Neura Market
Agent Q-Mix: Selecting the Right Action for LLM Multi-Agent Systems (April 2026) logo

Agent Q-Mix: Selecting the Right Action for LLM Multi-Agent Systems (April 2026)

Free

Reformulates topology selection as cooperative MARL — each agent selects communication actions that jointly induce round-wise communication graphs; improves coordination efficiency

FreeFree tier
Type
Open Source
Visit Website GitHub (40 stars)

About Agent Q-Mix: Selecting the Right Action for LLM Multi-Agent Systems (April 2026)

Agent Q-Mix is a reinforcement learning framework that reformulates topology selection in LLM multi-agent systems as a cooperative Multi-Agent Reinforcement Learning (MARL) problem. It uses QMIX value factorization to learn decentralized communication decisions, where each agent selects from a set of communication actions that jointly induce a round-wise communication graph. The architecture combines a topology-aware GNN encoder, GRU memory, and per-agent Q-heads under a Centralized Training with Decentralized Execution (CTDE) paradigm. The reward function balances task accuracy with token cost. The framework achieves the highest average accuracy across seven core benchmarks in coding, reasoning, and mathematics, and demonstrates superior token efficiency and robustness against agent failure. Notably, on the challenging Humanity's Last Exam (HLE) using Gemini-3.1-Flash-Lite as a backbone, Agent Q-Mix achieves 20.8% accuracy, outperforming existing methods like Microsoft Agent Framework (19.2%) and LangGraph (19.2%).

Key Features

Reformulates topology selection as cooperative Multi-Agent Reinforcement Learning (MARL) using QMIX value factorization
Learns decentralized communication decisions where each agent selects actions that jointly induce round-wise communication graphs
Architecture includes topology-aware GNN encoder, GRU memory, and per-agent Q-heads under CTDE paradigm
Reward function balances task accuracy with token cost
Achieves highest average accuracy across seven benchmarks in coding, reasoning, and mathematics
Demonstrates superior token efficiency and robustness against agent failure
Outperforms Microsoft Agent Framework and LangGraph on Humanity's Last Exam (HLE) with 20.8% accuracy using Gemini-3.1-Flash-Lite

Pros & Cons

Pros
  • Achieves highest average accuracy across multiple benchmarks compared to existing methods
  • Balances task accuracy with token efficiency, reducing computational costs
  • Robust to agent failures, making it suitable for unreliable multi-agent environments
  • Uses CTDE paradigm allowing decentralized execution after centralized training
  • Open access publication with full details on arXiv
Cons
  • Requires centralized training phase, which may be computationally intensive
  • Only evaluated on benchmarks; real-world deployment performance not yet studied
  • Limited to LLM-based agents; applicability to other agent types is unexplored

Best For

Improving coordination in LLM multi-agent systems for complex problem-solvingOptimizing communication topology in coding, reasoning, and mathematical tasksReducing token costs while maintaining high task accuracy in multi-agent setupsEnhancing robustness of multi-agent reasoning systems against agent failures

FAQ

What is Agent Q-Mix?
Agent Q-Mix is a reinforcement learning framework that reformulates topology selection in LLM multi-agent systems as a cooperative Multi-Agent Reinforcement Learning (MARL) problem using QMIX value factorization.
How does Agent Q-Mix work?
Each agent learns to select communication actions from a set, and these actions jointly induce a round-wise communication graph. The architecture uses a topology-aware GNN encoder, GRU memory, and per-agent Q-heads under Centralized Training with Decentralized Execution (CTDE).
What benchmarks were used to evaluate Agent Q-Mix?
The framework was evaluated on seven core benchmarks covering coding, reasoning, and mathematics tasks, including Humanity's Last Exam (HLE).
How does Agent Q-Mix compare to other multi-agent frameworks?
On HLE using Gemini-3.1-Flash-Lite, Agent Q-Mix achieves 20.8% accuracy, outperforming Microsoft Agent Framework (19.2%), LangGraph (19.2%), AutoGen, and Lobster.