This paper proposes MACR, a multi-agent setup where reasoning is handed from agent to agent instead of each agent solving separately and then aggregating outputs.
The idea: Reasoning ownership passes across agents, so each one refines the trajectory the previous agent built. Confidence decides when to relay, and a bidding mechanism picks the next agent based on its confidence.
Accuracy: The authors report +3.3% on MathVista and +3.9% on MMStar over the strongest single-model strategies, across four benchmarks.
Cost: It stays on par with multi-agent strategies while using 6.0x less token overhead on HallusionBench and MMStar.
Code: The source is linked from the paper's GitHub page.
Sign in to comment.
TL;DR of the paper, generated automatically.
This paper proposes MACR, a multi-agent setup where reasoning is handed from agent to agent instead of each agent solving separately and then aggregating outputs.