[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-141485-105":59,"doc-detail-141485-en":130},{"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":123,"head_meta":125,"extra_data":127,"updated_unix":129},105,"en","embedding-aligned-language-models-paper-conference","Embedding-Aligned Language Models - Paper Conference","","A proposal for training large language models to follow objectives defined inside a latent embedding space. The Embedding-Aligned Guided Language (EAGLE) agent uses reinforcement learning by treating a pretrained LLM as an environment, iteratively steering generation toward embedding regions optimized under a predefined criterion. Experiments on MovieLens 25M and Amazon Review demonstrate how EAGLE surfaces latent-demand content gaps and how a state-dependent action set improves efficiency. The work supports controlled, grounded text generation consistent with domain knowledge representations.",{"@graph":69,"@context":122},[70,84,105],{"@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/embedding-aligned-language-models-paper-conference/141485/",{"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/embedding-aligned-language-models-paper-conference/141485.png","ImageObject",300,407,{"name":92,"@type":93},"Maeve","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-09-19","2026-08-25",true,{"@type":102,"interactionType":103,"userInteractionCount":29},"InteractionCounter",{"@type":104},"ViewAction",{"@type":106,"mainEntity":107},"FAQPage",[108,114,118],{"name":109,"@type":110,"acceptedAnswer":111},"What is the core idea behind Embedding-Aligned Guided Language (EAGLE)?","Question",{"text":112,"@type":113},"EAGLE trains a language-based agent with reinforcement learning to iteratively steer an LLM’s generation toward optimal regions of a latent embedding space according to a predefined criterion.","Answer",{"name":115,"@type":110,"acceptedAnswer":116},"How does EAGLE evaluate generated outputs during training?",{"text":117,"@type":113},"After the agent modifies a textual description, the description is embedded into the latent space to assess quality with respect to the chosen objective (e.g., content gaps).",{"name":119,"@type":110,"acceptedAnswer":120},"Which datasets are used to validate the approach, and what is shown?",{"text":121,"@type":113},"MovieLens 25M and Amazon Review are used to demonstrate that EAGLE can surface content gaps aligned with latent user demand and that an efficient state-dependent action set improves performance.","https://schema.org",{"og:url":83,"og:type":124,"og:title":65,"og:site_name":95,"og:description":67},"article",{"robots":126,"canonical":83},"index,follow",{"doc_id":128,"site_id":62},141485,1787657404,{"code":4,"msg":5,"data":131},{"doc_id":128,"user_id":132,"nickname":92,"user_avatar":133,"doc_module":4,"category_id":39,"category_name":40,"doc_title":65,"doc_description":67,"doc_content":134,"file_id":135,"file_url":136,"file_type":137,"file_size":138,"view_count":29,"is_deleted":4,"is_public":8,"is_downloadable":8,"audit_status":8,"page_count":139,"language":140,"language_code":63,"site_id":62,"html_lang":63,"table_of_contents":141,"faqs":142,"seo_title":143,"seo_description":67,"update_tm":129,"read_time":144},5909877438554,"https://ap-avatar.wpscdn.com/avatar/5600025385ad2bf12a7?_k=1778553567797529272","Embedding-Aligned Language Models  \nGuy Tennenholtzy  Yinlam Chowz, Chih-Wei Hsuy, Lior Shaniy, Ethan Liangz, Craig Boutiliery  \ny Google Research, z Google Deepmind  \nAbstract  \nWe propose a novel approach for training large language models (LLMs) to adhere  \nto objectives defined within a latent embedding space. Our method leverages  \nreinforcement learning (RL), treating a pre-trained LLM as an environment. Our  \nembedding-aligned guided language (EAGLE) agent is trained to iteratively steer  \nthe LLM's generation towards optimal regions of the latent embedding space, w.r.t.  \nsome predefined criterion. We demonstrate the effectiveness of the EAGLE agent  \nusing the MovieLens 25M and Amazon Review datasets to surface content gaps  \nthat satisfy latent user demand. We also demonstrate the benefit ofusing an optimal  \ndesign of a state-dependent action set to improve EAGLE's efficiency. Our work  \npaves the way for controlled and grounded text generation using LLMs, ensuring  \nconsistency with domain-specific knowledge and data representations.  \n1 Introduction  \nLarge language models (LLMs) such as Gemini [Team et al., 2023] and GPT [Achiam et al., 2023] have revolutionized the field of natural language processing, achieving remarkable success in text generation, translation, comprehension, as well as expert-level performance on challenging tasks (e.g., exams, coding) . However, effectively applying (or fine-tuning) LLMs to domain-specific tasks often requires considerable domain knowledge and labeled human data [Jeong et al., 2023, Ouyang et al., 2022, Ziegler et al., 2019] . Fortunately, in many cases, domain knowledge is already captured and encoded within latent embedding spaces.  \nLatent embeddings are continuous vectors, ubiquitous in fields such as recommender systems [Hansen et al., 2020], reinforcement learning (RL) [Nabati et al., 2023, Pertsch et al., 2021], and imageclassification [Girdhar et al., 2023, Radford et al., 2021] . They offer a powerful means to represent entities, concepts, and relationships within a specific domain. For example, in recommender systems, embeddings of items and users encapsulate information about preferences and behavior [Zhao et al., 2023], embeddings of images can capture their content and style [Radford et al., 2021], while embeddings of scientific articles can represent their research area and findings [Taher Harikandehet al., 2023] . The utility of a latent embedding lies in its underlying compact representation of entities, concepts or relationships, and the associated metrics, allowing one to construct simpler, more efficient models or induce control over various processes [Arvanitidis et al., 2018, 2021, Radford et al., 2021, Tennenholtz and Mannor, 2022] . Importantly, latent embeddings are often pre-computed, and can thus serve as a readily available rich source of domain knowledge.  \nThis leads to the key question: can we leverage latent embeddings to better control and guide LLM generation? In this paper, we present a novel framework which accomplishes this by exploiting latent embedding spaces to define an objective function for an LLM in an iterative RL-driven process.  \nAs an example, consider the challenge of assisting content creators in generating valuable content within a recommender ecosystem (e.g., YouTube, Reddit, Spotify) [Boutilier et al., 2024] . An  \n􀀃 Correspondence to: guytenn@gmail.com  \n38th Conference on Neural Information Processing Systems (NeurIPS 2024) .  \nimportant aspect of this problem includes identifying and surfacing of content gaps, i.e., identifying hypothetical content which could potentially drive value for users, and subsequently, describing it to creators. Latent embeddings offer an effective way of defining a content gap. Informally, a content gap is a hypothetical (i.e., non-existing) content item, corresponding to some point in a latent embedding space for which: (1) no content currently exists, implying an unexplored area within the existing content ","cbCainDfyOKd17Hl","https://ap.wps.com/l/cbCainDfyOKd17Hl","pdf",1490524,54,"English","# Introduction\n## Latent embeddings as domain knowledge\n## Guiding LLM generation with embedding-aligned objectives\n## Content gap creation in recommender ecosystems\n## Contributions and validation","[{\"question\":\"What is the core idea behind Embedding-Aligned Guided Language (EAGLE)?\",\"answer\":\"EAGLE trains a language-based agent with reinforcement learning to iteratively steer an LLM’s generation toward optimal regions of a latent embedding space according to a predefined criterion.\"},{\"question\":\"How does EAGLE evaluate generated outputs during training?\",\"answer\":\"After the agent modifies a textual description, the description is embedded into the latent space to assess quality with respect to the chosen objective (e.g., content gaps).\"},{\"question\":\"Which datasets are used to validate the approach, and what is shown?\",\"answer\":\"MovieLens 25M and Amazon Review are used to demonstrate that EAGLE can surface content gaps aligned with latent user demand and that an efficient state-dependent action set improves performance.\"}]","Embedding-Aligned Language Models - Paper Conference | PDF",136]