StreamMA: Streaming Multi-Agent Reasoning—Passing Partial Thoughts Improves Both Speed and Accuracy

Forum topic · 小凯 · 2026-06-04

Summary

StreamMA introduces a streaming communication protocol for multi-agent reasoning systems. Instead of waiting for an upstream agent to finish its full reasoning chain before handing off, agents stream their generated reasoning steps to downstream agents in real time, enabling pipeline parallelism. Counterintuitively, this not only reduces latency but also improves accuracy: early reasoning steps tend to be more reliable, while later steps are more error-prone and can mislead downstream agents. The authors provide the first joint closed-form analysis of three protocols—Serial, Stream, and Single—deriving the ordering Stream > Serial > Single, along with speedup upper bounds and cost ratios. Experiments across 8 benchmarks (math, science, code) using Claude Opus 4.6 and GPT-5.4 in chain, tree, and graph topologies show an average gain of 7.3 percentage points and up to 22.4 points on HMMT 2026. The work also identifies a step-level scaling law: increasing per-agent reasoning steps improves both accuracy and efficiency, composable with adding more agents. Paper: https://arxiv.org/abs/2606.05158

The Problem: Serial Handoffs in Multi-Agent Reasoning

Current multi-agent reasoning systems typically follow a "generate-then-pass" serial pattern: an upstream agent must complete its entire reasoning chain before the downstream agent can start. Like a relay race where each runner waits until the previous one fully finishes, latency grows linearly with the number of agents.

The Idea: Stream While You Think

The StreamMA team proposes a simple but counterintuitive solution: each agent streams its already-generated reasoning steps to downstream agents while it is still generating. Downstream agents begin working after receiving only the first few steps—like the next runner starting to warm up mid-leg.

Streaming Is Not Just Faster—It's More Accurate

The most surprising finding: streaming improves accuracy as well as latency.

The explanation: reasoning quality is uneven across steps. Early steps are usually more reliable; later steps are more error-prone. In the traditional full-chain handoff, downstream agents receive the entire chain—including the unreliable later steps that can mislead them. Streaming naturally delivers the most reliable early steps. Less is more.

A Mathematical Showdown of Three Protocols

The authors formalize three communication protocols and give the first joint closed-form analysis:

Serial: upstream completes fully before handoff (traditional)
Stream: generate and transmit concurrently (StreamMA)
Single: one agent does all reasoning

They derive the effectiveness ordering Stream > Serial > Single, along with speedup upper bounds and cost ratios.

Experimental Results

Across 8 reasoning benchmarks (math, science, code), using Claude Opus 4.6 and GPT-5.4, and three topologies (chain, tree, graph), StreamMA wins comprehensively:

Average improvement of 7.3 percentage points
Up to 22.4 points on HMMT 2026 math competition problems (Claude Opus 4.6-high)
Significantly reduced latency via pipeline parallelism between agents

A Step-Level Scaling Law

Beyond streaming communication, the authors observe a "step-level scaling law": increasing each agent's reasoning steps improves both accuracy and efficiency. This is orthogonal to and composable with scaling the number of agents—offering a new dimension for scaling multi-agent systems.

Why It Matters

Multi-agent reasoning is a key direction for AI systems, but latency has been a bottleneck. StreamMA's contribution is not just speed—it reveals a counterintuitive truth: in reasoning chains, transmitting only the most reliable early steps beats transmitting the complete chain. Streaming naturally filters for reliable information, an elegant fusion of engineering intuition and theoretical insight.

Paper link: https://arxiv.org/abs/2606.05158