[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-85421-en":3,"doc-seo-85421-105":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},85421,1099513958762,"Logic","https://ap-avatar.wpscdn.com/avatar/1000023916a998db790?x-image-process=image/resize,m_fixed,w_180,h_180&k=1782109480056885918",8,"Research & Report","REAL: Reading Out Transformer Activations for Precise Localization in Language Model Steering","Inference-time steering alters an LLM’s responses without changing its parameters, but identifying the internal modules that govern the target behavior remains difficult. Existing methods often use simplistic cues or heuristics, causing suboptimal or unintended steering effects. REAL introduces a framework that trains a vector-quantized autoencoder on module hidden activations to separate behavior-relevant and behavior-irrelevant subspaces with a shared codebook and relevance scoring for module selection and steering strength. Experiments across eight LLMs show significant gains on truthfulness, QA with knowledge conflicts, and alignment tasks.","arXiv :2506 .08359v2 [ cs .CL] 1 Oct 2025  \nREAL: Reading Out Transformer Activations for Precise Localization in Language Model Steering  \nLi-Ming Zhan Bo Liu Yujie Feng Chengqiang Xie  \nJiannong Cao Xiao-Ming Wu  \nDepartment of Data Science and Artificial Intelligence  \nThe Hong Kong Polytechnic University  \nHong Kong S.A.R.  \n{lmzhan.zhan, bokelvin.liu, [yujie.feng}@connect.polyu.hk](yujie.feng}@connect.polyu.hk), [cqx.xie@gmail.com](cqx.xie@gmail.com)  \n[csjcao@comp.polyu.edu.hk](csjcao@comp.polyu.edu.hk)  \n[xiao-ming.wu@polyu.edu.hk](xiao-ming.wu@polyu.edu.hk)  \nAbstract  \nInference-time steering aims to alter an LLM’s responses without changing its parameters. A key challenge lies in selecting internal modules that most strongly govern the target behavior; existing approaches often rely on simplistic cues or ad hoc heuristics, leading to suboptimal or unintended effects. In this work, we introduce REAL, a novel framework for identifying behavior-relevant modules (heads or layers) in Transformers. For each module, we train a vector-quantized autoencoder (VQ-AE) on its hidden activations, partitioning the latent space into behaviorrelevant and behavior-irrelevant subspaces via a shared, learnable codebook. We quantify each module’s behavioral relevance by evaluating how effectively the VQ-AE encodings distinguish between behavior-aligned and behavior-violating responses using a binary classification metric. This relevance score informs both module selection and steering strength. We evaluate REAL across eight LLMs from two model families (LLAMA and QWEN) and nine datasets spanning truthfulness enhancement, open-domain question answering under knowledge conflicts, and general alignment tasks. REAL enables more effective inference-time interventions, yielding significant improvements on these steering tasks. Notably, it achieves an average relative improvement of 20%(up to 81.5%) over the seminal ITI method [1] on truthfulness steering. Moreover, the modules selected by our method exhibit strong zero-shot generalization in cross-domain truthfulness-steering scenarios.  \n1 Introduction  \nPost-hoc control of large language models (LLMs) has emerged as a critical area for advancing model alignment and safety. Activation (representation) steering [2, 3] achieves behavioral modulation by injecting an additive “steering vector” into intermediate activations during inference, without modifying any model parameters. In comparison to parameter-efficient fine-tuning [4] and neural knowledge editing [5], activation steering is minimally intrusive, computationally efficient, and often exhibits greater robustness to out-of-distribution scenarios [1, 6, 7, 8] . The steering process generally consists of two main steps: (I) identifying internal modules (e.g., attention heads) associated with the target behavior, and (II) constructing steering vectors to modify the corresponding activations. While substantial research has focused on developing more effective steering vectors [9, 10, 11], the questions of optimal intervention location (layer/head selection) and intervention strength (intensity scheduling) remain underexplored. Current steering methods often depend on simplistic linear probing [1, 11], ad-hoc empirical heuristics [12], or computationally intensive cross-validation [7,  \nPreprint. Under review.  \n3] for module selection. Internal components of LLMs, such as attention heads, are central to text generation and support diverse functions, including induction [13, 14] and long-range factual retrieval [15] . However, these polysemantic functions are often encoded in highly entangled hidden activations [16, 17, 11], making it difficult to isolate or detect behavior-specific features using simple linear probes.  \nTo highlight the critical importance of module selection and the limitations of linear-probe-based methods, we present a comparative analysis in Fig. 1 and Fig. 2. Specifically, Fig. 2 displays the top 48 attention heads iden","cbCaib1WVT0domvD","https://ap.wps.com/l/cbCaib1WVT0domvD","pdf",6264655,1,120,"English","en",105,"# Abstract\n# Introduction\n## Activation (Representation) Steering\n## Module Selection vs Linear Probes\n## REAL Framework Overview","[{\"question\":\"What problem does REAL target in inference-time steering?\",\"answer\":\"REAL targets the challenge of selecting internal transformer modules (heads or layers) that most strongly govern a desired behavior, since existing cue-based or heuristic methods can yield weak or unintended effects.\"},{\"question\":\"How does REAL identify behavior-relevant modules?\",\"answer\":\"For each module, REAL trains a vector-quantized autoencoder on its hidden activations, partitioning the latent space into behavior-relevant and behavior-irrelevant subspaces using a shared learnable codebook, then computes a relevance score via binary classification between behavior-aligned and behavior-violating responses.\"},{\"question\":\"What improvements does REAL achieve compared with ITI?\",\"answer\":\"On truthfulness steering, REAL reports an average relative improvement of 20% with results up to 81.5% over the seminal ITI method, and it also demonstrates strong zero-shot generalization in cross-domain truthfulness-steering scenarios.\"}]",1784203328,302,{"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},"real-reading-out-transformer-activations-for-precise-localization-in-language-model-steering","",{"@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/real-reading-out-transformer-activations-for-precise-localization-in-language-model-steering/85421/",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 problem does REAL target in inference-time steering?","Question",{"text":75,"@type":76},"REAL targets the challenge of selecting internal transformer modules (heads or layers) that most strongly govern a desired behavior, since existing cue-based or heuristic methods can yield weak or unintended effects.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How does REAL identify behavior-relevant modules?",{"text":80,"@type":76},"For each module, REAL trains a vector-quantized autoencoder on its hidden activations, partitioning the latent space into behavior-relevant and behavior-irrelevant subspaces using a shared learnable codebook, then computes a relevance score via binary classification between behavior-aligned and behavior-violating responses.",{"name":82,"@type":73,"acceptedAnswer":83},"What improvements does REAL achieve compared with ITI?",{"text":84,"@type":76},"On truthfulness steering, REAL reports an average relative improvement of 20% with results up to 81.5% over the seminal ITI method, and it also demonstrates strong zero-shot generalization in cross-domain truthfulness-steering scenarios.","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},[93,97,101,105,110,115,120,123,128,131,135],{"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":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":45,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":121,"slug":122},30,"research-report",{"id":124,"doc_module":4,"doc_module_name":45,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":45,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":132,"doc_module":4,"doc_module_name":45,"category_name":133,"show_sort_weight":132,"slug":134},10,"Lifestyle","lifestyle",{"id":136,"doc_module":4,"doc_module_name":45,"category_name":137,"show_sort_weight":106,"slug":138},19,"General","general"]