[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-81802-en":3,"doc-seo-81802-105":29,"detail-sidebar-cat-0-en-105":90},{"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":4,"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},81802,137441390410,"Hazel","https://ap-avatar.wpscdn.com/avatar/2000252f4ab5702993?_k=1776741390130283984",8,"Research & Report","Cache Merging as a Convergent Replicated State for Multi-Agent Latent Reasoning","Multi-agent latent reasoning composes multiple agents’ KV-cache contributions into a single context for a final agent. Prior work via BagMerge concatenates per-agent caches with RoPE re-encoding, but this composition is non-commutative and ordering sensitivity depends on deployment regime, latent-step budget, and model scale. The approach introduces convergent replicated-state cache exchange: CanonicalMerge deterministically renders a commutative, associative, idempotent, absorbing CvRDT from content-addressed latent fragments, absorbing duplicates and preserving accuracy guarantees while remaining robust to delivery order.","arXiv :2607 .0 1308v 1 [ cs .MA] 1 Jul 2026  \nCache Merging as a Convergent Replicated State for Multi-Agent Latent Reasoning  \nCarlos Baquero  \nFaculdade de Engenharia (FEUP), Universidade do Porto (UP), Portugal [cbm@fe.up.pt](cbm@fe.up.pt)  \nLuís Brito  \nEscola Superior de Tecnologia e Gestão (ESTG), Politécnico de Viana do Castelo (IPVC), Portugal [brito luis@estg. ipvc.pt](brito luis@estg. ipvc.pt)  \nAbstract  \nMulti-agent latent reasoning composes the KV-cache contributions of multiple agents into a single context for a final agent. Recent work (Agent Primitives) realises this composition by concatenating per-agent caches along the sequence axis with RoPE re-encoding, a construction we name BagMerge. BagMerge is non-commutative, and which input ordering is best is not predictable a priori: it shifts with the deployment regime, the latent-step budget, and even the model scale. We make this cache exchange a convergent replicated state.  \nFirst, CanonicalMerge fixes the layout: a content-determined ordering by mean Knorm at a middle transformer layer renders the merged cache byte-identical under any permutation of the inputs, verified algorithmically on synthetic tensors (arbitrary arity N ≤ 5) and bit-for-bit on the real KV state of Qwen3-1.7B (28 layers) and Qwen3-4B (36 layers) . Second, we separate the replicated state from decode-time layout: the durable object is a set of content-addressed latent fragments whose merge is set union, a state-based CvRDT (commutative, associative, idempotent, absorbing), and CanonicalMerge is its deterministic render. Because the render is byte-equivalent, every N = 2 accuracy number is inherited unchanged and re-delivered duplicate fragments are absorbed rather than reconcatenated. On a partitioned-reasoning benchmark (Qwen3-1.7B), CanonicalMerge lands within 4 percentage points of the best BagMerge ordering in every cell of a 12-cell regime × budget × ordering matrix and matches it without needing to know which ordering is best, trading a small, statistically insignificant accuracy margin for an unconditional structural guarantee; a four-cell Qwen3-4B scale check preserves the same qualitative result.  \nThe behaviour transfers to real multi-document QA (HotpotQA bridge-k = 2: no detectable degradation against a single-agent full-context baseline), while the closest training-free output-fusion baseline (PackLLM) loses by 45 points at matched generation budget, placing cache-level merging in a regime distinct from output-level fusion. Finally, at k > 2 we delimit the approach: cache merge transports and colocates latent traces but does not by itself compose them, motivating subsequent work.  \n1 Introduction  \nMulti-agent language-model systems increasingly reason in latent space: instead of exchanging naturallanguage messages, agents pass intermediate KV-cache state directly, giving a downstream agent access to richer, higher-bandwidth representations than decoded text (Zou et al., 2025; Jin et al., 2026) . This latent exchange is realized by composing the per-agent caches into a single context that a final agent attends over. As the number of contributing agents grows, and as settings move toward parallel reasoning, recursive delegation (Yang et al., 2026), and federated inference, this composition step becomes the load-bearing operation, yet the operator that performs it has been left largely implicit.  \nThe de facto standard for the parallel case, which we name BagMerge, concatenates the agents’ caches along the sequence axis and re-encodes their positions with RoPE (Jin et al., 2026) . Concatenation is ordered, and the RoPE re-encoding makes that order matter: BagMerge is non-commutative, so which agent’s fragment occupies the privileged position-0 prefix changes the decoded answer. Worse, which ordering is best is not predictable in advance: we find it shifts with the deployment regime, the latent-step budget, and even the model scale (Section 5), so a deployer cannot fix a safe ord","cbCaigLQkiAvloTP","https://ap.wps.com/l/cbCaigLQkiAvloTP","pdf",771889,1,27,"English","en",105,"# Abstract\n# Introduction\n# Background and Motivation\n# CanonicalMerge and Convergent Cache Exchange\n# Evaluation and Benchmarks\n# Multi-Document QA and Comparisons\n# Limitations and Future Directions","[{\"question\":\"What problem does the paper address in multi-agent latent reasoning?\",\"answer\":\"It addresses how to compose per-agent KV caches into one context for a final agent when multiple agents contribute latent state, aiming for order-independence and robustness to duplicate fragment delivery.\"},{\"question\":\"How is BagMerge characterized, and why is it insufficient?\",\"answer\":\"BagMerge concatenates agent caches along the sequence axis and uses RoPE re-encoding, making the merge non-commutative. Best ordering is unpredictable and duplicates are not absorbed, so it lacks coordinator-free convergence guarantees.\"},{\"question\":\"What are the key components of the proposed convergent replicated-state solution?\",\"answer\":\"It introduces CanonicalMerge, which fixes layout using a content-determined ordering, and separates a replicated state of content-addressed latent fragments combined by set union as a CvRDT; CanonicalMerge is the deterministic render.\"}]",1784176255,68,{"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":85,"head_meta":87,"extra_data":89,"updated_unix":27},"cache-merging-as-a-convergent-replicated-state-for-multi-agent-latent-reasoning","",{"@graph":35,"@context":84},[36,53,67],{"@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/cache-merging-as-a-convergent-replicated-state-for-multi-agent-latent-reasoning/81802/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":23,"description":14,"dateModified":61,"datePublished":61,"encodingFormat":60,"isAccessibleForFree":62,"interactionStatistic":63},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":40,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-16",true,{"@type":64,"interactionType":65,"userInteractionCount":4},"InteractionCounter",{"@type":66},"ViewAction",{"@type":68,"mainEntity":69},"FAQPage",[70,76,80],{"name":71,"@type":72,"acceptedAnswer":73},"What problem does the paper address in multi-agent latent reasoning?","Question",{"text":74,"@type":75},"It addresses how to compose per-agent KV caches into one context for a final agent when multiple agents contribute latent state, aiming for order-independence and robustness to duplicate fragment delivery.","Answer",{"name":77,"@type":72,"acceptedAnswer":78},"How is BagMerge characterized, and why is it insufficient?",{"text":79,"@type":75},"BagMerge concatenates agent caches along the sequence axis and uses RoPE re-encoding, making the merge non-commutative. Best ordering is unpredictable and duplicates are not absorbed, so it lacks coordinator-free convergence guarantees.",{"name":81,"@type":72,"acceptedAnswer":82},"What are the key components of the proposed convergent replicated-state solution?",{"text":83,"@type":75},"It introduces CanonicalMerge, which fixes layout using a content-determined ordering, and separates a replicated state of content-addressed latent fragments combined by set union as a CvRDT; CanonicalMerge is the deterministic render.","https://schema.org",{"og:url":51,"og:type":86,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":88,"canonical":51},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":91},[92,96,100,104,109,114,119,122,127,130,134],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":93,"show_sort_weight":94,"slug":95},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":97,"show_sort_weight":98,"slug":99},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":45,"category_name":101,"show_sort_weight":102,"slug":103},"Exam",70,"exam",{"id":105,"doc_module":4,"doc_module_name":45,"category_name":106,"show_sort_weight":107,"slug":108},5,"Comic",60,"comic",{"id":110,"doc_module":4,"doc_module_name":45,"category_name":111,"show_sort_weight":112,"slug":113},6,"Technology",50,"technology",{"id":115,"doc_module":4,"doc_module_name":45,"category_name":116,"show_sort_weight":117,"slug":118},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":120,"slug":121},30,"research-report",{"id":123,"doc_module":4,"doc_module_name":45,"category_name":124,"show_sort_weight":125,"slug":126},9,"Religion & Spirituality",20,"religion-spirituality",{"id":125,"doc_module":4,"doc_module_name":45,"category_name":128,"show_sort_weight":125,"slug":129},"World Cup","world-cup",{"id":131,"doc_module":4,"doc_module_name":45,"category_name":132,"show_sort_weight":131,"slug":133},10,"Lifestyle","lifestyle",{"id":135,"doc_module":4,"doc_module_name":45,"category_name":136,"show_sort_weight":105,"slug":137},19,"General","general"]