TacoMAS: Test-Time Co-Evolution of Topology and Capability in LLM-based Multi-Agent Systems
Summary
This paper introduces TacoMAS, a framework for test-time co-evolution of agent capabilities and communication topology in LLM-based multi-agent systems. It demonstrates that jointly adapting fast capability loops and slow topology loops improves performance and stability over existing baselines.
View Cached Full Text
Cached at: 05/13/26, 12:14 PM
Paper page - TacoMAS: Test-Time Co-Evolution of Topology and Capability in LLM-based Multi-Agent Systems
Source: https://huggingface.co/papers/2605.09539
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.
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 effectivetest-time evolutionrequires 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 ofonline 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 usingtrajectory-level feedback, while a slowmeta-LLM-driven topologyloop performsagents’ birth-death operationson MAS, includingedge edit,agent addition, andagent removal. We further show that thisfast-slow designdrives MAS evolution toward atask-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.
View arXiv pageView PDFGitHub1Add to collection
Get this paper in your agent:
hf papers read 2605\.09539
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2605.09539 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2605.09539 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2605.09539 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
Autonomous Topology Mutation: Safe Runtime Restructuring for Multi-Agent LLM Systems with Capability, State, and Shadow Invariants
This paper introduces Autonomous Topology Mutation (ATM), a runtime mechanism for multi-agent LLM systems that detects agent overload and safely restructures team topology using formal invariants. Empirical results on DeepSeek-V3 show large improvements in code-task success rates while maintaining safety constraints.
TMAS: Scaling Test-Time Compute via Multi-Agent Synergy
TMAS introduces a multi-agent framework that enhances large language model reasoning by scaling test-time compute through structured collaboration and hierarchical memory systems. The approach uses specialized agents, cross-trajectory information flow, and hybrid reward reinforcement learning to improve iterative scaling and stability on challenging reasoning benchmarks.
A Self-Evolving Framework for Efficient Terminal Agents via Observational Context Compression
TACO introduces a self-evolving compression framework that automatically learns to shrink redundant terminal interaction history, cutting token overhead ~10% while boosting accuracy 1-4% across TerminalBench and other code-agent benchmarks.
SkillCAT: Contrastive Assessment and Topology-Aware Skill Self-Evolution for LLM Agents
SkillCAT is a training-free framework for LLM agent skill self-evolution that addresses limitations of single-trace bias, unverified merging, and full corpus loading via three stages: Contrastive Causal Extraction, Assessment-Augmented Evolution, and Topology-Aware Task Execution, achieving up to 40.40% improvement on benchmarks.
Multi-agent Framework for Time-Sensitive Complementary Collaboration in Minecraft
The paper introduces TickingCollabBench, a Minecraft-based multi-agent benchmark for time-sensitive complementary collaboration tasks with dynamic environments, and demonstrates that LLMs frequently fail under such conditions compared to a global-knowledge oracle.