[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-86574-en":3,"doc-seo-86574-105":30,"detail-sidebar-cat-0-en-105":91},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":21,"is_downloadable":21,"audit_status":21,"page_count":22,"language":23,"language_code":24,"site_id":25,"html_lang":24,"table_of_contents":26,"faqs":27,"seo_title":13,"seo_description":14,"update_tm":28,"read_time":29},86574,1099514068365,"Aurelia","https://ap-avatar.wpscdn.com/avatar/10000253d8d9f28188e?_k=1776742907772140068",8,"Research & Report","Similarity Guided Curriculum Fine Tuning of LLMs for Neural Architecture Synthesis","Similarity-Guided Curriculum Fine-Tuning introduces a MinHash-based similarity scheduling framework for LLM-based neural architecture search (NAS). It builds a progressive curriculum over neural architecture code by partitioning a reference pool into similarity bands using 128-permutation MinHash signatures from normalized 7-gram source-code shingles. A best LoRA adapter from each stage is merged cumulatively into the backbone. Evaluated on OlympicCoder-7B in the LEMUR benchmark, it reports strong peak success at high-similarity without post-processing repair. Ablations and transfer to SVHN analyze interactions between curriculum warmup and interface repair, separating distinct failure modes.","Similarity-Guided Curriculum Fine-Tuning of LLMs for Neural Architecture Synthesis  \nAnujaya Vijayakumar∗, Radu Timofte, Dmitry Ignatov Computer Vision Lab, CAIDAS & IFI, University of W¨urzburg, Germany  \narXiv :2607 . 11591v1 [ cs .CV] 13 Jul 2026  \nAbstract  \nWe introduce a MinHash-based similarity scheduling framework that constructs a progressive curriculum over neural architecture code for LLM-based neural architecture search (NAS) . Using 128-permutation MinHash signatures over normalised 7-gram source code shingles, we partition the reference pool into similarity bands and present them in increasing architectural heterogeneity, with the best LoRA adapter from each stage merged cumulatively into the backbone. We evaluate the framework on OlympicCoder- 7B within the LEMUR benchmark on CIFAR-10 image classification, generating N=15 candidate architectures per epoch across six progressive fine-tuning steps. The curriculum achieves 60% peak success rate at the high-similarity level without post-processing repair. A 2×2 ablation at the most diverse level curriculum versus base model, with versus without partial interface repair reveals that without repair the base model (47% peak SR) substantially outperforms the curriculum model (7% SR), while adding partial repair brings both to 53% SR. This pattern is consistent with merge-level weight drift progressively erasing evaluator-interface priors, and suggests that interface repair and curriculum scheduling target distinct failure modes. We further report a cross-dataset transfer observation on SVHN, where direct basemodel generation without curriculum warmup yields 27% peak SR at substantially lower accuracy (60 .5%) than the CIFAR-10 equivalent, consistent with the increased synthesis difficulty of the unq-family anchor  \n∗ Corresponding author: anujaya.vijayakumar@stud[mail.uni-wuerzburg.de](mail.uni-wuerzburg.de)  \narchitecture.  \n1 Introduction  \nLarge language models (LLMs) have emerged as a viable alternative to classical neural architecture search (NAS) by reframing architecture design as a program synthesis problem. Instead of searching over predefined graph templates, an LLM can directly generate executable Python implementations of complete neural network classes. Frameworks such as NNGPT [1] and LEMUR [2, 3] have demonstrated this feasibility: when conditioned on prior successful architectures, a code-generating LLM can iteratively propose candidate models compiled, trained, and scored by an external evaluator.  \nA question that has received little attention in LLM-based NAS is how the reference architectures should be ordered during fine-tuning. Existing pipelines treat the reference pool as an unordered collection. Neural architecture code differs not only in downstream accuracy but also in module composition, parameter routing, residual connectivity, and tensor-shape conventions. Exposing all such structural variations simultaneously may destabilise the internal code priors the LLM must build to generate evaluator-compatible designs reliably.  \nWe investigate whether ordering references by structural similarity during fine-tuning affects generation reliability. We introduce a MinHash-based similarity scheduling framework that partitions the reference pool into bands of increasing architectural heterogeneity and presents them progressively, with sequential LoRA adapter merging to preserve  \ncoding conventions across stages. A 2×2 ablation examines how progressive ordering interacts with post-generation interface repair, revealing that the two components address distinct failure modes inevaluator-compatible code generation.  \nThis paper makes three contributions:  \n1. A code-level structural similarity pipeline implementing MinHash signature computation over normalised 7-gram source code shingles, integrated into the LEMUR database, enabling subsecond pairwise similarity retrieval across 13,023 CIFAR-10 neural architecture implementations.  \n2. A MinHash-based curri","cbCaijI3elIt7Rbi","https://ap.wps.com/l/cbCaijI3elIt7Rbi","pdf",648714,4,1,15,"English","en",105,"# Introduction\n# Related Work\n## LLM-Based Neural Architecture Search\n## Curriculum Learning","[{\"question\":\"What problem does similarity-guided curriculum fine-tuning address in LLM-based NAS?\",\"answer\":\"It studies how ordering reference architectures during fine-tuning affects generation reliability, since neural architecture code varies in module composition, routing, residual connectivity, and tensor conventions.\"},{\"question\":\"How is the MinHash curriculum constructed?\",\"answer\":\"The method computes 128-permutation MinHash signatures over normalized 7-gram source-code shingles, partitions the reference pool into similarity bands, and presents them progressively by increasing architectural heterogeneity.\"},{\"question\":\"What do the ablation results suggest about curriculum warmup versus interface repair?\",\"answer\":\"At the most diverse level, without repair the base model substantially outperforms the curriculum model, while adding partial repair raises both to similar success rates, indicating the two components target distinct failure modes.\"}]",1784212724,38,{"code":4,"msg":31,"data":32},"ok",{"site_id":25,"language":24,"slug":33,"title":13,"keywords":34,"description":14,"schema_data":35,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":28},"similarity-guided-curriculum-fine-tuning-of-llms-for-neural-architecture-synthesis","",{"@graph":36,"@context":85},[37,53,68],{"@type":38,"itemListElement":39},"BreadcrumbList",[40,44,48,51],{"item":41,"name":42,"@type":43,"position":21},"https://docshare.wps.com","Home","ListItem",{"item":45,"name":46,"@type":43,"position":47},"https://docshare.wps.com/document/","Document",2,{"item":49,"name":12,"@type":43,"position":50},"https://docshare.wps.com/document/research-report/",3,{"item":52,"name":13,"@type":43,"position":20},"https://docshare.wps.com/document/similarity-guided-curriculum-fine-tuning-of-llms-for-neural-architecture-synthesis/86574/",{"url":52,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":24,"description":14,"dateModified":61,"datePublished":62,"encodingFormat":60,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":41,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-27","2026-07-16",true,{"@type":65,"interactionType":66,"userInteractionCount":20},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"What problem does similarity-guided curriculum fine-tuning address in LLM-based NAS?","Question",{"text":75,"@type":76},"It studies how ordering reference architectures during fine-tuning affects generation reliability, since neural architecture code varies in module composition, routing, residual connectivity, and tensor conventions.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How is the MinHash curriculum constructed?",{"text":80,"@type":76},"The method computes 128-permutation MinHash signatures over normalized 7-gram source-code shingles, partitions the reference pool into similarity bands, and presents them progressively by increasing architectural heterogeneity.",{"name":82,"@type":73,"acceptedAnswer":83},"What do the ablation results suggest about curriculum warmup versus interface repair?",{"text":84,"@type":76},"At the most diverse level, without repair the base model substantially outperforms the curriculum model, while adding partial repair raises both to similar success rates, indicating the two components target distinct failure modes.","https://schema.org",{"og:url":52,"og:type":87,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":89,"canonical":52},"index,follow",{"doc_id":7,"site_id":25},{"code":4,"msg":5,"data":92},[93,97,101,105,110,115,120,123,128,131,135],{"id":21,"doc_module":4,"doc_module_name":46,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":47,"doc_module":4,"doc_module_name":46,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":20,"doc_module":4,"doc_module_name":46,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":46,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":46,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":46,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":46,"category_name":12,"show_sort_weight":121,"slug":122},30,"research-report",{"id":124,"doc_module":4,"doc_module_name":46,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":46,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":132,"doc_module":4,"doc_module_name":46,"category_name":133,"show_sort_weight":132,"slug":134},10,"Lifestyle","lifestyle",{"id":136,"doc_module":4,"doc_module_name":46,"category_name":137,"show_sort_weight":106,"slug":138},19,"General","general"]