[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-201813-105":59,"doc-detail-201813-en":131},{"code":4,"msg":5,"data":6},0,"success",[7,13,18,23,28,33,38,43,48,51,55],{"id":8,"doc_module":4,"doc_module_name":9,"category_name":10,"show_sort_weight":11,"slug":12},1,"Document","Story & Novel",90,"story-novel",{"id":14,"doc_module":4,"doc_module_name":9,"category_name":15,"show_sort_weight":16,"slug":17},2,"Literature",80,"literature",{"id":19,"doc_module":4,"doc_module_name":9,"category_name":20,"show_sort_weight":21,"slug":22},4,"Exam",70,"exam",{"id":24,"doc_module":4,"doc_module_name":9,"category_name":25,"show_sort_weight":26,"slug":27},5,"Comic",60,"comic",{"id":29,"doc_module":4,"doc_module_name":9,"category_name":30,"show_sort_weight":31,"slug":32},6,"Technology",50,"technology",{"id":34,"doc_module":4,"doc_module_name":9,"category_name":35,"show_sort_weight":36,"slug":37},7,"Healthcare",40,"healthcare",{"id":39,"doc_module":4,"doc_module_name":9,"category_name":40,"show_sort_weight":41,"slug":42},8,"Research & Report",30,"research-report",{"id":44,"doc_module":4,"doc_module_name":9,"category_name":45,"show_sort_weight":46,"slug":47},9,"Religion & Spirituality",20,"religion-spirituality",{"id":46,"doc_module":4,"doc_module_name":9,"category_name":49,"show_sort_weight":46,"slug":50},"World Cup","world-cup",{"id":52,"doc_module":4,"doc_module_name":9,"category_name":53,"show_sort_weight":52,"slug":54},10,"Lifestyle","lifestyle",{"id":56,"doc_module":4,"doc_module_name":9,"category_name":57,"show_sort_weight":24,"slug":58},19,"General","general",{"code":4,"msg":60,"data":61},"ok",{"site_id":62,"language":63,"slug":64,"title":65,"keywords":66,"description":67,"schema_data":68,"social_meta":124,"head_meta":126,"extra_data":128,"updated_unix":130},105,"en","provable-robustness-to-spurious-correlations-via-invariant-data-for-robust-finetuning","Provable Robustness to Spurious Correlations via Invariant Data - For Robust Finetuning","","Scaling models on simple predictive objectives often fails under distribution shifts because spurious correlations in training data lead to degraded out-of-distribution generalization. Domain Generalization techniques seek invariant representations but can be computationally expensive and may underperform Empirical Risk Minimization. This work introduces Geometric Robustness via Invariant Training (GRIT), a data-centric fine-tuning method that enforces geometric invariance using noisy invariant pairs, with a finite-sample analysis and theoretical robustness guarantees.",{"@graph":69,"@context":123},[70,84,106],{"@type":71,"itemListElement":72},"BreadcrumbList",[73,77,79,82],{"item":74,"name":75,"@type":76,"position":8},"https://docshare.wps.com","Home","ListItem",{"item":78,"name":9,"@type":76,"position":14},"https://docshare.wps.com/document/",{"item":80,"name":40,"@type":76,"position":81},"https://docshare.wps.com/document/research-report/",3,{"item":83,"name":65,"@type":76,"position":19},"https://docshare.wps.com/document/provable-robustness-to-spurious-correlations-via-invariant-data-for-robust-finetuning/201813/",{"url":83,"name":65,"@type":85,"image":86,"author":91,"headline":65,"publisher":94,"fileFormat":97,"inLanguage":63,"description":67,"dateModified":98,"datePublished":99,"encodingFormat":97,"isAccessibleForFree":100,"interactionStatistic":101},"DigitalDocument",{"url":87,"@type":88,"width":89,"height":90},"https://docshare.wps.com/thumbnails/provable-robustness-to-spurious-correlations-via-invariant-data-for-robust-finetuning/201813.png","ImageObject",300,407,{"name":92,"@type":93},"Alex Sinclair","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-11-04","2026-09-04",true,{"@type":102,"interactionType":103,"userInteractionCount":105},"InteractionCounter",{"@type":104},"ViewAction",18,{"@type":107,"mainEntity":108},"FAQPage",[109,115,119],{"name":110,"@type":111,"acceptedAnswer":112},"What problem does the paper address in finetuning foundation models?","Question",{"text":113,"@type":114},"It addresses out-of-distribution failures caused by spurious correlations learned during training, where standard ERM on observational data is insufficient for robust generalization.","Answer",{"name":116,"@type":111,"acceptedAnswer":117},"How does GRIT enforce robustness without explicit causal modeling?",{"text":118,"@type":114},"GRIT uses geometric invariance constraints during fine-tuning based on a small set of noisy invariant pairs that implicitly encode invariance properties.",{"name":120,"@type":111,"acceptedAnswer":121},"What theoretical guarantee does the paper provide for GRIT?",{"text":122,"@type":114},"The paper provides a finite-sample analysis showing GRIT’s robust generalization improves at a rate of O(1/√k) with the number of invariant pairs k.","https://schema.org",{"og:url":83,"og:type":125,"og:title":65,"og:site_name":95,"og:description":67},"article",{"robots":127,"canonical":83},"index,follow",{"doc_id":129,"site_id":62},201813,1788538967,{"code":4,"msg":5,"data":132},{"doc_id":129,"user_id":133,"nickname":92,"user_avatar":134,"doc_module":4,"category_id":39,"category_name":40,"doc_title":65,"doc_description":67,"doc_content":135,"file_id":136,"file_url":137,"file_type":138,"file_size":139,"view_count":105,"is_deleted":4,"is_public":8,"is_downloadable":8,"audit_status":8,"page_count":140,"language":141,"language_code":63,"site_id":62,"html_lang":63,"table_of_contents":142,"faqs":143,"seo_title":144,"seo_description":67,"update_tm":130,"read_time":145,"preview_page_count":4},1099523882182,"https://ap-avatar.wpscdn.com/davatar_6f874abed73319feea01a86fa6f0fab8","Provable Robustness to Spurious Correlations via Invariant Data  \nFor Robust Finetuning  \nRuqi Bai 1 , Yao Ji2 , Mingyu Kim 1 , Easton Currie 1 , Zeyu Zhou 1 , David I. Inouye 1  \n1 Purdue University, 2 Georgia Institute of Technology  \nAbstract  \nScaling models on simple predictive objectives is often insufficient to overcome spurious correlations that degrade out-of-distribution generalization. While Domain Generalization (DG) methods aim to learn invariant representations–often based on causality principles–they can be computationally expensive and underperform simple Empirical Risk Minimization (ERM) . We propose a data-centric alternative: Geometric Robustness via Invariant Training (GRIT) . Instead of explicit causal modeling, GRIT enforces a geometric constraint during finetuning based on a small set of noisy invariant pairs, which implicitly encode an invariance property. We provide the first finitesample analysis of this setting, showing that our framework generalizes latent linear causal models. We prove GRIT achieves robust generalization that scales at a rate of O(1/ √k)  with the number of pairs k, offering a scalable alternative to ERM or explicit causal modeling for out-of-distribution robustness.  \n1 Introduction  \nAdapting pre-trained foundation models via finetuning has become ubiquitous. While scaling these feature spaces captures rich semantic information, they still suffer from spurious correlations—misleading patterns that hold in the training data but fail in new environments [Sagawa et al. , 2019] . Addressing these failures reveals a critical limitation: simply scaling standard ERM on observational data is often insufficient to induce true out-of-distribution robustness.  \nStandard approaches fall under Robust Fine-Tuning or Domain Generalization (DG) . Methods like WiSEFT [Wortsman et al. , 2022] require expensive retraining. Many DG approaches aim to learn invariant representations and are explicitly inspired by causality;  \nfor example, Invariant Risk Minimization (IRM) [Arjovsky et al. , 2019] builds on invariant causal prediction (ICP) [Peters et al. , 2016], and MatchDG [Mahajan et al. , 2021] explicitly assumes a causal model. However, these methods are often unstable, underperform simple ERM [Gulrajani and Lopez-Paz, 2021], and rely on strong assumptions. While a few prior works explore out-of-distribution robustness by leveraging paired data [Mahajan et al. , 2021 , Nguyen et al. , 2021 , Surner et al. , 2025], they lack theoretical guarantees addressing noisy and finite data pairs. (See Section A for expanded related works.)  \nWe shift the focus from explicit causal modeling by asking: Can we provide theoretic guarantees for the robustness of a predictor on a fixed feature space using a small number of noisy invariant pairs—samples that share the same semantic label but differ in spurious attributes (e.g., counterfactuals)?  \nWe operate under the geometric intuition that spurious correlations reside in specific subspaces of the feature embedding. Importantly, our framework generalizes latent linear causal models, demonstrating the connection to causal models. While grounded in linear analysis, our framework naturally extends to kernelized non-linear predictors. Crucially, our framework explicitly models pair noise and provides a finitesample analysis with respect to the number of pairs. Our contributions are threefold:  \n• We propose GRIT, a fine-tuning approach that enforces geometric invariance via noisy pairs.  \n• We prove that GRIT robustness improves at a rate of O(1/ √k  ) with the number of pairs k.  \n• We demonstrate that GRIT outperforms baselines on real-world datasets (Waterbirds, ColoredMNIST) using frozen CLIP embeddings.  \n2 Problem Setup  \nWe analyze the robustness of a model trained on a frozen representation feature space z = ϕ (x) ∈ Rd.  \nGeometric Structure and Distribution Shifts.  \nWe assume the representation space possesses a geometric structure where predictive signals","cbCaid7cCp67bJen","https://ap.wps.com/l/cbCaid7cCp67bJen","pdf",1592970,31,"English","# Introduction\n## Problem Setup\n## Geometric Structure and Distribution Shifts\n## Observation Model: Noisy Invariant Pairs","[{\"question\":\"What problem does the paper address in finetuning foundation models?\",\"answer\":\"It addresses out-of-distribution failures caused by spurious correlations learned during training, where standard ERM on observational data is insufficient for robust generalization.\"},{\"question\":\"How does GRIT enforce robustness without explicit causal modeling?\",\"answer\":\"GRIT uses geometric invariance constraints during fine-tuning based on a small set of noisy invariant pairs that implicitly encode invariance properties.\"},{\"question\":\"What theoretical guarantee does the paper provide for GRIT?\",\"answer\":\"The paper provides a finite-sample analysis showing GRIT’s robust generalization improves at a rate of O(1/√k) with the number of invariant pairs k.\"}]","Provable Robustness to Spurious Correlations via Invariant Data - For Robust Finetuning | PDF",78]