[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-512776-105":59,"doc-detail-512776-en":131},{"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":124,"head_meta":126,"extra_data":128,"updated_unix":130},105,"en","dnd-boosting-large-language-models-with-dynamic-nested-depth-512776","DND: Boosting Large Language Models with Dynamic Nested Depth","","Dynamic Nested Depth (DND) introduces a token-level reprocessing method for off-the-shelf LLMs. DND selects critical tokens at the end of each transformer layer using a router, then feeds the selected subset for an additional nested round of processing. This “review” approach increases understanding of difficult tokens while avoiding redundant computation for easier ones. The method uses a router controlling loss to improve token selection distinguishability and a threshold control scheme to ensure selection stability, validated through post-training integration on dense and MoE models.",{"@graph":69,"@context":123},[70,84,106],{"@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/dnd-boosting-large-language-models-with-dynamic-nested-depth-512776/512776/",{"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/dnd-boosting-large-language-models-with-dynamic-nested-depth-512776/512776.png","ImageObject",300,407,{"name":92,"@type":93},"Connor ","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-11-11","2026-10-06",true,{"@type":102,"interactionType":103,"userInteractionCount":105},"InteractionCounter",{"@type":104},"ViewAction",12,{"@type":107,"mainEntity":108},"FAQPage",[109,115,119],{"name":110,"@type":111,"acceptedAnswer":112},"What problem does Dynamic Nested Depth (DND) target in large language model inference?","Question",{"text":113,"@type":114},"DND targets the uneven difficulty across tokens, where some tokens require deeper computation while others can be processed more cheaply.","Answer",{"name":116,"@type":111,"acceptedAnswer":117},"How does DND decide which tokens to reprocess?",{"text":118,"@type":114},"DND uses a router at the end of transformer blocks to select critical tokens whose outputs satisfy a threshold-based criterion, then reprocesses only the selected subset.",{"name":120,"@type":111,"acceptedAnswer":121},"What strategies does DND use to make token selection stable and reliable during post-training?",{"text":122,"@type":114},"DND improves selection distinguishability via a router controlling loss, and stabilizes the selection ratio using a threshold control scheme, along with a normalized fusion strategy to integrate the nested output with the original forward output.","https://schema.org",{"og:url":83,"og:type":125,"og:title":65,"og:site_name":95,"og:description":67},"article",{"robots":127,"canonical":83},"index,follow",{"doc_id":129,"site_id":62},512776,1791335092,{"code":4,"msg":5,"data":132},{"doc_id":129,"user_id":133,"nickname":92,"user_avatar":134,"doc_module":4,"category_id":39,"category_name":40,"doc_title":65,"doc_description":67,"doc_content":135,"file_id":136,"file_url":137,"file_type":138,"file_size":139,"view_count":105,"is_deleted":4,"is_public":8,"is_downloadable":8,"audit_status":8,"page_count":140,"language":141,"language_code":63,"site_id":62,"html_lang":63,"table_of_contents":142,"faqs":143,"seo_title":144,"seo_description":67,"update_tm":145,"read_time":146,"preview_page_count":19,"preview_urls":147},687207022233,"https://ap-avatar.wpscdn.com/davatar_155a257f0dc6eb9ab79c44ca47cae57d","arXiv :2510 . 11001v3 [ cs .CL] 27 Jan 2026  \nDND: Boosting Large Language Models with Dynamic  \nNested Depth  \nTieyuan Chen2,1,3∗ , Xiaodong Chen4,1, Haoxing Chen1, Zhenzhong Lan1,5, Weiyao Lin2,3†, Jianguo Li1†  \n1Inclusion AI 2Shanghai Jiao Tong University  \n3ZhongguanCun Academy 4Renmin University of China 5Westlake University  \nAbstract  \nWe introduce Dynamic Nested Depth (DND), a novel method that improves performance for off-the-shelf LLMs by selecting critical tokens to reprocess in a nested depth manner.  \nSpecifically, at the end of the given transformer layer, DND identifies more critical tokens with a router and feeds them back for an extra round of processing, effectively “reviewing”  \ndifficult tokens while avoiding redundant computation for easier ones. The dynamic selection mechanism is tailored for precise control via two novel strategies: a router controlling loss to enhance token selection distinguishability, and a threshold control scheme to ensure selection stability. We demonstrate the effectiveness of DND by directly integrating it into pre-trained dense and MoE models during a post-training phase. On diverse benchmarks, DND boosts the performances of the dense Qwen3-1.7B, Llama3.2-1B, and Gemma3-1Bby 1.88%, 2.61%, and 2.50% and the MoE Qwen3-30B-A3B by 0.87%, all with a minimal parameter and computing increase.  \n1 introduction  \nLarge Language Models (LLMs) have transformed artificial intelligence with their powerful abilities. The main strategy for improving them has been scaling, as empirical laws show that model performance predictably increases with more parameters, data, and computation (Achiam et al., 2023; Team et al., 2024; Yang et al., 2025; Liu et al., 2024) . However, this scaling paradigm has significant drawbacks. The computational overhead for both training and inference grows exponentially with model size. This trend underscores a critical need for more efficient approaches to enhance model performance beyond simple brute-force scaling.  \nA key insight from (Gloeckle et al., 2024) is that prediction difficulty varies significantly across tokens; some are trivial to predict, while others demand deep computational processing. This disparity motivates token-level adaptive computation, where models can focus resources on the most critical inputs. A foundational version of this approach is token pruning, which has been shown to be effective across language understanding (Bae et al., 2025), model compression (Yang et al., 2022b), and vision (Hojjat et al., 2025) . By filtering out uninformative tokens, these methods reduce computational overhead and can even improve robustness by mitigating noise. This establishes a binary choice: a token is either discarded or processed normally. Furthermore, we propose the natural next step: instead of merely retaining challenging tokens for standard processing, we should allocate additional computation to them, ensuring these critical tokens are properly understood.  \nOur choice of how to allocate additional computation is inspired by latent strategies in test-time scaling (Hao et al., 2024; Saunshi et al., 2025) . Unlike using explicit output expansion like COT, these methods recur computation in hidden states, scaling inference without extra text generation. They observed that reasoning tokens place uneven demands on computation: as illustrated in Fig. 1, most tokens serve fluency, while a few critical ones drive complex planning or logical transitions.  \nInspired by these two perspectives, we propose to integrate token-level selection with latent-space deepening.  \n∗ Work done at Ant Group.†Corresponding Authors. Accepted by ICLR 2026,‡Codes will be released soon.  \n...  \n Nested Depth  \nFigure 1: DND Motivation. The tokens highlighted in red denote critical elements in the QA pair. We propose a strategy within the transformer layers to identify and allocate additional computation to these critical tokens. Ls and Le indicate the starting and end layers t","cbCaid3AuM0CvafT","https://ap.wps.com/l/cbCaid3AuM0CvafT","pdf",1498380,17,"English","# Introduction\n## Token-level adaptive computation and token pruning\n## Latent test-time scaling and recurrent hidden-state deepening\n## DND overview and motivation\n## Dynamic selection and recalculation of critical tokens\n## Router design and normalized fusion for post-training integration","[{\"question\":\"What problem does Dynamic Nested Depth (DND) target in large language model inference?\",\"answer\":\"DND targets the uneven difficulty across tokens, where some tokens require deeper computation while others can be processed more cheaply.\"},{\"question\":\"How does DND decide which tokens to reprocess?\",\"answer\":\"DND uses a router at the end of transformer blocks to select critical tokens whose outputs satisfy a threshold-based criterion, then reprocesses only the selected subset.\"},{\"question\":\"What strategies does DND use to make token selection stable and reliable during post-training?\",\"answer\":\"DND improves selection distinguishability via a router controlling loss, and stabilizes the selection ratio using a threshold control scheme, along with a normalized fusion strategy to integrate the nested output with the original forward output.\"}]","DND: Boosting Large Language Models with Dynamic Nested Depth | PDF",1791275456,43,[148,152,156,160],{"page_no":8,"url":149,"width":150,"height":151},"https://res-academy.cache.wpscdn.com/upload/kdocshare/2026-11-17/2801c3e49bdb4d5194b43c4b7a4a01fe.webp",983,1272,{"page_no":14,"url":153,"width":154,"height":155},"https://res-academy.cache.wpscdn.com/upload/kdocshare/2026-11-17/488c50349c5ddef3198b7c6b452d6ae4.webp",982,1270,{"page_no":81,"url":157,"width":158,"height":159},"https://res-academy.cache.wpscdn.com/upload/kdocshare/2026-11-17/44518e0624e5c066bf90e938e966c371.webp",981,1269,{"page_no":19,"url":161,"width":162,"height":163},"https://res-academy.cache.wpscdn.com/upload/kdocshare/2026-11-17/55d50449ee7001789b4f3d1e57b699f6.webp",980,1268]