Route, Communicate, and Reason: Gated Routing and Adaptive Depth for Efficient Multi-Agent Reasoning

2026-07-12

AI Summary

This work addresses the high computational cost and insufficient control over agent selection, reasoning depth, and communication timing in multi-agent reasoning by proposing GRADE, a hierarchical multi-agent system. GRADE jointly optimizes agent selection, reasoning depth, communication decisions, and branch pruning through four lightweight learning gates. It further introduces CoGRPO—a critic-free collaborative policy optimization method—along with a hot-swappable expert pool and a calibration mechanism to significantly enhance flexibility and deployment efficiency. With an average of approximately 17 billion activated parameters, GRADE consistently outperforms existing baselines across GSM8K, MMLU-Pro, and GPQA benchmarks, achieving a 4.8-point lead over the strongest baseline on MMLU-Pro while halving computational cost, and remains competitive on AIME-2025.

Abstract

Multi-agent ensembling multiplies active parameters and inference cost without answering three basic questions: which agents to consult, how deeply a query should traverse a hierarchy of agents, and when inter-agent communication is worth its cost. We present GRADE (Gated Routing and Adaptive Depth for Efficient Reasoning), a hierarchical multi-agent system in which four lightweight learned gates jointly govern agent selection, hierarchy depth, inter-agent communication, and branch pruning. Training uses CoGRPO (Collaborative Group-Relative Policy Optimization), a novel critic-free recipe that adapts GRPO to multi-agent hierarchies and assigns a shared advantage signal to every gate and agent that participated in a rollout. Agent models are drawn from a hot-swappable Expert Registry; per-agent calibration maps allow experts to be replaced at inference time without retraining. At $\sim$17B average active parameters, GRADE outperforms all baselines on GSM8K, MMLUPro, and GPQA, surpassing the strongest baseline by 4.8 points on MMLUPro at half the active compute. On AIME-2025, where model depth dominates, GRADE remains competitive to existing frameworks. Ablations isolate the hierarchy and masked cross-attention as the largest contributors to accuracy, and show that per-agent calibration is necessary for safe hot-swapping.

Problem

• multi-agent reasoning
• agent selection
• hierarchy depth
• inter-agent communication
• inference efficiency

Innovation

• Gated Routing
• Adaptive Depth
• Multi-Agent Reasoning
• CoGRPO
• Hot-Swappable Experts