[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-83217-en":3,"doc-seo-83217-105":29,"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":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":13,"seo_description":14,"update_tm":27,"read_time":28},83217,962075114765,"Quinn","https://ap-avatar.wpscdn.com/davatar_a8503ba1806abce46bf441b54a3ca4cd",8,"Research & Report","Naming the Concepts Classifiers Rely On Language Anchored Decomposition for Faithful Explanation","Deep neural networks are increasingly used in high-stakes visual settings where explanations must be both interpretable and faithful. Existing concept methods either discover unnamed factors faithful to model behavior or attach human-readable names by retraining. Language-Anchored Decomposition (LAD) is a post-hoc framework producing named concepts without modifying the model, localizing language-proposed vocabularies via CLIP similarity and learning a fixed reconstruction basis to match frozen activations while preserving decision relevance across benchmarks.","Naming the Concepts Classifiers Rely On: Language-Anchored Decomposition  \nfor Faithful Explanation  \nAhsan Habib Akash 1 Dipkamal Bhusal2 Stacey Jones3  \nDonald A. Adjeroh 1 Binod Bhattarai4 ,5 ,6 Prashnna Kumar Gyawali 1 *  \n1West Virginia University, USA 2Rochester Institute of Technology, USA 3 O Analytics  \n4University of Aberdeen, UK 5Fogsphere (Redev.AI Ltd, UK)  \n6 University College London, UK  \narXiv :2607 .07264v 1 [ cs .CV] 8 Jul 2026  \nAbstract  \nDeep neural networks are widely deployed in high-stakes visual applications where interpretability is critical, yet existing explanations face a trade-off: post-hoc concept methods recover factors that are faithful to a model’s behavior but unnamed, while naming and by-design methods attach human-readable concepts only by retraining or altering the classifier. We propose Language-Anchored Decomposition (LAD), a post-hoc framework that delivers concepts which are simultaneously named, faithful, and obtained without modifying the model. For each class, a large language model proposes a concept vocabulary that CLIP-based similarity maps localize across image regions. Inverting standard non-negative matrix factorization, LAD fixes these language-grounded maps as the coefficient matrix and learns only a concept basis that reconstructs the frozen encoder’s activations, so naming becomes a structural constraint and the model’s own feature geometry determines which concepts are retained. Removing this anchor preserves accuracy but collapses attribution faithfulness. Across natural-image, scene, and medical-imaging benchmarks, LAD produces spatially precise explanations that are decision-relevant under both concept insertion and deletion, while uniquely providing stable, human-interpretable concept names. Code is available at: [https://github. com/machine](https://github. com/machine)intelligence-lab-wvu/LAD.  \n1. Introduction  \nInterpreting the decision mechanisms of deep visual models is crucial for diagnosing failures, mitigating biases, and ensuring responsible deployment in high-stakes applications. Post-hoc attribution methods such as Grad-CAM [20] and  \n* Corresponding author: [prashnna.gyawali@mail.wvu.edu](prashnna.gyawali@mail.wvu.edu)  \nFigure 1 . Qualitative illustration of LAD. Each image is decomposed into spatially localized, named concepts, including partlevel object concepts (e.g.,“Pointy Ears,”“Green Eyes”) and contextual scene concepts (e.g.,“Crowd in the Background,” “Player Dribbling”) . Because each concept is anchored to language before decomposition, the explanations are both human-interpretable and faithful to the frozen classifier’s predictions.  \nIntegrated Gradients [24] visualize the input regions most influential to a prediction, but they operate at the pixel level and emphasize low-level activations rather than the highlevel cues humans reason with, offering limited semantic insight into a model’s decision process [14] . Conceptbased explanations address this by attributing predictions to human-interpretable units such as object parts and visual attributes [6, 7, 9] .  \nExisting concept-based methods, however, fall into two families that each pay a price. The first is post-hoc concept discovery. Building on TCAV [9], unsupervised methods such as ACE [7], ICE [28], and CRAFT [6] extract concepts via clustering or Non-Negative Matrix Factorization (NMF), and FACE [3] further aligns the decomposition with classifier logits to improve faithfulness. These methods are faithful to the model, but the recovered factors are unnamed: the same factor index can correspond to different visual elements across images, so a human must label each  \nfactor after the fact and consistency is not guaranteed. The second family attaches names directly. Concept Bottleneck Models [10], their label-free variants [16], and spatial extensions such as Show and Tell [2] predict named concepts before classification, while neuron-naming methods such as LaViSE [27] train an auxiliary","cbCaimxQ8tu2ZTss","https://ap.wps.com/l/cbCaimxQ8tu2ZTss","pdf",24958926,1,20,"English","en",105,"# Abstract\n# Introduction\n## Problem: Faithful interpretability in deep visual models\n## Existing concept methods and limitations\n## Proposed approach: Language-Anchored Decomposition (LAD)","[{\"question\":\"What limitation do existing concept-based explanation methods share?\",\"answer\":\"They either provide faithful factors without stable human-readable names, or provide names by altering the model and/or without verifying causal responsibility. As a result, no prior method delivers concepts that are simultaneously named, faithful, and obtained without modifying the deployed classifier.\"},{\"question\":\"How does LAD generate named, faithful concepts without retraining the classifier?\",\"answer\":\"For each class, a large language model proposes a concept vocabulary; CLIP similarity localizes each concept across image regions. LAD then fixes these language-grounded maps as coefficients and learns only a concept basis that reconstructs the frozen encoder’s activations, making naming a structural constraint.\"},{\"question\":\"Why is the “language anchor” important for attribution faithfulness?\",\"answer\":\"If the fixed language-grounded coefficients are replaced with learned ones, accuracy can remain but attribution faithfulness collapses. The anchor steers discovered basis directions toward classifier-meaningful structure and suppresses concepts inconsistent with the encoder’s evidence, reducing biases in CLIP-style setups.\"}]",1784186012,50,{"code":4,"msg":30,"data":31},"ok",{"site_id":24,"language":23,"slug":32,"title":13,"keywords":33,"description":14,"schema_data":34,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":27},"naming-the-concepts-classifiers-rely-on-language-anchored-decomposition-for-faithful-explanation","",{"@graph":35,"@context":85},[36,53,68],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,47,50],{"item":40,"name":41,"@type":42,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":44,"name":45,"@type":42,"position":46},"https://docshare.wps.com/document/","Document",2,{"item":48,"name":12,"@type":42,"position":49},"https://docshare.wps.com/document/research-report/",3,{"item":51,"name":13,"@type":42,"position":52},"https://docshare.wps.com/document/naming-the-concepts-classifiers-rely-on-language-anchored-decomposition-for-faithful-explanation/83217/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":23,"description":14,"dateModified":61,"datePublished":62,"encodingFormat":60,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":40,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-17","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 limitation do existing concept-based explanation methods share?","Question",{"text":75,"@type":76},"They either provide faithful factors without stable human-readable names, or provide names by altering the model and/or without verifying causal responsibility. As a result, no prior method delivers concepts that are simultaneously named, faithful, and obtained without modifying the deployed classifier.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How does LAD generate named, faithful concepts without retraining the classifier?",{"text":80,"@type":76},"For each class, a large language model proposes a concept vocabulary; CLIP similarity localizes each concept across image regions. LAD then fixes these language-grounded maps as coefficients and learns only a concept basis that reconstructs the frozen encoder’s activations, making naming a structural constraint.",{"name":82,"@type":73,"acceptedAnswer":83},"Why is the “language anchor” important for attribution faithfulness?",{"text":84,"@type":76},"If the fixed language-grounded coefficients are replaced with learned ones, accuracy can remain but attribution faithfulness collapses. The anchor steers discovered basis directions toward classifier-meaningful structure and suppresses concepts inconsistent with the encoder’s evidence, reducing biases in CLIP-style setups.","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":24},{"code":4,"msg":5,"data":92},[93,97,101,105,110,114,119,122,126,129,133],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":45,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":45,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":45,"category_name":112,"show_sort_weight":28,"slug":113},6,"Technology","technology",{"id":115,"doc_module":4,"doc_module_name":45,"category_name":116,"show_sort_weight":117,"slug":118},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":120,"slug":121},30,"research-report",{"id":123,"doc_module":4,"doc_module_name":45,"category_name":124,"show_sort_weight":21,"slug":125},9,"Religion & Spirituality","religion-spirituality",{"id":21,"doc_module":4,"doc_module_name":45,"category_name":127,"show_sort_weight":21,"slug":128},"World Cup","world-cup",{"id":130,"doc_module":4,"doc_module_name":45,"category_name":131,"show_sort_weight":130,"slug":132},10,"Lifestyle","lifestyle",{"id":134,"doc_module":4,"doc_module_name":45,"category_name":135,"show_sort_weight":106,"slug":136},19,"General","general"]