Multi-Agent Relay Thinking
Abstract Overview
This paper proposes Multi-Agent Confidence-guided Relay (Macr), a multi-agent collaboration framework that replaces parallel group thinking with a relay-style process in which reasoning ownership is transferred across agents during generation. The method calibrates local confidence scores across heterogeneous vision-language models so that their confidence values become comparable, then uses these scores to choose the starting model, detect reasoning bottlenecks, and select the next agent through online bidding. Experiments are conducted with a three-model pool on four multimodal benchmarks: MathVista, LogicVista, MMStar, and HallusionBench. The evidence indicates that relay-based collaboration can combine complementary model strengths across different reasoning stages while reducing unnecessary generation overhead.
Novelty
The main novelty is the introduction of a relay thinking topology for multi-agent reasoning, where agents sequentially inherit and refine a shared reasoning trajectory instead of producing separate full solutions for later aggregation. A second distinctive contribution is the calibration of window-level confidence across heterogeneous models, enabling real-time cross-model handoffs based on comparable local confidence.
Results
Macr consistently outperforms the strongest single-model strategies in reported accuracy, including gains of +3.30% on MathVista and +3.87% on MMStar. It achieves accuracy comparable to stronger multi-agent baselines while using far fewer tokens, achieving 6.0x lower token overhead than H-DMAD on HallusionBench and MMStar. Additional analyses show that the models exhibit complementary strengths across capabilities and reasoning stages, and that confidence-guided relays outperform fixed initiation, routing, or trigger variants.
Key Points
- Macr uses calibrated sliding-window confidence to decide when generation should be relayed and which model should take over.
- The framework shifts multi-agent collaboration from parallel answer aggregation to sequential refinement of a shared reasoning chain.
- Empirical results show improved accuracy over single-model baselines and substantially better efficiency than heavier multi-agent methods.