[论文解读] MiTa: A Hierarchical Multi-Agent Collaboration Framework with Memory-integrated and Task Allocation

XiaoJie Zhang, JianHan Wu+2 more|arXiv (Cornell University)|Jan 30, 2026

Multi-Agent Systems and Negotiation被引用 0

一句话总结

MiTa 引入一个具有记忆的分层管理者–成员多智能体框架，配合记忆整合的情节摘要和考虑谈判的全局任务分配器，以提升协作与长期任务表现。即使成员模型较弱，也优于基线。

ABSTRACT

Recent advances in large language models (LLMs) have substantially accelerated the development of embodied agents. LLM-based multi-agent systems mitigate the inefficiency of single agents in complex tasks. However, they still suffer from issues such as memory inconsistency and agent behavioral conflicts. To address these challenges, we propose MiTa, a hierarchical memory-integrated task allocative framework to enhance collaborative efficiency. MiTa organizes agents into a manager-member hierarchy, where the manager incorporates additional allocation and summary modules that enable (1) global task allocation and (2) episodic memory integration. The allocation module enables the manager to allocate tasks from a global perspective, thereby avoiding potential inter-agent conflicts. The summary module, triggered by task progress updates, performs episodic memory integration by condensing recent collaboration history into a concise summary that preserves long-horizon context. By combining task allocation with episodic memory, MiTa attains a clearer understanding of the task and facilitates globally consistent task distribution. Experimental results confirm that MiTa achieves superior efficiency and adaptability in complex multi-agent cooperation over strong baseline methods.

研究动机与目标

• 通过解决记忆不一致和智能体之间冲突，推动多智能体系统的协同改进。
• 提出一个带有 Allocation（分配）与 Summary（摘要）模块的集中式管理者，实现全局任务分配和情节记忆整合。
• 通过情节记忆摘要实现对长期任务的跟踪，同时通过智能体谈判保持灵活性。
• 在多种大型语言模型及代理配置下，展示鲁棒性与效率提升。

提出的方法

• 将代理结构化为包含感知、记忆、谈判与执行模块的管理者与成员。
• 在管理者中增加 Allocation 与 Summary 模块，以执行全局任务分配与情节记忆整合。
• 使用谈判阶段，让每个代理通过大语言模型推理提出下一步，再由全局分配器选择一致的联合行动。
• 通过在任务进展改变时对协作历史进行摘要，实施情节记忆整合，存储简明的协作摘要以便未来规划。
• 将任务表述为带集中式管理者的多智能体部分可观测马尔可夫决策过程（MPOMDP），基于跨代理上下文分配联合行动。
• 提供一个基于伪代码驱动的记忆增强摘要过程，在进展更新时触发以保持长远上下文。

实验结果

研究问题

• RQ1 集中式管理者结合专门的分配与记忆模块，如何改善分层多智能体系统中的协作？
• RQ2 情节记忆摘要是否能保持长时间范围的上下文并降低长期任务中的信息丢失？
• RQ3 谈判驱动的提议与全局分配对不同LLM及代理配置下的任务效率有何影响？
• RQ4 相比基线，MiTa在弱代理模型与资源受限设置下的鲁棒性如何？

主要发现

Num	Method	Wash Dishes	Put Groceries	Prepare a Meal	Set up Table	Prepare Tea	Average
1	MHP	102.9	101.3	94.3	90.7	140.8	106.1
2	MHP	73.2 (+29%)	65.4 (+35%)	66.4 (+30%)	56.9 (+37%)	95.7 (+32%)	71.5 (+33%)
2	CoELA	47.3 (+54%)	42.3 (+58%)	50.7 (+46%)	47.0 (+48%)	69.1 (+51%)	51.3 (+51%)
2	ProAgent	45.5 (+55%)	54.4 (+46%)	46.4 (+51%)	54.3 (+40%)	69.2 (+51%)	53.9 (49%)
2	MiTa (ours)	51.1 (+49%)	45.4 (+55%)	46.3 (+51%)	47.0 (+48%)	54.6 (+61%)	48.8 (+54%)
3	MHP	58.2 (+43%)	55.8 (+45%)	61.3 (+35%)	48.7 (+46%)	85.6 (+39%)	61.9 (+41%)
3	CoELA	40.5 (+60%)	34.7 (+65%)	35.9 (+62%)	37.9 (+58%)	47.7 (+66%)	39.3 (+63%)
3	ProAgent	46.3 (+55%)	36.4 (+64%)	36.8 (+61%)	41.3 (+54%)	58.1 (+59%)	43.8 (+58%)
3	MiTa (ours)	39.7 (+ 61% )	30.9 ( +69% )	28.4 ( +70% )	33.2 (+ 63% )	39.7 (+ 72% )	34.4 (+ 68% )

• MiTa在复杂多智能体协作中相较强基线具有更高的效率与适应性，尤其在三个代理的情形下。
• MiTa在大多数任务中优于 CoELA 和 ProAgent，即使成员使用较弱模型也表现出近乎最优的性能。
• 用更强的背骨LLM替换后端模型，显著减少所需的步数，显示对LLM选择的鲁棒性。
• 在资源受限设置下，成员较弱时 MiTa 的步数提升很小，但当管理者较弱时协作性能下降更明显。
• 消融实验表明移除分配或摘要模块会降低性能，验证集中式协调与记忆整合的重要性。
• 三代理配置下获得最大收益，MiTa在所有五个任务中均取得最佳结果。