Agents that Matter: Optimizing Multi-Agent LLMs via Removal-Based Attribution
Proposes removal-based attribution framework for multi-agent LLMs; LOO achieves 7.5× efficiency, boosting task performance by 17%.
Key Findings
Methodology
The paper formalizes agent attribution as a cooperative game parameterized by coalition distribution, removal protocol, and target metric. Using LOO, it reduces computational costs by up to 7.5× while accurately identifying bottleneck agents. Introduces model replacement protocol to evaluate agent backbone contributions, improving task performance by up to 17% and reducing costs by 35%. Experiments span code generation and medical diagnosis, demonstrating broad applicability. Different removal protocols (agent ablation, introspective judgment, model replacement) show distinct behaviors across communication topologies, emphasizing protocol choice based on task goals and budget.
Key Results
- LOO matches the performance of combinatorial methods in identifying bottleneck agents but at a fraction of the cost, reducing token usage by roughly 7.6× across datasets. Model replacement boosts task success by up to 17% and cuts costs by 35%. In medical systems, it uncovers decoupling between diagnostic accuracy and ethical behavior, enabling targeted interventions that enhance ethics while maintaining performance.
- Across communication topologies (centralized, decentralized, hybrid), attribution results vary significantly, with agent ablation emphasizing structural importance, and model replacement focusing on capability. LOO consistently offers a favorable efficiency-effectiveness balance, making it the recommended attribution approach.
- Cost analysis shows LOO's token usage is substantially lower than Shapley/Owen and Myerson values, making it suitable for large-scale systems. Experimental results confirm that attribution-guided model replacement reduces inference costs and improves success rates, especially under resource constraints.
Significance
This work provides a rigorous theoretical foundation and practical tools for optimizing multi-agent systems, addressing the high computational cost and interpretability issues of previous methods. By formalizing attribution as a cooperative game, it clarifies the impact of different protocols and coalition choices, enabling more effective debugging, system tuning, and ethical auditing. Its broad applicability in software engineering, healthcare, and beyond paves the way for safer, more efficient AI systems, fostering trust and transparency in complex multi-agent deployments.
Technical Contribution
The paper introduces a unified cooperative game framework for agent attribution, explicitly modeling coalition distribution, removal protocols, and behavior metrics. It demonstrates that LOO offers a near-optimal balance of efficiency and accuracy, outperforming combinatorial methods. The novel model replacement protocol extends attribution beyond simple ablation, allowing capability assessment and cost-effective intervention. Theoretical analysis and extensive experiments across diverse topologies and tasks validate these contributions, establishing a new standard for scalable, interpretable multi-agent attribution.
Novelty
This is the first systematic formalization of multi-agent attribution as a cooperative game with explicit protocol parameterization. The introduction of model replacement as a non-destructive, capability-focused protocol is innovative, enabling nuanced evaluation of agent contributions. The comprehensive comparison of removal protocols and coalition distributions across communication topologies provides new insights into attribution behavior, setting a foundation for future scalable, interpretable multi-agent system analysis.
Limitations
- The current framework assumes static communication topologies; dynamic or evolving networks require further validation. Model replacement relies on pre-trained models, which may be limited by their quality and availability. Attribution of complex ethical behaviors remains challenging, needing integration with causal inference and multi-modal data. Computational costs, while reduced, still pose challenges for extremely large systems, suggesting ongoing optimization is necessary.
Future Work
Future research will explore adaptive attribution methods for dynamic topologies, integrating causal inference to improve ethical behavior attribution. Combining multi-modal data and reinforcement learning could enhance interpretability and robustness. Extending the framework to real-time, large-scale systems and exploring automated intervention strategies are promising directions to further advance multi-agent AI deployment.
AI Executive Summary
As multi-agent systems (MAS) grow in complexity, understanding individual agent contributions becomes crucial for system optimization. Traditional attribution methods like Shapley values, while theoretically sound, are computationally prohibitive for large systems. This paper introduces a novel framework that formalizes agent attribution as a cooperative game parameterized by coalition distribution, removal protocol, and target metric. The core idea leverages Leave-One-Out (LOO), which isolates bottleneck agents efficiently—reducing costs by up to 7.5×—while maintaining high accuracy in identifying critical agents.
The authors further propose a model replacement protocol, substituting low-contribution models with alternatives to evaluate backbone capabilities. This approach improves task performance by up to 17% and reduces costs by 35%, demonstrating significant practical benefits. Extensive experiments across code generation and medical diagnosis benchmarks validate the framework's effectiveness, showing that different communication topologies (centralized, decentralized, hybrid) influence attribution results and intervention strategies.
One key insight is that different removal protocols—agent ablation, introspective judgment, and model replacement—answer distinct attribution questions and exhibit varying behaviors depending on system structure. The study highlights that agent ablation emphasizes structural importance, often causing system failures when removing essential roles, while model replacement provides a capability-oriented perspective, enabling cost-effective upgrades.
The framework's versatility extends to real-world applications, such as identifying non-contributing agents in medical diagnosis, where targeted interventions can improve ethical alignment without sacrificing accuracy. Overall, this work offers a scalable, principled approach to MAS attribution, balancing computational efficiency, interpretability, and practical utility. Future directions include dynamic topology handling, causal inference integration, and real-time system optimization, promising to advance the deployment of trustworthy multi-agent AI systems.
Deep Analysis
Background
多智能体系统（MAS）在自动化、机器人、软件工程等领域得到广泛应用。早期研究主要关注单一模型优化，代表性工作如Shapley值和Owen值用于贡献评估，但计算复杂度高，难以扩展。近年来，LLMs的崛起推动了多智能体交互的研究，强调系统的协作与调优。现有方法多依赖全局或实例级归因，缺乏统一理论框架，且在复杂场景中效率不足。随着系统规模扩大，识别关键代理、优化系统性能成为亟待解决的问题。
Core Problem
核心挑战在于如何高效、准确地评估每个代理对整体任务的贡献，尤其在多代理交互复杂、角色多样的情况下。传统方法如Shapley值虽理论完备，但计算成本随代理数指数增长，难以应用于大规模系统。此外，不同通信拓扑和角色分配影响归因结果，缺乏统一分析框架，限制了系统调优的效果。如何在保证效率的同时，提供可信的贡献评估，是当前亟待突破的难题。
Innovation
提出合作博弈的归因框架，将归因参数化为联盟分布、移除协议和目标指标，统一分析不同方法。引入LOO方法，显著降低计算成本（低至7倍），同时保持识别瓶颈代理的准确性。模型替换协议创新性地用不同模型版本评估代理能力，兼顾成本与性能。系统比较不同通信拓扑下的归因差异，为多场景应用提供理论依据。这些创新极大丰富了多智能体归因的理论体系。
Methodology
- �� 定义多智能体系统为合作博弈，参数包括代理集、架构和效用函数。• 采用联盟分布、移除协议（代理消融、内省判定、模型替换）和目标指标（任务成功、成本）三要素，参数化归因问题。• 利用LOO方法，将归因转化为单一代理移除，降低计算复杂度。• 引入模型替换协议，通过替换模型评估能力贡献。• 比较不同通信拓扑（集中、分散、混合）下的归因效果，分析协议差异。• 设计多场景实验，验证方法在代码生成和医疗诊断中的有效性。
Experiments
采用PlanCraft、WorkBench和BrowseComp-Plus三大基准，模拟不同通信拓扑（独立、分散、集中、混合），配置不同角色模型。评估指标包括AUC、任务成功率和成本。对比LOO、Shapley、Owen和Myerson值的归因效果，分析不同协议的差异。通过消融实验验证模型替换的性能提升，观察归因对干预策略的指导作用。多轮实验确保结果的稳健性和可重复性。
Results
LOO在所有场景中表现出与组合方法相当的归因效果，但计算成本低至7倍，显著节省资源。模型替换在三项任务中最高提升17%的任务性能，成本降低35%。在医疗系统中，识别出诊断与伦理行为的解耦，干预低效角色后伦理得分提升，同时保持诊断准确性。不同通信拓扑下，归因结果差异明显，强调协议选择的重要性。整体上，LOO成为效率与效果兼备的首选方案。
Applications
该框架可广泛应用于软件工程、医疗、金融等领域的多智能体系统调优。通过高效识别关键代理，优化系统性能和伦理表现，降低成本。未来结合动态拓扑和因果推断，可实现更复杂场景的实时调优，推动AI在实际应用中的安全性和可靠性。
Limitations & Outlook
目前方法主要适用于静态通信拓扑，动态变化场景尚未充分验证。模型替换依赖预训练模型，受限于模型质量和可用性。对复杂伦理行为的归因仍存在挑战，未来需结合多模态数据和因果推断技术提升解释性。计算成本虽降低，但在超大规模系统中仍有优化空间。
Plain Language Accessible to non-experts
想象一个工厂里有很多工人，每个人负责不同的任务。有时候，工厂出问题了，不知道哪个工人出了差错。为了找出问题所在，工厂经理会暂时让一些工人休息，看看问题是否解决。这个过程就像我们用“移除”某个工人，观察工厂的表现变化。通过这种方法，我们可以知道哪个工人对工厂的正常运转最重要。这个方法可以帮我们更好地理解每个工人在整个系统中的作用，就像找出工厂中最关键的工人一样。
ELI14 Explained like you're 14
想象你在学校里有一组朋友一起完成一个大项目。每个人都负责不同的部分，比如画画、写字、做模型。有时候，项目出了问题，你会想知道哪个朋友的部分最重要。你可以试着让某个朋友暂时不帮忙，看看项目还能不能完成。如果还能，那这个朋友可能不是那么关键；如果不能，那他就是关键人物。这就像我们用“移除”朋友的方法，来找出谁对项目最重要。这样一来，你就知道该多花时间帮忙那些真正关键的朋友啦！
Glossary
合作博弈 (Cooperative Game)
一种分析多个代理合作贡献的模型，强调联盟和合作关系，帮助分配整体效用。
本文将归因问题形式化为合作博弈，定义联盟和贡献指标。
Leave-One-Out (LOO)
通过逐个移除代理，评估其对系统性能的边际贡献，计算简单且效率高。
作者采用LOO作为核心归因工具，显著降低计算成本。
模型替换 (Model Replacement)
用不同能力的模型替代原模型，以评估其在系统中的贡献和潜在升级空间。
引入模型替换协议，提升任务性能，降低成本。
通信拓扑 (Communication Topology)
描述代理间信息交流结构的图模型，包括集中式、分散式和混合式拓扑。
不同拓扑影响归因结果和干预策略。
联盟分布 (Coalition Distribution)
定义在合作博弈中不同代理子集的概率分布，用于计算边际贡献。
本文分析了LOO、Shapley、 Owen等不同联盟分布的效果。
Open Questions Unanswered questions from this research
- 1 如何在动态变化的通信拓扑中保持归因的准确性仍未解决，尤其在实时系统中，模型的稳定性和解释性需要进一步研究。
- 2 现有方法对复杂伦理行为的归因仍有限，未来需结合多模态数据和因果推断技术提升解释能力。
Abstract
As multi-agent systems (MAS) become increasingly complex, identifying the contributions of individual agents is critical for system optimization. However, existing approaches lack a rigorous, unified framework for credit assignment. In this work, we formalize agent attribution as a cooperative game, parameterized by the coalition distribution, removal protocol, and target metric. Using this framework, we show that Leave-One-Out (LOO) identifies bottleneck agents as effectively as combinatorial methods, but at a fraction of the computational cost. We also demonstrate that removal protocols induce distinct games: Agent ablation isolates structural bottlenecks, whereas introspective LLM judges fail to faithfully approximate this behavior. Furthermore, to evaluate the utility of specific agent backbones, we introduce attribution via model replacement. By substituting underlying models of low-contribution agents, we improve task performance by up to 17% while reducing cost by up to 35% across three benchmarks. Finally, we apply our framework to audit a medical MAS, revealing that agent contributions to diagnostic accuracy and ethical behavior are often decoupled. By intervening on counterproductive roles, we observe an increase in ethics alignment while maintaining diagnostic accuracy. Overall, this work provides a principled approach for cost-effective MAS attribution and intervention.