DeLM outperforms Claude Code and Codex on coding tasks, achieving up to 2.49× faster execution
Summary
A new paper from Stanford has introduced Decentralized Language Models (DeLM), a multi-agent system that outperforms Claude Code and Codex in complex coding tasks, achieving speeds up to 2.49 times faster and increasing accuracy by up to 19.2 percentage points on long-horizon benchmarks. This framework addresses inefficiencies in traditional multi-agent systems by utilizing shared context and task queues, enabling agents to work independently and collaborate more effectively without a central coordinator. The open-source DeLM code, along with trajectories and an integration plugin for existing coding models, has also been released, promoting wider accessibility and experimentation in AI research.
Analysis
DeLM: DeLM is a coordination framework for multi-agent systems built on top of existing LLM agent harnesses. It enables asynchronous agent collaboration through a shared context log and task queue without a central orchestrating agent. In this news, DeLM is presented as a new approach that improves speed and accuracy on long-horizon coding tasks by reducing redundant work and waiting among agents. Codex: Codex is an AI model for code generation and agentic tasks. It serves as a primary baseline in evaluations of multi-agent coding systems. The news highlights DeLM's performance gains over both vanilla Codex and its native subagent configurations on multiple benchmarks. Jerry Gu: Jerry Gu is a co-author of the DeLM paper on decentralized multi-agent systems with shared context. He contributed to the research demonstrating improved agent collaboration on coding benchmarks. Yuzhen Mao: Yuzhen Mao is a researcher and the lead author of the arXiv paper introducing DeLM. His work centers on decentralized coordination mechanisms for scaling LLM agents on complex tasks. Claude Code: Claude Code is Anthropic's coding assistant and CLI tool designed for agentic workflows. It functions as a key baseline harness in the reported experiments. The news shows DeLM delivering faster execution and higher accuracy than Claude Code setups with and without subagents. Hangoo Kang: Hangoo Kang is a co-author of the paper on Decentralized Language Models. His contributions support the evaluation of DeLM across multiple agentic coding benchmarks. Aadi Chauhan: Aadi Chauhan is a co-author of the paper presenting the DeLM framework. His involvement focuses on advancing asynchronous multi-agent techniques for long-horizon software engineering problems. Qizheng Zhang: Qizheng Zhang is a co-author of the DeLM research. He helped develop the shared context and task queue approach that eliminates coordination bubbles in agent teams. Azalia Mirhoseini: Azalia Mirhoseini is a co-author of the DeLM paper. She participated in the work showing how decentralized agent coordination can outperform centralized and synchronous baselines. AI Research: New decentralized frameworks are being developed to improve efficiency in multi-agent LLM systems for complex tasks. Open Source: The DeLM code, trajectories, and an integration plugin are publicly released for use with existing coding models. Agent Coordination: Shared context and task queues allow agents to build directly on each other's progress without a central planner.
Categories
aiai_agentsmachine_learningtechvirtuals