HELENA Runs a Union of Multi-Agent Topologies but Activates Only a Sparse Subgraph, Gaining 3.47% Average and 10.34% on MMLU-Pro

arXiv 2608.04634·medium signalXBlueskyLinkedInCopy link

Most LLM multi-agent systems optimize a single topology, which narrows the reasoning trajectory, while naively merging topologies propagates redundant noise across irrelevant edges. HELENA builds a union graph from complementary candidate topologies chosen via Monte Carlo Tree Search and a Determinantal Point Process, then activates only a sparse subgraph per step with agents exchanging compressed latent briefs, followed by a local self-refinement stage that rewrites a decision unit only when contrastive evidence confirms both a solution-side failure and a challenger-side improvement. Across eight benchmarks it reports state-of-the-art on all, averaging 3.47% over the strongest baseline and up to 10.34% on MMLU-Pro, with larger gains on harder benchmarks.