[논문 리뷰] Self-Resource Allocation in Multi-Agent LLM Systems

Alfonso Amayuelas, Junjie Yang외 5명|ArXiv.org|2025. 04. 02.

Scheduling and Optimization Algorithms인용 수 3

한 줄 요약

다중 에이전트 시스템에서 LLM이 다수의 에이전트 워커 간 작업을 배분하는 오케스트레이터나 플래너로 작동하는 방법을 평가하고, Hungarian algorithm 베이스라인 및CuisineWorld 벤치마크와 비교하며, 작업자 역량이 계획 효율성에 미치는 영향을 분석한다.

ABSTRACT

With the development of LLMs as agents, there is a growing interest in connecting multiple agents into multi-agent systems to solve tasks concurrently, focusing on their role in task assignment and coordination. This paper explores how LLMs can effectively allocate computational tasks among multiple agents, considering factors such as cost, efficiency, and performance. In this work, we address key questions, including the effectiveness of LLMs as orchestrators and planners, comparing their effectiveness in task assignment and coordination. Our experiments demonstrate that LLMs can achieve high validity and accuracy in resource allocation tasks. We find that the planner method outperforms the orchestrator method in handling concurrent actions, resulting in improved efficiency and better utilization of agents. Additionally, we show that providing explicit information about worker capabilities enhances the allocation strategies of planners, particularly when dealing with suboptimal workers.

연구 동기 및 목표

• Assess how LLMs can allocate computational tasks among multiple agents under cost, efficiency, and performance constraints.
• Compare LLM-based orchestrator and planner approaches against optimal Hungarian algorithm solutions.
• Evaluate resource allocation in dynamic, multi-task environments (CuisineWorld) with delayed rewards.
• Analyze how worker capabilities influence planner effectiveness and how explicit capability information affects results.

제안 방법

• Formalize multi-agent resource allocation as a combinatorial optimization problem with costs, capabilities, and tasks.
• Compare three experimental architectures: Individual, Orchestrator (centralized), and Planner (semi-decentralized).
• Use Hungarian algorithm as ground truth for single-turn assignment problems.
• Evaluate on CuisineWorld-based concurrent task scenarios with delayed rewards.
• Assess capability-aware allocation by varying worker model backbones and providing explicit capability hints.

실험 결과

연구 질문

• RQ1 Can LLMs generate near-optimal task allocations in a standard assignment problem compared to the Hungarian algorithm?
• RQ2 How do orchestrator and planner approaches compare in handling concurrent actions and resource utilization in multi-agent tasks?
• RQ3 Does providing explicit worker capability information improve planner-based allocation, especially with suboptimal workers?
• RQ4 How does planner-based allocation perform under heterogeneous worker capabilities within CuisineWorld?

주요 결과

• LLMs achieve higher validity and accuracy with larger models, but at substantial computational and monetary cost.
• Planner method outperforms orchestrator in handling concurrent actions and improves efficiency and agent utilization.
• Providing explicit worker capability information enhances planner allocations, particularly with suboptimal workers.
• Planner achieves higher efficiency (completed orders per cost) and fewer idle actions compared to orchestrator and individual approaches.
• Resource awareness reveals that heterogeneous teams can outperform homogeneous ones when at least one strong model is included.