---
title: "GRADE: Efficient Multi-Agent Reasoning via Gated Routing and Adaptive Depth"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2607.10836v1"
permalink: "https://feed.astralor.com/items/itm_77c14f4adcd8237d87f4f0a9"
published: "2026-07-12T16:58:21.000Z"
score: 81
channel: "ai"
tags: ["CoGRPO","GRADE","Multi-agent Systems","Hot-Swapping","Adaptive Depth","Gated Routing"]
---

# GRADE: Efficient Multi-Agent Reasoning via Gated Routing and Adaptive Depth

## AI commentary

> 这篇论文对于研究多智能体协作和大模型推理效率的从业者来说很有价值。它提出了实用的门控机制和训练方法，并支持模型热替换，有望降低实际系统的部署成本。尤其是CoGRPO的无评论家训练思路，或可启发类似场景。

## At a glance

This paper proposes GRADE, a hierarchical multi-agent system where four lightweight learned gates jointly govern agent selection, hierarchy depth, inter-agent communication, and branch pruning. Training uses CoGRPO, a critic-free algorithm. With ~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. Ablations identify hierarchy and masked cross-attention as key contributors, and show per-agent calibration is necessary for safe hot-swapping.

## Key points

- GRADE uses four gates to dynamically select agents, control reasoning depth, decide communication timing, and prune branches.
- CoGRPO training method is critic-free and adapted to multi-agent hierarchies.
- With ~17B active parameters, GRADE outperforms stronger baselines on multiple benchmarks, doubling compute efficiency.
- Per-agent calibration enables hot-swapping of expert models at inference without retraining.

## Full text

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.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_77c14f4adcd8237d87f4f0a9