[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-84007-en":3,"doc-seo-84007-105":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},84007,7971461740909,"Levi","https://ap-avatar.wpscdn.com/davatar_155a257f0dc6eb9ab79c44ca47cae57d",8,"Research & Report","Controlling Tool Use with Heading-Specific Activation Steering","Tool-augmented large language models extend capabilities with external tools, yet often invoke them unnecessarily. The work studies whether tool-use decisions have stable internal representations that can be extracted and manipulated, given that tools are provided only in context and not encoded in model weights. Heading-anchored steering vectors enable bidirectional causal control of tool invocation across multiple open-source models and domains, strongly suppressing overuse when parametric reasoning suffices. Geometric analysis shows diffuse, bimodal alignment and low cross-tool feature overlap, suggesting non-parametric tool behavior.","Controlling Tool Use with Heading-Specific Activation Steering  \nYuqi Chen 1 Vincent Siu 1 Yang Liu 1 Dawn Song 2 Chenguang Wang 1  \narXiv :2607 .05790v 1 [ cs .AI ] 7 Jul 2026  \nAbstract  \nTool-augmented large language models extend their capabilities beyond parametric knowledge through external tools, but tend to invoke them unnecessarily. We investigate whether tool-use decisions have any stable internal representation that can be extracted and manipulated, a question that is non-trivial given that tools exist entirely in context at inference time and have no direct encoding in model weights. We show that steering vectors extracted from heading-anchors positions exert bidirectional causal control over tool-invocation behavior across five open-source models and three domains, suppressing unnecessary tool use most effectively in domains where parametric reasoning suffices. However, geometric analysis reveals that this causal effectiveness does not correspond to clean linear structure: toolinvocation steps exhibit diffuse, bimodal alignment with the suppression vector rather than the consistent negative alignment a linear encoding account would predict, and different tool types recruit largely distinct internal signatures with low cross-tool feature overlap. We hypothesize these geometric properties are indicative of thenon-parametric nature of tools, and distinguish tool-use steering vectors from those extracted for parametrically grounded concepts. The relationship between this geometric irregularity and the observed causal effectiveness remains an open question.  \n1. Introduction  \nTool-augmented LLMs have become a dominant paradigm for complex reasoning and real-world tasks (Yao et al., 2023 ; Levy et al., 2024 ; Schick et al., 2023) . External tools such as web search,  code execution, and user-interaction modules  \n1Department of Computer Science and Engineering, UC Santa Cruz 2Department of Computer Science, UC Berkeley. Correspondence to: Chenguang Wang \u003C[chenguangwang@ucsc.edu](chenguangwang@ucsc.edu) >.  \nPublished at the Second Workshop on Agents in the Wild: Safety, Security, and Beyond (AIWILD) at ICML 2026 . Copyright 2026 by the author(s) .  \nextend model capabilities beyond parametric knowledge, improving performance on tasks requiring up-to-date information, precise computation, or clarification of underspecified intent (Schick et al., 2023 ; Schneider, 2025) .  \nHowever, tool augmentation also introduces a persistent failure mode: tool overuse. Models may invoke tools even when internal reasoning would suffice, repeat calls without meaningful progress, or rely on tools in ways that increase latency and cost while providing limited benefit (Qian et al., 2025 ; Shen et al., 2024) . Unnecessary tool calls can expose the system to noisy retrieval, execution failures, and avoidable error propagation (Wang et al., 2025) . Existing mitigations either require expensive retraining (Qian et al., 2025 ; Shen et al., 2024) or operate at the output interface rather than on the internal activations underlying tool selection (Wang et al., 2025), leaving the latent decision to invoke a tool unaddressed.  \nWe ask whether tool-use decisions have any stable internal representation that can be extracted and manipulated. This is non-obvious: unlike concepts such as sentiment orfactuality that are encoded in model weights through training, tools exist entirely in context at inference time, and standard assumptions from the linear representation hypothesis (Park et al., 2024) and representation engineering (Zou et al., 2023) do not straightforwardly apply to concepts with no parametric grounding. Prior steering work has targeted concepts with clear parametric grounding (sentiment, refusal, truthfulness) (Panickssery et al., 2023 ; Siu et al., 2025a); we are among the first to apply it to a behavior that is explicitly non-parametric by construction. Following contrastive activation addition (Turner et al., 2023 ; Panickssery et al.,","cbCaij5mLvMDTnnb","https://ap.wps.com/l/cbCaij5mLvMDTnnb","pdf",1390248,3,1,15,"English","en",105,"# Abstract\n# Introduction\n## Motivation: Tool Overuse\n## Goal: Extractable Internal Representation for Tool Use\n## Method: Heading-Specific Steering Vectors\n## Results: Bidirectional Causal Control and Suppression","[{\"question\":\"What problem does the document address in tool-augmented LLMs?\",\"answer\":\"It addresses tool overuse: models may call external tools even when internal reasoning would suffice, or repeat calls with limited progress, increasing latency and cost.\"},{\"question\":\"How do heading-specific activation steering vectors affect tool invocation?\",\"answer\":\"Activation addition suppresses tool use below a baseline, while orthogonalization amplifies tool use above it, demonstrating bidirectional causal control over tool-selection behavior.\"},{\"question\":\"What does the geometric analysis suggest about how tool-use decisions are represented?\",\"answer\":\"Causal effectiveness does not follow clean linear structure: tool-invocation steps show diffuse, bimodal alignment with the suppression vector and different tool types activate largely distinct internal signatures with low cross-tool feature overlap.\"}]",1784191983,38,{"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},"controlling-tool-use-with-heading-specific-activation-steering","",{"@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/controlling-tool-use-with-heading-specific-activation-steering/84007/",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","2026-07-25","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 the document address in tool-augmented LLMs?","Question",{"text":75,"@type":76},"It addresses tool overuse: models may call external tools even when internal reasoning would suffice, or repeat calls with limited progress, increasing latency and cost.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How do heading-specific activation steering vectors affect tool invocation?",{"text":80,"@type":76},"Activation addition suppresses tool use below a baseline, while orthogonalization amplifies tool use above it, demonstrating bidirectional causal control over tool-selection behavior.",{"name":82,"@type":73,"acceptedAnswer":83},"What does the geometric analysis suggest about how tool-use decisions are represented?",{"text":84,"@type":76},"Causal effectiveness does not follow clean linear structure: tool-invocation steps show diffuse, bimodal alignment with the suppression vector and different tool types activate largely distinct internal signatures with low cross-tool feature overlap.","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":136,"doc_module":4,"doc_module_name":46,"category_name":137,"show_sort_weight":106,"slug":138},19,"General","general"]