[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-117618-en":3,"doc-seo-117618-105":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":4,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":27,"seo_description":14,"update_tm":28,"read_time":29},117618,962075114101,"Seraphina","https://ap-avatar.wpscdn.com/avatar/e000253a75eb197efd?x-image-process=image/resize,m_fixed,w_180,h_180&k=1780044092746381165",8,"Research & Report","Provably Reliable Machine Learning Systems - Dissertation","Machine learning systems built on deep neural networks (DNNs) underpin safety-critical and compound AI workflows, yet automated, provably reliable verification remains limited. Neural network verification is NP-hard and expensive to rerun when models change in iterative deployment cycles. In compound systems, large language models (LLMs) can also emit outputs that break syntactic or semantic constraints, triggering cascading failures. This dissertation presents a unified framework combining incremental DNN verification and constrained LLM decoding to achieve general, precise, and scalable reliability.","© 2025 Shubham Ugare  \nPROVABLY RELIABLE MACHINE LEARNING SYSTEMS  \nBY  \nSHUBHAM UGARE  \nDISSERTATION  \nSubmitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Computer Science in the Graduate College of the University of Illinois Urbana-Champaign, 2025  \nUrbana, Illinois  \nDoctoral Committee:  \nAssociate Professor Sasa Misailovic, Chair  \nAssistant Professor Gagandeep Singh  \nAssociate Professor Lingming Zhang  \nProfessor Swarat Chaudhuri, UT Austin  \nABSTRACT  \nMachine learning systems, which primarily use deep neural networks (DNNs), serve as critical components in safety-critical applications and compound AI systems. Despite their ubiquity, automated formal reasoning about their reliability has lagged significantly. Neural network verification is NP-hard, requiring expensive end-to-end recomputation whenever networks are modified during iterative deployment cycles. Simultaneously, Large Language Models (LLMs) in compound systems frequently generate outputs that violate syntactic and semantic specifications, leading to cascading failures in automated workflows. Thus, developing reliability techniques for machine learning systems that are simultaneously general, precise, and scalable remains a challenging task. To address these challenges, this dissertation develops a comprehensive framework for provably reliable machine learning systems by establishing incremental verification techniques for DNNs and constrained decoding methods specific to LLMs.  \nTo obtain verification techniques with the desired scalability for pre-deployment validation, Part I proposes incremental verification techniques that fundamentally change the verification workflow from repeatedly running expensive end-to-end verification to efficiently reusing and adapting existing proofs. Chapter 2 presents FANC, which introduces proof transfer for sound and incomplete verification methods through template-based abstractions. Chapter 3 extends incremental techniques to sound and complete verification through Ivan, which leverages specification trees to guide branch-and-bound verification while maintaining completeness guarantees. Chapter 4 addresses the most scalable verification paradigm through IRS, which develops the first incremental approach for randomized smoothing.  \nTo obtain reliable structured output generation during deployment, Part II proposes constrained decoding frameworks that ensure LLM outputs adhere to formal specifications. Chapter 5 presents SynCode, which addresses syntactic correctness through a parsingbased approach that constrains LLM decoding using context-free grammars. Chapter 6 extends beyond syntax through IterGen, which provides grammar symbol-level navigation enabling programmatic checking and correction of semantic violations through bidirectional generation.  \nACKNOWLEDGMENTS  \nMy PhD years have been a period of immense growth both personally and professionally. Sasa and Gagandeep have been instrumental advisers. Sasa’s focus on choosing impactful problems and communicating research clearly has shaped how I approach my work. Gagandeep’s determination in tackling hard problems and his calmness in adversity taught me resilience when some research problem felt impossible. I still think back to the all-nighterin the office for my first submission deadline and the countless PLDI presentation rehearsalsin a hotel room the night before the talk. I was a bit nervous during those times, but their support turned them into memories I now look back on very fondly. I am deeply indebted to both of them for the patience, guidance, and confidence they invested in me over these years.  \nI am grateful to both Lingming and Swarat for participating in my PhD committee. The consistency of high-quality research output that they have produced has been an inspiration for me. Their thoughtful feedback and encouragement helped me refine my work and think more deeply about its broader impact.  \nI owe a great deal to","cbCaibh2XMc9Y8dY","https://ap.wps.com/l/cbCaibh2XMc9Y8dY","pdf",6980705,1,195,"English","en",105,"# Abstract\n# Part I: Incremental Verification for DNNs\n## Chapter 2: FANC\n## Chapter 3: Ivan\n## Chapter 4: IRS\n# Part II: Constrained Decoding for LLMs\n## Chapter 5: SynCode\n## Chapter 6: IterGen\n# Acknowledgments","[{\"question\":\"Why is provable reliability difficult for machine learning systems in practice?\",\"answer\":\"Neural network verification is NP-hard, so iterative model changes require costly end-to-end recomputation. Additionally, LLMs in compound workflows may violate syntactic or semantic specifications, causing cascading failures.\"},{\"question\":\"How does the dissertation improve scalability for pre-deployment validation?\",\"answer\":\"It introduces incremental verification methods that reuse and adapt existing proofs instead of repeatedly running expensive end-to-end verification, including approaches for sound/incomplete, sound/complete, and randomized smoothing.\"},{\"question\":\"How does the dissertation ensure reliable structured output generation from LLMs during deployment?\",\"answer\":\"It proposes constrained decoding frameworks that enforce formal specifications, using grammar-based syntactic constraints (SynCode) and grammar symbol-level navigation for semantic checking and correction (IterGen).\"}]","Provably Reliable Machine Learning Systems - Dissertation | PDF",1785677308,491,{"code":4,"msg":31,"data":32},"ok",{"site_id":24,"language":23,"slug":33,"title":13,"keywords":34,"description":14,"schema_data":35,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":28},"provably-reliable-machine-learning-systems-dissertation","",{"@graph":36,"@context":85},[37,54,68],{"@type":38,"itemListElement":39},"BreadcrumbList",[40,44,48,51],{"item":41,"name":42,"@type":43,"position":20},"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":53},"https://docshare.wps.com/document/provably-reliable-machine-learning-systems-dissertation/117618/",4,{"url":52,"name":13,"@type":55,"author":56,"headline":13,"publisher":58,"fileFormat":61,"inLanguage":23,"description":14,"dateModified":62,"datePublished":62,"encodingFormat":61,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":57},"Person",{"url":41,"name":59,"@type":60},"DocShare","Organization","application/pdf","2026-08-02",true,{"@type":65,"interactionType":66,"userInteractionCount":4},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"Why is provable reliability difficult for machine learning systems in practice?","Question",{"text":75,"@type":76},"Neural network verification is NP-hard, so iterative model changes require costly end-to-end recomputation. Additionally, LLMs in compound workflows may violate syntactic or semantic specifications, causing cascading failures.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How does the dissertation improve scalability for pre-deployment validation?",{"text":80,"@type":76},"It introduces incremental verification methods that reuse and adapt existing proofs instead of repeatedly running expensive end-to-end verification, including approaches for sound/incomplete, sound/complete, and randomized smoothing.",{"name":82,"@type":73,"acceptedAnswer":83},"How does the dissertation ensure reliable structured output generation from LLMs during deployment?",{"text":84,"@type":76},"It proposes constrained decoding frameworks that enforce formal specifications, using grammar-based syntactic constraints (SynCode) and grammar symbol-level navigation for semantic checking and correction (IterGen).","https://schema.org",{"og:url":52,"og:type":87,"og:title":13,"og:site_name":59,"og:description":14},"article",{"robots":89,"canonical":52},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":92},[93,97,101,105,110,115,120,123,128,131,135],{"id":20,"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":53,"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"]