StreamMA: Multi-agent reasoning doesn't need to wait — streaming as you think is actually more accurate.
Have you experienced this: three people on a team complete a report in relay — the second person can only start after the first finishes entirely, and the third waits for the second. Total time is the sum of everyone's time. This is the plight of current multi-agent reasoning systems: the "generate-then-pass" serial mode makes latency grow linearly with the number of agents.
But wait — why not stream while writing?
Researchers from the StreamMA team propose a seemingly simple yet counterintuitive solution: each agent streams its already-generated reasoning steps to downstream agents while still generating. Downstream agents don't wait for upstream completion; they start working from the first few steps. Like a relay race where the next runner starts warming up while the previous one is only halfway through their leg.
Streaming: Not just faster, but more accurate
The most surprising finding: streaming transmission not only reduces latency but also improves reasoning accuracy.
Why? The researchers offer a sharp explanation: the quality of multi-step reasoning is uneven. Early steps tend to be more reliable; later steps are more error-prone. In the traditional "generate-then-pass" mode, downstream agents receive the complete reasoning chain — including those unreliable later steps, which can mislead them.
In streaming mode, downstream agents receive the upstream agent's early steps — exactly the most reliable part. It's like asking a colleague for advice: their first few suggestions are usually the soundest, and the later ones drift. Listening to just the first few is actually better.
A mathematical showdown of three protocols
The researchers don't just run experiments — they provide rigorous mathematical analysis. They define three communication protocols:
- Serial protocol: upstream fully completes before passing downstream (traditional)
- Stream protocol: transmit while generating (StreamMA's approach)
- Single-agent protocol: one agent does all reasoning
- Average improvement of 7.3 percentage points
- Up to 22.4 percentage points on HMMT 2026 math competition problems (with Claude Opus 4.6-high)
- Significantly reduced latency via pipeline parallelism between agents
They derive the effectiveness ordering: Stream > Serial > Single, and give upper bounds on speedup and cost ratios. This is the first work to provide a joint closed-form analysis of these three protocols.
Experimental results: comprehensive wins
Across 8 reasoning benchmarks (math, science, code), using Claude Opus 4.6 and GPT-5.4, with three topologies (chain, tree, graph), StreamMA wins across the board:
Step-level scaling law: a new dimension of scaling
Beyond streaming communication, the researchers found an interesting phenomenon: increasing the number of reasoning steps per agent improves both accuracy and efficiency. They call it the "step-level scaling law."
This is orthogonal to and composable with traditional agent-count scaling. You can increase both the number of agents and the steps per agent for a double benefit — a new dimension for scaling multi-agent systems.
Why this matters
Multi-agent reasoning is a key direction for AI systems, but latency has been a bottleneck. StreamMA's contribution isn't just "faster" — it reveals a counterintuitive truth: in reasoning chains, less is more. Passing only the most reliable early steps beats passing the full chain.
It echoes an old saying: talk too much and you'll slip. In multi-agent communication, more information isn't necessarily better — what matters is whether the information passed is reliable. StreamMA uses streaming to naturally filter for the most reliable information — an elegant combination of engineering intuition and theoretical insight.
Paper link: https://arxiv.org/abs/2606.05158