E2-Explainer uses causal inference to prune LLM agent communication costs

E2-Explainer, a new framework from Junzhi Li and colleagues, tackles a key problem in LLM-based multi-agent systems (MAS): communication topologies are usually learned via black-box optimization, offering no insight into why specific agent-to-agent edges are chosen. The paper, submitted to AAAI 2027 and posted on arXiv (2608.12921), formulates topology explanation as a causal attribution problem. It identifies compact communication subgraphs using edge-level evidence, measured by a Granger-style objective that masks each channel to observe changes in task outcomes and response stability. This evidence is then distilled into an amortized explainer, allowing fast post-hoc explanations without repeated edge evaluations.

Experiments across multiple reasoning and coding benchmarks show E2-Explainer not only identifies critical subgraphs that preserve collaboration, but also allows direct execution of these subgraphs to prune redundant edges. This substantially reduces communication overhead while keeping task performance competitive. For practitioners building multi-agent pipelines, this means lower latency, reduced token costs, and a clearer understanding of which agent conversations are essential—a step toward more efficient and auditable AI systems.