Orchestrating Intelligence: Confidence-Aware Routing for Efficient Multi-Agent Collaboration across Multi-Scale Models

TL;DR

This work addresses the inefficiency of fixed backbones in multi-agent reasoning by introducing OI-MAS, a dynamic conductor that jointly selects agent roles and backbone scales per reasoning step. It uses a state-dependent routing mechanism and a confidence-aware optimization objective to allocate computational resources where needed, reducing waste on easy subtasks while escalating to stronger backbones for complex steps. Across mathematics, professional reasoning, and programming tasks, OI-MAS achieves notable accuracy improvements (up to about 12.88% in the abstract) and substantial cost savings (up to about 79.78%), while also reducing latency. The approach advances practical multi-agent reasoning by enabling more reliable, efficient collaboration among heterogeneous LLMs, with robust generalization to out-of-distribution tasks and clear implications for latency-sensitive deployments.

Abstract

While multi-agent systems (MAS) have demonstrated superior performance over single-agent approaches in complex reasoning tasks, they often suffer from significant computational inefficiencies. Existing frameworks typically deploy large language models (LLMs) uniformly across all agent roles, failing to account for the varying cognitive demands of different reasoning stages. We address this inefficiency by proposing OI-MAS framework, a novel multi-agent framework that implements an adaptive model-selection policy across a heterogeneous pool of multi-scale LLMs. Specifically, OI-MAS introduces a state-dependent routing mechanism that dynamically selects agent roles and model scales throughout the reasoning process. In addition, we introduce a confidence-aware mechanism that selects appropriate model scales conditioned on task complexity, thus reducing unnecessary reliance on large-scale models. Experimental results show that OI-MAS consistently outperforms baseline multi-agent systems, improving accuracy by up to 12.88\% while reducing cost by up to 79.78\%.

Figures (8)

• Figure 1: Paradigm comparison of static multi-agent systems, dynamic multi-agent routing with a shared LLM backbone, and dynamic multi-agent routing across a multi-scale LLM pool.
• Figure 2: Overview of the proposed OI-MAS framework. The top part shows a per-turn routing policy that coordinates agent roles and assigns model capacity from a multi-scale LLM pool based on the current reasoning state; the bottom part illustrates how the system evolves across turns.
• Figure 3: The comparison of accuracy and inference cost across four benchmarks, where different marker shapes denote different baseline categories.
• Figure 4: Wall-clock latency of OI-MAS and baselines on the GPQA benchmark.
• Figure 5: Model selection distribution across the five difficulty levels on the MATH dataset.