[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-86105-en":3,"doc-seo-86105-105":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},86105,1374391974468,"Eden","https://ap-avatar.wpscdn.com/davatar_29158cc5080c5b710cf443261637dec0",8,"Research & Report","TabPFN Beyond Tabular Data Calibration and Accuracy on Multimodal Embeddings","Few-shot multimodal classification often uses a frozen pretrained encoder with a lightweight head (k-nearest neighbors, logistic regression, or a linear SVM). Though efficient, such heads frequently output poorly calibrated confidence scores, undermining reliability when calibration matters. This work evaluates TabPFN as a training-free, plug-and-play, zero-gradient classification head for frozen image, text, and audio encoders. Across 22,820 episodes over 14 datasets, 11 encoders, and three modalities, TabPFN attains the best mean rank on NLL and ECE among nine heads, reducing NLL by 48–62% and ECE by 2.1–5.3× versus baselines while preserving or improving accuracy under moderate-to-high shots and low-to-moderate feature dimensions.","TabPFN beyond Tabular Data: Calibration and Accuracy on Multimodal  \nEmbeddings  \nJingxiang Zhang 1 * , Lujia Zhong 1 * , Zijie Zhu 1 , Shuo Huang 1 , Yuang Xu 1 ,  \n1University of Southern California  \n[jzhang92@usc.edu](jzhang92@usc.edu), [lujiazho@usc.edu](lujiazho@usc.edu), [zzijiezhu@usc.edu](zzijiezhu@usc.edu), [shuohuan@usc.edu](shuohuan@usc.edu), [yuangxu@usc.edu](yuangxu@usc.edu)  \narXiv :2607 . 1 1007v 1 [ cs .LG] 13 Jul 2026  \nAbstract  \nFew-shot multimodal classification commonly attaches a lightweight head, such as k-nearest neighbors, logistic regression, or a linear SVM, to a frozen pretrained encoder. Although computationally efficient, these heads can produce poorly calibrated confidence scores, limiting their reliability in calibration-sensitive applications. We evaluate TabPFN asa plug-and-play, zero-gradient classification head for frozen image, text, and audio encoders. Across 22,820 evaluation episodes spanning 14 datasets, 11 encoders, and three modalities, TabPFN achieves the best mean rank among nine classification heads on both negative log-likelihood (NLL) and expected calibration error (ECE) . At a representative setting, it reduces NLL by 48–62% and ECE by 2.1–5.3 × relative to the average of the eight baselines while matching or exceeding their average accuracy. Its accuracy advantage is conditional, concentrating at moderate-to-high shot countsand low-to-moderate feature dimensions (k ≥ 50, d ≤ 32), and diminishing when labeled data are scarce, feature dimensions are high, or competing methods approach ceiling accuracy. In targeted backbone-adaptation experiments, replacing the trained linear head with TabPFN substantially improves calibration while preserving competitive accuracy. These results provide empirical guidance for using TabPFNas a training-free head in calibration-sensitive multimodal classification. To support transparency and reproducibility, we publicly release the source code, experiment configurations, and evaluation scripts in our GitHub repository: [https://github.com/Jingxiang-Zhang/tabpfn-multimodal](https://github.com/Jingxiang-Zhang/tabpfn-multimodal)embeddings.  \nIntroduction  \nAttaching a lightweight classification head to a frozen pretrained encoder is a widely adopted pipeline for few-shot multimodal classification. The workflow is to embed data with a pretrained vision, language, or audio encoder; fit ahead such as k-nearest neighbors (kNN), logistic regression, or linear SVM; and deploy (Sun et al. 2019; Radford et al. 2021; Br et al. 2024; Abubakar et al. 2026; Wang et al. 2023) . This gradient-free approach requires minimal computational overhead, enabling rapid deployment.  \nMost classification heads share a systematic flaw: their confidence scores are poorly calibrated. Standard benchmarks rarely expose this issue because they focus solely on  \n*These authors contributed equally.  \naccuracy. A classifier is well-calibrated if its stated confidence matches empirical accuracy; that is, a prediction labeled 70% confident should be correct 70% of the time. However, these classification heads often fail to satisfy this property (Niculescu-Mizil and Caruana 2005; Dormann 2020; Silva Filho et al. 2023) . In safety-critical settings such as medical imaging triage and clinical decision support, a miscalibrated head is unsafe regardless of its accuracy rank (Pias et al. 2025; Rezaeian, Bayrak, and Asan 2026) . Despite this risk, calibration across both frozen-encoder and fine-tuned adaptation pipelines remains underexplored.  \nThis lack of attention masks a fundamental technical challenge: standard classification methods struggle to naturally produce calibrated probabilities. In zero-training pipelines, classical heads such as kNN, logistic regression, and linear SVM are commonly used, but they are not explicitly designed to produce well-calibrated predictive uncertainty. Even methods that natively output probabilities, including logistic regression and kNN, can still produ","cbCaio9k9DSU2bEJ","https://ap.wps.com/l/cbCaio9k9DSU2bEJ","pdf",7721579,3,1,19,"English","en",105,"# Abstract\n# Introduction\n## Calibration problem in lightweight heads\n## Limits of zero-training and probabilistic outputs\n## Post-hoc calibration trade-offs\n## Motivation for TabPFN in multimodal settings","[{\"question\":\"What issue does the paper identify with common few-shot multimodal classification heads?\",\"answer\":\"Lightweight heads attached to frozen encoders often produce confidence scores that are not well calibrated, meaning stated confidence does’t match empirical accuracy. This harms reliability in calibration-sensitive scenarios even when accuracy rankings look good.\"},{\"question\":\"How does TabPFN differ from traditional lightweight heads?\",\"answer\":\"TabPFN is used as a training-free, plug-and-play classification head with a zero-gradient setup. It is evaluated directly on frozen image, text, and audio encoders rather than requiring head training or extra calibration optimization.\"},{\"question\":\"Under what conditions does TabPFN’s accuracy and calibration advantage hold best?\",\"answer\":\"The advantage is strongest with moderate-to-high shot counts and low-to-moderate feature dimensions (k ≥ 50, d ≤ 32). It diminishes when labeled data are scarce, feature dimensions are high, or competing methods reach similar ceiling accuracy.\"}]",1784208552,48,{"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},"tabpfn-beyond-tabular-data-calibration-and-accuracy-on-multimodal-embeddings","",{"@graph":36,"@context":85},[37,53,68],{"@type":38,"itemListElement":39},"BreadcrumbList",[40,44,48,50],{"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":20},"https://docshare.wps.com/document/research-report/",{"item":51,"name":13,"@type":43,"position":52},"https://docshare.wps.com/document/tabpfn-beyond-tabular-data-calibration-and-accuracy-on-multimodal-embeddings/86105/",4,{"url":51,"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","2026-07-24","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 issue does the paper identify with common few-shot multimodal classification heads?","Question",{"text":75,"@type":76},"Lightweight heads attached to frozen encoders often produce confidence scores that are not well calibrated, meaning stated confidence does’t match empirical accuracy. This harms reliability in calibration-sensitive scenarios even when accuracy rankings look good.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How does TabPFN differ from traditional lightweight heads?",{"text":80,"@type":76},"TabPFN is used as a training-free, plug-and-play classification head with a zero-gradient setup. It is evaluated directly on frozen image, text, and audio encoders rather than requiring head training or extra calibration optimization.",{"name":82,"@type":73,"acceptedAnswer":83},"Under what conditions does TabPFN’s accuracy and calibration advantage hold best?",{"text":84,"@type":76},"The advantage is strongest with moderate-to-high shot counts and low-to-moderate feature dimensions (k ≥ 50, d ≤ 32). It diminishes when labeled data are scarce, feature dimensions are high, or competing methods reach similar ceiling accuracy.","https://schema.org",{"og:url":51,"og:type":87,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":89,"canonical":51},"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":52,"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":22,"doc_module":4,"doc_module_name":46,"category_name":136,"show_sort_weight":106,"slug":137},"General","general"]