[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-327256-105":59,"doc-detail-327256-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","the-state-prediction-separation-hypothesis","The State-Prediction Separation Hypothesis","","Transformers use the same forward computation stream to both predict the next token and store state for future token predictions. The paper proposes the state-prediction separation hypothesis, arguing that disentangling these roles improves language modeling. It introduces a Transformer variant with two computation streams and runs pretraining experiments across multiple scales. Results show consistent gains in data and compute efficiency, lowering validation loss and improving downstream performance by 2–3 percentage points on average.",{"@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/the-state-prediction-separation-hypothesis/327256/",{"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/the-state-prediction-separation-hypothesis/327256.png","ImageObject",300,407,{"name":92,"@type":93},"Lucas Martin","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-10-31","2026-09-21",true,{"@type":102,"interactionType":103,"userInteractionCount":39},"InteractionCounter",{"@type":104},"ViewAction",{"@type":106,"mainEntity":107},"FAQPage",[108,114,118],{"name":109,"@type":110,"acceptedAnswer":111},"What problem does the State-Prediction Separation Hypothesis target in Transformers?","Question",{"text":112,"@type":113},"It targets the entanglement of two roles in the same forward computation stream: next-token prediction and storing reusable state for later predictions.","Answer",{"name":115,"@type":110,"acceptedAnswer":116},"How does the proposed SPS Transformer separate prediction and state?",{"text":117,"@type":113},"It inserts an additional computation time step so that the model forms a persistent state in one step and emits the next token in the following step, then manages key-value cache usage accordingly.",{"name":119,"@type":110,"acceptedAnswer":120},"What do the experiments show about SPS compared with standard Transformers?",{"text":121,"@type":113},"State-prediction separation consistently improves data and compute efficiencies, reducing validation loss and outperforming standard Transformers by about 2–3 percentage points on average on downstream tasks.","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},327256,1790704841,{"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":39,"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":144,"read_time":145,"preview_page_count":4},8796095360427,"https://ap-avatar.wpscdn.com/davatar_994ba38a5ba835b3df7d355c54d3ed8d","arXiv :2607 .0 12 18v 1 [ cs .CL] 1 Jul 2026  \nThe State-Prediction Separation Hypothesis  \nGiovanni Monea† Nathan Godey† Kianté Brantley⋄ Yoav Artzi††Cornell University ⋄Harvard University  \n[giovanni@cs.cornell.edu](giovanni@cs.cornell.edu) , {ng554, [yoavartzi}@cornell.edu](yoavartzi}@cornell.edu)[kdbrantley@g.harvard.edu](kdbrantley@g.harvard.edu)  \nAbstract  \nTransformers use the same forward computation stream to both predict the next token and store useful state for future token predictions. We formulate the stateprediction separation hypothesis: disentangling the two roles yields better language modeling performance. We design a Transformer variant that uses two computation streams to separate the two functions, and conduct pretraining experiments across various scales. Our experiments show that state-prediction separation consistently offers better data and compute efficiencies, improving validation loss and outperforming standard Transformers by 2–3 percentage points on average on downstream tasks. We also conduct extensive empirical analysis that rules out potential confounders and demonstrates the fundamental difference in the gradientsour design entails.  \nPre-Decay NLL  \n2.61 2.58 2.55  \n2.52  \n KV cached activations  \nKV ephemeral activations  \n10B 20B 30B 40B  \nTraining Tokens  \n~~ ~~ STANDARD (47B tokens)  STATE-PREDICTION SEPARATION (18B tokens)  \nFigure 1: Standard versus State-Prediction Separation Transformer. Top: The standard Transformer uses the same hidden states for both memory and prediction. Our variant separates these roles: input token xi time steps form a persistent state, while prediction token steps ρi produce next-token predictions. Bottom: At 1.6B parameters, State-Prediction Separation matches the validation loss of a standard Transformer trained on 47B tokens while using 2.6 × fewer tokens (pre-decay) . At an 18B-token pre-decay budget, it already achieves ∆NLL = −0 .071 versus standard.  \nPreprint.  \n1 Introduction  \nAttention-based architectures, including the Transformer [Vaswani et al., 2017] and earlier recurrent designs [Bahdanau et al., 2015], have dual use for the activations computed at each time step: they are used to predict the output of that time step (i.e., token in language models) and are attended to by subsequent steps. The first role is focused on prediction; the second on capturing state information to be reused later on. Generally, these two functions are entangled in the same representation and computation stream (i.e., forward path) . In this paper, we propose and study the following hypothesis in large language models (LLMs):  \nHypothesis: State-Prediction Separation (SPS; informal)  \nThe next-token prediction computation and state representation compete when forced through the same computation. Routing them through separate streams yields better language modeling.  \nTechnically, we separate the state and prediction functionalities by inserting an additional computation time step before predicting the next token (Figure 1) . Time steps then appear in pairs: first, the token previously generated is processed, but no new token is emitted. An additional time step follows, which emits the next token to generate. The key-value (KV) entries from the first of the two steps are added to the KV cache, while the entries of the latter of the two are discarded.1 This design distinguishes between two streams: a state stream and a prediction stream.  \nWe conduct extensive experiments pretraining a set ofLLMs at common research scales, from 53M to 1.678B parameters. The main result is that SPS significantly improves pretraining performance over standard Transformers in both token-equivalent and compute-equivalent settings. Figure 1 illustrates one of the key results: non-separating baselines cannot match the training loss of an SPS Transformer even with double the number of training tokens. We also show that state-prediction separation outperforms several variants controlling for SP","cbCaicfIu3uZkQwc","https://ap.wps.com/l/cbCaicfIu3uZkQwc","pdf",1547777,15,"English","# Abstract\n# Introduction\n# Prediction and State Preparation","[{\"question\":\"What problem does the State-Prediction Separation Hypothesis target in Transformers?\",\"answer\":\"It targets the entanglement of two roles in the same forward computation stream: next-token prediction and storing reusable state for later predictions.\"},{\"question\":\"How does the proposed SPS Transformer separate prediction and state?\",\"answer\":\"It inserts an additional computation time step so that the model forms a persistent state in one step and emits the next token in the following step, then manages key-value cache usage accordingly.\"},{\"question\":\"What do the experiments show about SPS compared with standard Transformers?\",\"answer\":\"State-prediction separation consistently improves data and compute efficiencies, reducing validation loss and outperforming standard Transformers by about 2–3 percentage points on average on downstream tasks.\"}]","The State-Prediction Separation Hypothesis | PDF",1789976837,38]