<a href=\"https://cdn-uploads.huggingface.co/production/uploads/66614d9472f344f198f88d11/bvuLsWqAnp7k2qkhL9b0q.jpeg\" rel=\"nofollow\"><img src=\"https://cdn-uploads.huggingface.co/production/uploads/66614d9472f344f198f88d11/bvuLsWqAnp7k2qkhL9b0q.jpeg\" alt=\"Snipaste_2026-05-12_11-58-54\"></a></p>\n<p><a href=\"https://cdn-uploads.huggingface.co/production/uploads/66614d9472f344f198f88d11/iNVF0BuF2rOmaZF_-ZMPh.jpeg\" rel=\"nofollow\"><img src=\"https://cdn-uploads.huggingface.co/production/uploads/66614d9472f344f198f88d11/iNVF0BuF2rOmaZF_-ZMPh.jpeg\" alt=\"Snipaste_2026-05-12_11-59-28\"></a></p>\n","updatedAt":"2026-05-12T03:59:46.923Z","author":{"_id":"66614d9472f344f198f88d11","avatarUrl":"/avatars/f5dec0177b33b0d2692e78c2df6c7783.svg","fullname":"Xinyu Lin","name":"Lxyhaha","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":1,"isUserFollowing":false}},"numEdits":1,"identifiedLanguage":{"language":"en","probability":0.2547594904899597},"editors":["Lxyhaha"],"editorAvatarUrls":["/avatars/f5dec0177b33b0d2692e78c2df6c7783.svg"],"reactions":[],"isReport":false}},{"id":"6a0443dbdeaeb1383bcb3de5","author":{"_id":"66614d9472f344f198f88d11","avatarUrl":"/avatars/f5dec0177b33b0d2692e78c2df6c7783.svg","fullname":"Xinyu Lin","name":"Lxyhaha","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":1,"isUserFollowing":false},"createdAt":"2026-05-13T09:26:51.000Z","type":"comment","data":{"edited":false,"hidden":false,"latest":{"raw":"Multi-agent systems (MAS) have emerged as a promising paradigm for solving complex tasks.\nRecent work has explored self-evolving MAS that automatically optimize agent capabilities or\ncommunication topologies. However, existing methods either learn a topology that remains\nfixed at inference time or adapt only topology or capability during inference. We empirically\nand theoretically show that effective test-time evolution requires jointly adapting both axes, but\non different time scales: capabilities should update rapidly to handle emerging subtasks, while\nthe topology should evolve more slowly to preserve coordination stability. We then introduce\nTacoMAS, a test-time co-evolution framework for dynamic MAS. TacoMAS formulates MAS\ninference as a task of online graph adaptation, where nodes represent agents with role-specific\ncapability and edges define their communication topology. During inference, a fast capability\nloop updates agent expertise using trajectory-level feedback, while a slow meta-LLM-driven\ntopology loop performs agents’ birth-death operations on MAS, including edge edit, agent\naddition, and agent removal. We further show that this fast–slow design drives MAS evolution\ntoward a task-conditioned stable equilibrium. Experiments on four benchmarks demonstrate\nthat TacoMAS outperforms nearly 20 multi-agent baselines, achieving an average improvement\nof 13.3% over the strongest baseline.","html":"<p>Multi-agent systems (MAS) have emerged as a promising paradigm for solving complex tasks.<br>Recent work has explored self-evolving MAS that automatically optimize agent capabilities or<br>communication topologies. However, existing methods either learn a topology that remains<br>fixed at inference time or adapt only topology or capability during inference. We empirically<br>and theoretically show that effective test-time evolution requires jointly adapting both axes, but<br>on different time scales: capabilities should update rapidly to handle emerging subtasks, while<br>the topology should evolve more slowly to preserve coordination stability. We then introduce<br>TacoMAS, a test-time co-evolution framework for dynamic MAS. TacoMAS formulates MAS<br>inference as a task of online graph adaptation, where nodes represent agents with role-specific<br>capability and edges define their communication topology. During inference, a fast capability<br>loop updates agent expertise using trajectory-level feedback, while a slow meta-LLM-driven<br>topology loop performs agents’ birth-death operations on MAS, including edge edit, agent<br>addition, and agent removal. We further show that this fast–slow design drives MAS evolution<br>toward a task-conditioned stable equilibrium. Experiments on four benchmarks demonstrate<br>that TacoMAS outperforms nearly 20 multi-agent baselines, achieving an average improvement<br>of 13.3% over the strongest baseline.</p>\n","updatedAt":"2026-05-13T09:26:51.254Z","author":{"_id":"66614d9472f344f198f88d11","avatarUrl":"/avatars/f5dec0177b33b0d2692e78c2df6c7783.svg","fullname":"Xinyu Lin","name":"Lxyhaha","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":1,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8881892561912537},"editors":["Lxyhaha"],"editorAvatarUrls":["/avatars/f5dec0177b33b0d2692e78c2df6c7783.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2605.09539","authors":[{"_id":"6a02a4e0b823258e76123574","name":"Chen Xu","hidden":false},{"_id":"6a02a4e0b823258e76123575","name":"Yicheng Hu","hidden":false},{"_id":"6a02a4e0b823258e76123576","name":"Ruizi Wang","hidden":false},{"_id":"6a02a4e0b823258e76123577","name":"Xinyu Lin","hidden":false},{"_id":"6a02a4e0b823258e76123578","name":"Wenjie Wang","hidden":false},{"_id":"6a02a4e0b823258e76123579","name":"Dongrui Liu","hidden":false},{"_id":"6a02a4e0b823258e7612357a","name":"Fuli Feng","hidden":false}],"mediaUrls":["https://cdn-uploads.huggingface.co/production/uploads/66614d9472f344f198f88d11/Bjm2KqB1RMphozgcs0RCt.jpeg"],"publishedAt":"2026-05-10T00:00:00.000Z","submittedOnDailyAt":"2026-05-13T00:00:00.000Z","title":"TacoMAS: Test-Time Co-Evolution of Topology and Capability in LLM-based Multi-Agent Systems","submittedOnDailyBy":{"_id":"66614d9472f344f198f88d11","avatarUrl":"/avatars/f5dec0177b33b0d2692e78c2df6c7783.svg","isPro":false,"fullname":"Xinyu Lin","user":"Lxyhaha","type":"user","name":"Lxyhaha"},"summary":"Multi-agent systems (MAS) have emerged as a promising paradigm for solving complex tasks. Recent work has explored self-evolving MAS that automatically optimize agent capabilities or communication topologies. However, existing methods either learn a topology that remains fixed at inference time or adapt only the topology or capability during inference. We empirically and theoretically show that effective test-time evolution requires jointly adapting both axes, but on different time scales: capabilities should update rapidly to handle emerging subtasks, while the topology should evolve more slowly to preserve coordination stability. We then introduce TacoMAS, a test-time co-evolution framework for dynamic MAS. TacoMAS formulates MAS inference as a task of online graph adaptation, where nodes represent agents with role-specific capabilities and edges define their communication topology. During inference, a fast capability loop updates agent expertise using trajectory-level feedback, while a slow meta-LLM-driven topology loop performs agents' birth-death operations on MAS, including edge edit, agent addition, and agent removal. We further show that this fast-slow design drives MAS evolution toward a task-conditioned stable equilibrium. Experiments on four benchmarks demonstrate that TacoMAS outperforms nearly 20 multi-agent baselines, achieving an average improvement of 13.3% over the strongest baseline. The codes are released at https://github.com/chenxu2-gif/TacoMAS-MultiAgent.","upvotes":2,"discussionId":"6a02a4e0b823258e7612357b","githubRepo":"https://github.com/chenxu2-gif/TacoMAS-MultiAgent","githubRepoAddedBy":"user","ai_summary":"Test-time co-evolution framework for multi-agent systems that jointly adapts agent capabilities and communication topology at different time scales to achieve task-conditioned stability and improved performance.","ai_keywords":["multi-agent systems","self-evolving systems","test-time evolution","online graph adaptation","fast-slow design","trajectory-level feedback","meta-LLM-driven topology","agents' birth-death operations","edge edit","agent addition","agent removal","task-conditioned stable equilibrium"],"githubStars":1},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"66614d9472f344f198f88d11","avatarUrl":"/avatars/f5dec0177b33b0d2692e78c2df6c7783.svg","isPro":false,"fullname":"Xinyu Lin","user":"Lxyhaha","type":"user"},{"_id":"6824c001b02b1366297dffe7","avatarUrl":"/avatars/7353dd2c6a41e1c67db471d9c1d67cd9.svg","isPro":false,"fullname":"Anonymous","user":"getbetterhyccc","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2605/2605.09539.md"}">
TacoMAS: Test-Time Co-Evolution of Topology and Capability in LLM-based Multi-Agent Systems
Abstract
Test-time co-evolution framework for multi-agent systems that jointly adapts agent capabilities and communication topology at different time scales to achieve task-conditioned stability and improved performance.
AI-generated summary
Multi-agent systems (MAS) have emerged as a promising paradigm for solving complex tasks. Recent work has explored self-evolving MAS that automatically optimize agent capabilities or communication topologies. However, existing methods either learn a topology that remains fixed at inference time or adapt only the topology or capability during inference. We empirically and theoretically show that effective test-time evolution requires jointly adapting both axes, but on different time scales: capabilities should update rapidly to handle emerging subtasks, while the topology should evolve more slowly to preserve coordination stability. We then introduce TacoMAS, a test-time co-evolution framework for dynamic MAS. TacoMAS formulates MAS inference as a task of online graph adaptation, where nodes represent agents with role-specific capabilities and edges define their communication topology. During inference, a fast capability loop updates agent expertise using trajectory-level feedback, while a slow meta-LLM-driven topology loop performs agents' birth-death operations on MAS, including edge edit, agent addition, and agent removal. We further show that this fast-slow design drives MAS evolution toward a task-conditioned stable equilibrium. Experiments on four benchmarks demonstrate that TacoMAS outperforms nearly 20 multi-agent baselines, achieving an average improvement of 13.3% over the strongest baseline. The codes are released at https://github.com/chenxu2-gif/TacoMAS-MultiAgent.
Community
Multi-agent systems (MAS) have emerged as a promising paradigm for solving complex tasks.
Recent work has explored self-evolving MAS that automatically optimize agent capabilities or
communication topologies. However, existing methods either learn a topology that remains
fixed at inference time or adapt only topology or capability during inference. We empirically
and theoretically show that effective test-time evolution requires jointly adapting both axes, but
on different time scales: capabilities should update rapidly to handle emerging subtasks, while
the topology should evolve more slowly to preserve coordination stability. We then introduce
TacoMAS, a test-time co-evolution framework for dynamic MAS. TacoMAS formulates MAS
inference as a task of online graph adaptation, where nodes represent agents with role-specific
capability and edges define their communication topology. During inference, a fast capability
loop updates agent expertise using trajectory-level feedback, while a slow meta-LLM-driven
topology loop performs agents’ birth-death operations on MAS, including edge edit, agent
addition, and agent removal. We further show that this fast–slow design drives MAS evolution
toward a task-conditioned stable equilibrium. Experiments on four benchmarks demonstrate
that TacoMAS outperforms nearly 20 multi-agent baselines, achieving an average improvement
of 13.3% over the strongest baseline.
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2605.09539 in a model README.md to link it from this page.
Cite arxiv.org/abs/2605.09539 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2605.09539 in a Space README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.