DeLM: Decentralized Language Models
A research framework that replaces a single central orchestrator with a shared, verified context so parallel agent threads can accumulate reusable progress on a coding benchmark.
DeLM is a research implementation for decentralized multi-agent systems that share context. Instead of one central orchestrator, it runs a configurable number of solver threads in parallel on a SWE-bench task, each with its own planner agent, one or more implementer agents, a Docker container, and local memory.
Each thread writes compact typed notes into a shared lessons store that peer threads read before planning or implementing, so threads can avoid duplicating work. The reported pass at N score is the best result among the parallel solver threads, and the average at one score is the per-thread success rate used for leaderboard comparison.
It requires Linux or another Docker-capable environment, a running Docker daemon, Python 3.11 or newer, and network access for Hugging Face datasets, SWE-bench Docker images, and LLM API calls. Model access is configured through a gitignored YAML file or environment variables.