DeLM framework decentralized multi-agent coordination cuts costs 50%, boosts reasoning

The paper, published on arXiv on June 9, 2026, tackles a fundamental bottleneck in multi-agent LLM systems: centralized control. Existing architectures rely on a main agent to assign tasks, collect outputs, and merge results, which becomes a communication and integration bottleneck as subtask count grows. DeLM eliminates this by introducing a shared verified context that acts as a common communication substrate. Agents claim subtasks from a queue, read accumulated progress, perform local reasoning, and write back verified updates without routing through a central controller. This asynchronous, decentralized design enables agents to build on each other's verified work in parallel, dramatically reducing overhead.

Empirically, the results are striking. On SWE-bench Verified (a software-engineering test-time scaling benchmark), DeLM achieves top performance across all evaluated metrics—Average@1, Pass@2, and Pass@4—with gains up to 10.5 percentage points over the strongest baseline (which uses a centralized orchestrator). Even more impressive, it reduces cost per task by roughly 50%. On LongBench-v2 Multi-Doc QA, DeLM achieves the highest average accuracy across four frontier model families (including GPT-4, Claude, Gemini, and Llama), improving over the strongest baseline by up to 5.7 percentage points. The code is open-sourced on the project website, making this approach immediately accessible for production use.