Agents that Matter: Optimizing Multi-Agent LLMs via Removal-Based Attribution

A new formal framework treats multi-agent system attribution as a cooperative game, enabling researchers to pinpoint individual agent contributions and optimize systems at lower computational cost, according to a paper submitted on 26 May 2026 [1][2]. The work addresses a growing challenge as multi-agent systems built on large language models become more complex. Large language models are neural networks trained on vast text corpora that underpin modern chatbots and other generative AI applications [3][4]. The paper formalizes agent attribution through three parameters: the coalition distribution, the removal protocol, and the target metric [2]. Using this structure, the authors demonstrate that the Leave-One-Out method identifies bottleneck agents as effectively as more expensive combinatorial approaches, but at a fraction of the computational cost [2]. The study also shows that the choice of removal protocol produces fundamentally different games. Agent ablation, which physically removes an agent from the system, isolates structural bottlenecks. In contrast, introspective LLM judges—where a language model is asked to simulate an agent's absence—fail to faithfully approximate the behavior observed under ablation [2]. This finding carries implications for system auditing, as LLMs can generate unreliable output when trained on biased or inaccurate data [3]. To assess the value of specific underlying models, the researchers introduce attribution via model replacement. By swapping out the backbones of agents deemed low-contribution, the team improved task performance by up to 17 percent while cutting costs by up to 35 percent across three benchmarks [2]. Generative AI has seen rapid adoption across sectors including software development, healthcare, and finance since the AI boom of the 2020s, making such cost-performance trade-offs commercially significant [4]. The framework was also applied to audit a medical multi-agent system. The audit revealed that an agent's contribution to diagnostic accuracy is often decoupled from its contribution to ethical behavior. By intervening on agents playing counterproductive roles, the researchers increased ethics alignment while maintaining diagnostic accuracy [2]. The field of artificial intelligence, which dates to 1956, has long grappled with ethical concerns and safety, and these issues have intensified with the rise of generative AI [5]. The paper provides a principled method for cost-effective attribution and targeted intervention in multi-agent systems [2].