CMAT: A Multi-Agent Collaboration Tuning Framework for Enhancing Small Language Models

尽管大型语言模型（LLMs）在各种任务中展现出了令人印象深刻的性能，但它们的有效操作仍然严重依赖于人类输入以准确指导对话流程，而代理调整是一种关键的优化技术，涉及到人类对模型的调整，以便更好地响应这种指导。为了解决这种依赖性，我们的工作引入了TinyAgent模型，它是在一个精心策划的高质量数据集上训练的。我们还提出了协作多代理调整（CMAT）框架，这是一个创新的系统，旨在通过基于环境反馈的自适应权重更新来增强语言代理的能力。该框架促进了多个智能代理之间的协作学习和实时适应，增强了它们的上下文感知和长期记忆。在这项研究中，我们提出了一个新的通信代理框架，将多代理系统与环境反馈机制相结合，提供了一种可扩展的方法来探索合作行为。值得注意的是，我们的TinyAgent-7B模型表现与GPT-3.5相当，尽管参数更少，这意味着LLMs的效率和有效性有了实质性的改进。

TinyAgent model and Collaborative Multi-Agent Tuning (CMAT) framework are proposed to address the heavy reliance on human input in the effective operation of large language models (LLMs) for better response to guidance.

TinyAgent model is trained on a high-quality dataset and CMAT framework enhances context-awareness and long-term memory of multiple intelligent agents through adaptive weight updates based on environmental feedback.

TinyAgent-7B model exhibits performance on par with GPT-3.5 with fewer parameters, offering a substantial improvement in the efficiency and effectiveness of LLMs. The framework fosters collaborative learning and real-time adaptation among multiple intelligent agents, offering a scalable method to explore cooperative behaviors.

Recent related research includes advancements in large language models and multi-agent systems, such as GPT-3, GShard, and COMET. Additionally, there is ongoing research in exploring cooperative behaviors among multiple agents in various environments.