[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-85255-en":3,"doc-seo-85255-105":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},85255,1374391974564,"Clementine","https://ap-avatar.wpscdn.com/avatar/14000253aa45c000a9e?x-image-process=image/resize,m_fixed,w_180,h_180&k=1779874745381141002",8,"Research & Report","Route Communicate and Reason Gated Routing and Adaptive Depth for Efficient Multi-Agent Reasoning","Multi-agent ensembling increases active parameters and inference cost while still lacking answers to three core issues: which agents to consult, how deeply to traverse an agent hierarchy, and when communication trade-offs justify inter-agent exchange. GRADE introduces hierarchical gated routing with adaptive depth, using four lightweight learned gates to select agents, set depth, enable selective communication, and prune low-utility branches. Training with CoGRPO jointly optimizes all gates and agents. With ~17B active parameters, GRADE improves GSM8K, MMLUPro, and GPQA, with notable gains on MMLUPro and cost-aware efficiency.","arXiv :2607 . 10836v 1 [ cs .AI] 12 Jul 2026  \n lcs2 , iit delhi July 2026  \nRoute, Communicate, and Reason: Gated Routing and Adaptive Depth for Efficient Multi-Agent Reasoning  \nSudipto Ghosh 1 Tanmoy Chakraborty 1,2  \n1 Yardi School of Artificial Intelligence, Indian Institute of Technology Delhi  \n2 Department of Electrical Engineering, Indian Institute of Technology Delhi  \n[sudipto.ghosh@scai.iitd.ac.in](sudipto.ghosh@scai.iitd.ac.in) , [tanchak@iitd.ac.in](tanchak@iitd.ac.in)  \n[a](a github.com/parmanu-lcs2/grade)[ github.com/parmanu-lcs2/grade](a github.com/parmanu-lcs2/grade)  \nAbstract Multi-agent ensembling multiplies active parameters and inference cost without answering three basic questions: which agents to consult, how deeply a query should traverse a hierarchy of agents, and when inter-agent communication is worth its cost. We present GRADE (Gated Routing and Adaptive Depth for Efficient Reasoning), a hierarchical multi-agent system in which four lightweight learned gates jointly govern agent selection, hierarchy depth, inter-agent communication, and branch pruning. Training uses CoGRPO (Collaborative Group-Relative Policy Optimization), a novel critic-free recipe that adapts GRPO to multi-agent hierarchies and assigns a shared advantage signal to every gate and agent that participated in a rollout. Agent models are drawn from a hot-swappable Expert Registry; per-agent calibration maps allow experts to be replaced at inference time without retraining. At ∼ 17B average active parameters, GRADE outperforms all baselines on GSM8K, MMLUPro, and GPQA, surpassing the strongest baseline by 4.8 points on MMLUPro at half the active compute. On AIME-2025, where model depth dominates, GRADE remains competitive to existing frameworks. Ablations isolate the hierarchy and masked cross-attention as the largest contributors to accuracy, and show that per-agent calibration is necessary for safe hot-swapping.  \n 1 Introduction and Prior Art  \nEach successive doubling of parameters in Large Language Models (LLMs) yields diminishing returns on popular benchmarks while linearly inflating memory and inference latency [1] . This cost-quality tradeoff has pushed the field toward multi-agent LLM coordination, where a pool of smaller, specialized models collaborates to outperform a larger monolith at a fraction of the active-parameter budget.  \nPrior multi-agent coordination strategies span several paradigms. General-purpose agent frameworks such as AutoGen [2], MetaGPT [3], CAMEL [4], and ChatDev [5] orchestrate role-specialized LLMs through hand-designed communication protocols, but their topology and routing are fixed by the prompt rather than learned. Debate and critique methods [6, 7] allow multiple model instances to exchange arguments before reaching a consensus, improving factuality at the cost of requiring multiple rounds of interaction and back-and-forth communication with exploding context lengths. Iterative self-refinement [8] uses extra compute to improve a single model’s output, but cannot exploit a heterogeneous pool; reasoning-and-acting loops [9] interleave tool calls but still operate on a single model. Ensemble and fusion approaches [10, 11] rank, vote over, or fuse outputs from independently queried models; LLM-Blender [10] adds pairwise reranking and generative fusion, but routing is fixed rather than learned. Mixture-of-Agents [12] uses layered aggregation to refine outputs across rounds, but every agent processes every query, so compute is not adaptive. Evolutionary orchestration methods such as EvoAgent [13] and Puppeteer [14], restructure agent graphs dynamically; Puppeteer achieves strong results via RL-trained orchestration of a flat pool of parameter-matched models. On the training side, MAPoRL [15] jointly trains multiple LLMs with multi-agent PPO, an influence-aware verification reward, and a GAE critic, establishing that joint co-training is necessary – SFTon expert trajectories alone does not produce colla","cbCaiuMtbD7DILF3","https://ap.wps.com/l/cbCaiuMtbD7DILF3","pdf",705474,1,20,"English","en",105,"# Introduction and Prior Art\n## Overview of GRADE\n## Limitations in Prior Systems","[{\"question\":\"What problem does GRADE address in multi-agent reasoning systems?\",\"answer\":\"GRADE targets inefficiency caused by ensembling, specifically deciding which agents to consult, how far to traverse an agent hierarchy, and when inter-agent communication is worth its cost.\"},{\"question\":\"How does GRADE control routing, depth, and communication?\",\"answer\":\"It uses four learned gates to govern agent selection, hierarchy depth, selective cross-agent communication, and pruning of low-utility branches before aggregation.\"},{\"question\":\"How is GRADE trained and what optimization signals are used?\",\"answer\":\"GRADE is trained with CoGRPO, a critic-free adaptation of GRPO that samples coordination rollouts and assigns a shared group-relative advantage signal to every gate and participating agent.\"}]",1784202103,50,{"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},"route-communicate-and-reason-gated-routing-and-adaptive-depth-for-efficient-multi-agent-reasoning","",{"@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/route-communicate-and-reason-gated-routing-and-adaptive-depth-for-efficient-multi-agent-reasoning/85255/",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 GRADE address in multi-agent reasoning systems?","Question",{"text":75,"@type":76},"GRADE targets inefficiency caused by ensembling, specifically deciding which agents to consult, how far to traverse an agent hierarchy, and when inter-agent communication is worth its cost.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How does GRADE control routing, depth, and communication?",{"text":80,"@type":76},"It uses four learned gates to govern agent selection, hierarchy depth, selective cross-agent communication, and pruning of low-utility branches before aggregation.",{"name":82,"@type":73,"acceptedAnswer":83},"How is GRADE trained and what optimization signals are used?",{"text":84,"@type":76},"GRADE is trained with CoGRPO, a critic-free adaptation of GRPO that samples coordination rollouts and assigns a shared group-relative advantage signal to every gate and participating agent.","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},[93,97,101,105,110,114,119,122,126,129,133],{"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":28,"slug":113},6,"Technology","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":21,"slug":125},9,"Religion & Spirituality","religion-spirituality",{"id":21,"doc_module":4,"doc_module_name":45,"category_name":127,"show_sort_weight":21,"slug":128},"World Cup","world-cup",{"id":130,"doc_module":4,"doc_module_name":45,"category_name":131,"show_sort_weight":130,"slug":132},10,"Lifestyle","lifestyle",{"id":134,"doc_module":4,"doc_module_name":45,"category_name":135,"show_sort_weight":106,"slug":136},19,"General","general"]