Streaming Communication in Multi-Agent Reasoning

Streaming Communication in Multi-Agent Reasoning introduces STREAM MA, a multi-agent protocol from HKUST(GZ), Alibaba Group, and Zhejiang University that replaces the usual generate-then-transfer rule with step-level forwarding: each reasoning step is streamed downstream immediately, allowing adjacent agents to overlap execution instead of waiting for a full response. The paper gives a closed-form analysis of three execution modes—Single, Serial, and Stream—showing that Stream is optimal when early steps are more reliable than later ones, while also deriving a speedup upper bound and an exact cost ratio under LLM serving assumptions. Empirically, across eight benchmarks spanning AIME 2025, AIME 2026, HMMT 2026, GPQA-Diamond, HLE, and LiveCodeBench, using Claude Opus 4.6 and GPT-5.4 over Chain, Tree, and Graph topologies, STREAM MA improves average accuracy by +7.3 pp on Claude and +1.5 pp on GPT versus Serial, with a peak gain of +22.4 pp on HMMT 2026. The study also reports a step-level scaling law: at A=64, increasing per-agent steps S from auto-decided length to 64 raises accuracy from 68.2% to 73.5% and yields 26.9× wall-clock speedup, reaching 83% of the theoretical 32.3× bound. With A=S=4 on Claude Opus 4.6, the protocol achieves about 2.30× speedup, and under full prefix caching it is 7.5% cheaper than Serial, while the no-cache regime reverses that advantage.