Most MoA systems freeze routing while agents learn — or freeze agents while the router is tuned. CERA-MoA closes that loop: a familiarity-based router and N LoRA agents co-evolve with RL. Mid-layer hidden states estimate each agent’s competence without full rollouts; cumulative-threshold routing then activates the smallest capable subset. Code : <a href=\"https://github.com/michaeljiang0530/CERA-MoA\" rel=\"nofollow\">https://github.com/michaeljiang0530/CERA-MoA</a></p>\n","updatedAt":"2026-09-17T11:39:08.873Z","author":{"_id":"69c3d6894f60bf51a1d0690c","avatarUrl":"/avatars/969cd683fdc94ea9a87490c1435a5d43.svg","fullname":"Jiaxuan","name":"mj0530","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":1,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.7499290108680725},"editors":["mj0530"],"editorAvatarUrls":["/avatars/969cd683fdc94ea9a87490c1435a5d43.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2609.18779","authors":[{"_id":"6aab52f71d9cc4dec7962450","user":{"_id":"69c3d6894f60bf51a1d0690c","avatarUrl":"/avatars/969cd683fdc94ea9a87490c1435a5d43.svg","isPro":false,"fullname":"Jiaxuan","user":"mj0530","type":"user","name":"mj0530"},"name":"Jiaxuan Jiang","status":"claimed_verified","statusLastChangedAt":"2026-09-17T09:09:51.939Z","hidden":false},{"_id":"6aab52f71d9cc4dec7962451","name":"Liyuan He","hidden":false},{"_id":"6aab52f71d9cc4dec7962452","name":"Zhixuan Fang","hidden":false}],"publishedAt":"2026-09-16T00:00:00.000Z","submittedOnDailyAt":"2026-09-17T00:00:00.000Z","title":"CERA-MoA: Co-Evolving Routing Mechanisms with Continually Learning LLM Agents","submittedOnDailyBy":{"_id":"69c3d6894f60bf51a1d0690c","avatarUrl":"/avatars/969cd683fdc94ea9a87490c1435a5d43.svg","isPro":false,"fullname":"Jiaxuan","user":"mj0530","type":"user","name":"mj0530"},"summary":"Current Mixture-of-Agents (MoA) paradigms generally treat query routing and agent fine-tuning as separate processes, limiting their ability to respond to evolving agent capabilities. This disconnect prevents routing strategies from adapting to evolving agent capabilities during post-training and prevents agents from achieving synergistic data-driven specialization. To resolve this, we introduce CERA-MoA (Co-Evolving Router with continually learning Agents for Mixture-of-Agents), an iterative reinforcement learning framework where the dynamic router and independent agent policies co-evolve. We design a predictive familiarity estimator that leverages mid-layer hidden states to evaluate semantic competence among agents, avoiding the overhead of full rollouts. Based on these familiarity scores, a cumulative-threshold adaptive routing mechanism dynamically activates a tailored minimal agent subset, achieving a trade-off between task performance and efficiency. By proactively allocating targeted training samples to agents based on their evolving competence, CERA-MoA promotes capability differentiation. Extensive experiments across various domains demonstrate that CERA-MoA outperforms state-of-the-art static-agent routing and fix-workflow fine-tuning baselines.","upvotes":3,"discussionId":"6aab52f71d9cc4dec7962453","githubRepo":"https://github.com/michaeljiang0530/CERA-MoA","githubRepoAddedBy":"user","githubStars":0,"organization":{"_id":"628735cbc83a2d6ab8d14a66","name":"Tsinghua","fullname":"Tsinghua University","avatar":"https://www.gravatar.com/avatar/6c5c1441e3283e7543342e59277ea219?d=retro&size=100"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"69c3d6894f60bf51a1d0690c","avatarUrl":"/avatars/969cd683fdc94ea9a87490c1435a5d43.svg","isPro":false,"fullname":"Jiaxuan","user":"mj0530","type":"user"},{"_id":"6a2da6c8ca070ee12c6e396c","avatarUrl":"/avatars/0355287dcabaa67dbc7f0b10b87451f9.svg","isPro":false,"fullname":"Joe Mama","user":"JoeMama123123123","type":"user"},{"_id":"6270324ebecab9e2dcf245de","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6270324ebecab9e2dcf245de/cMbtWSasyNlYc9hvsEEzt.jpeg","isPro":false,"fullname":"Kye Gomez","user":"kye","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"628735cbc83a2d6ab8d14a66","name":"Tsinghua","fullname":"Tsinghua University","avatar":"https://www.gravatar.com/avatar/6c5c1441e3283e7543342e59277ea219?d=retro&size=100"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2609/2609.18779.md","query":{}}">

Get this paper in your agent:

Don't have the latest CLI?

Cite arxiv.org/abs/2609.18779 in a dataset README.md to link it from this page.

Cite arxiv.org/abs/2609.18779 in a Space README.md to link it from this page.

Add this paper to a collection to link it from this page.