DeLM uses a shared knowledge base as a common communication substrate and a task queue, letting parallel agents read verified progress, failures, and evidence without routing every interaction through a central agent.

Stanford researchers introduce DeLM, a decentralized language model framework that coordinates multiple agents without a central controller, aiming to reduce inference costs and coordination latency.

Traditional centralized multi-agent systems suffer bottlenecks, information dilution, and repeated interactions as subtasks scale, leading to slower coordination and higher costs.

The framework shows particular promise for software engineering test-time scaling, long-context reasoning, and multi-document question-answering, where parallel hypotheses can be explored with coordinated progress.

DeLM challenges the assumption that every multi-agent workflow needs a central controller, presenting a decentralized approach that is reportedly faster, more accurate, and roughly half the cost based on benchmarks.

The DeLM workflow proceeds from initialization to parallel execution, then compression and verification into reusable gists, optional unfolding for detail, and final consolidation by the last-returning agent.

DeLM’s shared context stores gists—verified findings, partial results, and documented failures—allowing agents to build on prior discoveries while preserving evidence for needed revisits.

Key advantages include avoiding redundant exploration, faster adaptation as subtasks scale, and robust coordination through shared constraints and failures, with compact progress kept via unfoldable gists.

Empirical results on SWE-bench Verified show DeLM outperformed the strongest baseline by about 10.5% and reduced cost per task by roughly 50%; on LongBench-v2 Multi-Doc QA it achieved top accuracy across four model families, including GPT-5.4 and Claude Sonnet.

The approach’s benefits contrast with centralized systems by reducing bottlenecks, information dilution, and repeated interactions as tasks scale.