Existing multi-agent systems (MAS) for large language model reasoning typically rely on a central controller that assigns tasks and merges results, creating a bottleneck as subtasks increase. This paper introduces DeLM, a decentralized framework where agents asynchronously claim subtasks from a queue, read a shared verified context, perform local reasoning, and write back compact updates. This design eliminates the need for all communication to pass through a central point.
DeLM achieves state-of-the-art results on SWE-bench Verified, with gains of up to 10.5 percentage points over the strongest baseline while reducing cost per task by roughly 50%. On LongBench-v2 Multi-Doc QA, it improves accuracy by up to 5.7 percentage points across four frontier model families. These results demonstrate that decentralizing coordination can improve both performance and efficiency in test-time scaling and long-context reasoning tasks.