AGTBRIEF
Researchers propose E2 Explainer, an after the fact analysis tool that identifies which message channels between collaborating large language model (LLM) agents carry task signal.
When a team of large language models (LLMs) tackles a coding or reasoning problem, the agents can pass hundreds of messages back and forth, and only some of those messages move the answer forward. A preprint posted Thursday on arXiv proposes a way to ask, after the fact, which edges in that communication graph actually matter.
The framework, called E2-Explainer, treats each message channel as a candidate cause and tests it with a Granger-style procedure: mask one edge, then measure both the change in task accuracy and the shift in the agents' result-level semantic entropy. Edges whose removal tanks accuracy are flagged as load-bearing; edges that drop entropy without hurting the answer are doing real collaboration work. The survivors are distilled into a compact, budgeted subgraph scoreable in a single forward pass at deployment, as described in the paper.
The paper applies the method to topology generators like G-Designer, ARG-Designer, AgentDropout, and AgentPrune, and tests it on reasoning and coding benchmarks including MMLU and HumanEval. Their Figure 1 experiment, reproduced in the HTML version, masks 5% to 80% of edges on G-Designer graphs across 10 seeds: most random pruning degrades accuracy, while only a handful of masks preserve or improve it. That pattern is evidence that load-bearing edges exist but are not reachable by chance.
The caveat: this is a post-hoc explanation, not a preventive control, and the gains are benchmark-scoped, not from a deployed agent product. The 18-page preprint, with 4 figures, has been submitted to AAAI 2027 and is not yet peer-reviewed.