Preprint proposes causal method to prune LLM-agent communication links

A new arXiv preprint describes E2-Explainer, a framework for identifying communication links that matter in LLM-based multi-agent systems. The method measures how masking individual channels changes task outcomes and final-response stability, then distills selected subgraphs into an explainer for deployment. The authors report that the resulting subgraphs preserved successful collaboration on reasoning and coding benchmarks while allowing redundant communication edges to be pruned. The results are preliminary and have not been independently verified.