Dynamic capability-aware MPNNs

Develop message-passing neural networks that dynamically adapt message-passing depth, embedding dimensions, or aggregation responsibilities to runtime changes in agents’ computational power, memory, and bandwidth.

Background

Current asynchronous MPNN methods generally assume homogeneous agents with static capabilities. Real deployments experience changing workloads and network conditions, requiring coordination of computational-tree traversal among agents operating at different speeds. The survey identifies runtime capability-aware adaptation as unresolved.

References

While prior work balances static capability constraints~\citep{liu2021glint,chen2022graph,zeng2022gnn}, dynamically adapting message-passing depth, embedding dimensions, or aggregation responsibilities to runtime capability changes remains unexplored.

— From Euclidean to Graph-Structured Data: A Survey of Collaborative Learning  (2609.02984 - Bourgerie et al., 2 Sep 2026) in Section 6.2, paragraph “Open challenges”