. Gupta, B\. P\. Majumder, K\. Hermann, S\. Welleck, A\. Yazdanbakhsh, and P\. Clark\(2023\)Self\-refine: iterative refinement with self\-feedback\.InAdvances in Neural Information Processing Systems,Vol\.36\.Cited by:[§2](https://arxiv.org/html/2607.05704#S2.SS0.SSS0.Px3.p1.1)\.
- \[18\]M\. U\. Nasir, S\. Earle, C\. Cleghorn, S\. James, and J\. Togelius\(2023\)LLMatic: neural architecture search via large language models and quality diversity optimization\.arXiv preprintarXiv:2306\.01102\.External Links:[Link](https://arxiv.org/abs/2306.01102)Cited by:[§2](https://arxiv.org/html/2607.05704#S2.SS0.SSS0.Px5.p1.1)\.
- \[19\]E\. Real, A\. Aggarwal, Y\. Huang, and Q\. V\. Le\(2019\)Regularized evolution for image classifier architecture search\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.33,pp\. 4780–4789\.External Links:[Document](https://dx.doi.org/10.1609/aaai.v33i01.33014780)Cited by:[§2](https://arxiv.org/html/2607.05704#S2.SS0.SSS0.Px1.p1.1)\.
- \[20\]E\. Real, C\. Liang, D\. R\. So, and Q\. V\. Le\(2020\)AutoML\-Zero: evolving machine learning algorithms from scratch\.arXiv preprintarXiv:2003\.03384\.External Links:[Link](https://arxiv.org/abs/2003.03384)Cited by:[§2](https://arxiv.org/html/2607.05704#S2.SS0.SSS0.Px1.p1.1)\.
- \[21\]N\. Shinn, F\. Cassano, E\. Berman, A\. Gopinath, K\. Narasimhan, and S\. Yao\(2023\)Reflexion: language agents with verbal reinforcement learning\.InAdvances in Neural Information Processing Systems,Vol\.36\.Cited by:[§2](https://arxiv.org/html/2607.05704#S2.SS0.SSS0.Px3.p1.1)\.
- \[22\]M\. Tan and Q\. V\. Le\(2019\)EfficientNet: rethinking model scaling for convolutional neural networks\.InProceedings of the 36th International Conference on Machine Learning,pp\. 6105–6114\.Cited by:[§2](https://arxiv.org/html/2607.05704#S2.SS0.SSS0.Px1.p1.1)\.
- \[23\]T\. A\. Uzun, W\. Khalid, S\. U\. Din, S\. R\. Mulukuledu, A\. Singh, C\. Vysyaraju, R\. Duvvuri, A\. Goyal, Y\. R\. Lukhi, A\. Hussain, K\. Jesani, U\. Shrestha, Y\. Mittal, R\. Kochnev, P\. Kadam, M\. Ikram, H\. R\. Moradiya, A\. Arslanian, D\. Ignatov, and R\. Timofte\(2026\)LEMUR 2: unlocking neural network diversity for AI\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops \(CVPRW\),Note:to appearCited by:[§2](https://arxiv.org/html/2607.05704#S2.SS0.SSS0.Px6.p1.1)\.
- \[24\]C\. Yang, X\. Wang, Y\. Lu, H\. Liu, Q\. V\. Le, D\. Zhou, and X\. Chen\(2023\)Large language models as optimizers\.arXiv preprintarXiv:2309\.03409\.External Links:[Link](https://arxiv.org/abs/2309.03409)Cited by:[§2](https://arxiv.org/html/2607.05704#S2.SS0.SSS0.Px5.p1.1)\.
- \[25\]S\. Yao, J\. Zhao, D\. Yu, N\. Du, I\. Shafran, K\. Narasimhan, and Y\. Cao\(2023\)ReAct: synergizing reasoning and acting in language models\.InInternational Conference on Learning Representations,Cited by:[§2](https://arxiv.org/html/2607.05704#S2.SS0.SSS0.Px4.p1.1)\.
- \[26\]B\. Zoph and Q\. V\. Le\(2016\)Neural architecture search with reinforcement learning\.arXiv preprintarXiv:1611\.01578\.External Links:[Link](https://arxiv.org/abs/1611.01578)Cited by:[§2](https://arxiv.org/html/2607.05704#S2.SS0.SSS0.Px1.p1.1)\.