Multi-Agent Relay Thinking

Multi-Agent Relay Thinking. By Xu Liu and four others, from Zhejiang University, Nanyang Technological University, and one other institution. Published on the twenty-seventh of October.

Most methods for multi-agent collaboration have explored one arrangement, which the authors call group thinking. Each agent produces its own solution independently, and the outputs are then aggregated. The agents work side by side, not in sequence, and none of them builds directly on another's reasoning as it unfolds. Whatever one agent works out along the way stays with that agent, and only the finished answers meet at the end.

This work investigates a different topology: relay thinking. Here ownership of the reasoning is handed from one agent to the next, dynamically. A later agent inherits the reasoning trajectory so far and refines it, so the trajectory evolves as it passes along, much like a baton in a relay race. The framework that makes this possible is called Multi-Agent Confidence-guided Relay, or MACR.

A relay raises two questions. When should the reasoning be passed on, and which agent should take it next? MACR answers both with confidence. Relay is triggered by confidence, and the next agent is chosen through a bidding mechanism, where each candidate's bid is based on its own confidence. The same signal decides both the timing of a handoff and the agent who receives it.

The evaluation covers four benchmarks. Against the strongest single-model strategies, MACR is more accurate, for example by three point three percent on MathVista and three point nine percent on MMStar. Against multi-agent strategies it stays on par in accuracy, while using six times less token overhead on HallusionBench and MMStar.

Further analysis shows that relay collaboration lets heterogeneous agents progressively integrate complementary insights, combining their different capabilities across stages of the reasoning. Each agent adds what it is good at, at the point in the trajectory where it takes over.

The comparison that matters here is between two shapes of teamwork. Group thinking pays for several full solutions and then reconciles them. Relay thinking pays for one trajectory that changes hands when confidence says it should. On these benchmarks, the second shape keeps pace with the first on accuracy at a fraction of the tokens.