Graph-based multi-agents smash long-context benchmarks–89% MMLU-Pro on 8B models (opens in new tab)

Dynamic graph structures for multi-agent collab—agents as nodes on data chunks, edges as comms. Crushes long-context (2K tokens vs. 128K) and hits 89% on MMLU-Pro with fewer agents than fixed teams. Model-agnostic, emergent smarts from graph topology. ICML paper, no hype, pure efficiency hack. Paper: arXiv:2509.21848 ICML: openreview.net/forum?id=34cANdsHKV