StreamMA: Making Multi-Agent Systems Faster and More Accurate!

Hey everyone! Our team just released StreamMA. It is a new way to make multi-agent systems both faster and more accurate. The core idea is simple but powerful: Instead of waiting for one agent to finish its full response, we now send information step by step. Each reasoning step goes to the next agent right away.

Why does this work so well? The upstream agent sends one step, and the downstream agents start working immediately. This makes everything much faster. Surprisingly, the reasoning quality also gets better. Complex reasoning steps are not all the same quality. Early steps are usually reliable and clear. Later steps are more likely to have errors. StreamMA lets downstream agents start with the good early steps. By the time any mistakes arrive, the other agents have already built their own strong reasoning path. This naturally reduces the impact of errors.

Key Contributions: ① Stream Protocol: We change communication from full responses to single steps. This improves both speed and performance. ② Three closed-form theorems: We show when Stream works best, the theoretical speedup limit, and the cost ratio. ③ Step-level scaling law: With the same number of agents, adding more steps improves both accuracy and speed. It works well together with traditional agent scaling.

Strong Experimental Results: • Performance: Average +7.3 percentage points better on 8 math, science, and code benchmarks (Opus 4.6) • Speedup: With 64 agents and 64 steps, we get 26.9× faster inference (theoretical max is 32.3×, we reached 83% of it) (HMMT26 GPT5.4)

Want to learn more? Paper: https://huggingface.co/papers/2606.05158… Project Page: https://zhenyangcs.github.io/StreamMA-website… GitHub: https://github.com/EnVision-Research/StreamMA…

If you work on multi-agent systems, distributed reasoning, or long-chain inference, we would love to hear your thoughts! This streaming idea brings new possibilities. #StreamMA #MultiAgent #AI #Reasoning #LLM

Paper page - Streaming Communication in Multi-Agent Reasoning

Source: https://huggingface.co/papers/2606.05158

Abstract

StreamMA enables efficient multi-agent reasoning by streaming intermediate results and leveraging reliable early steps to improve both latency and effectiveness across various reasoning tasks.

Multi-agent reasoning systemsadopt a “generate-then-transfer” paradigm that forcesend-to-end latencyto scale linearly with pipeline depth. We introduce StreamMA, a multi-agent reasoning system that streams each reasoning step to downstream agents as soon as it is generated,pipeliningadjacent agents and thus reducing latency. Surprisingly, thispipeliningalso improves effectiveness: because multi-step reasoning quality is non-uniform and early steps are more reliable than later ones, working with these reliable early steps instead of the fullchainprevents error-prone late steps from misleading downstream agents. We formalize both advantages with the first closed-form joint analysis of stream, serial, andsingle protocols, deriving theeffectiveness ordering,speedup upper bound, andcost ratio. Across eightreasoning benchmarksspanning mathematics, science, and code, two frontierLLMs(Claude Opus 4.6 and GPT-5.4), and three topologies (Chain,Tree,Graph), StreamMA outperforms both baselines (avg. +7.3 pp, max +22.4 pp on HMMT 2026; Claude Opus 4.6-high). Beyond these contributions, we discover a “step-level scaling law”: increasing per-agent steps consistently improves both effectiveness and efficiency, a new scaling dimension orthogonal to and composable withagent-count scaling.

View arXiv pageView PDFProject pageGitHub25Add to collection

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2606.05158 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2606.05158 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2606.05158 in a Space README.md to link it from this page.

Collections including this paper1