MoAA：将多模型集体智能蒸馏进单一高效模型
原标题：Mixture-of-Agents Alignment: Harnessing the Collective Intelligence of Open-Source LLMs to Improve Post-Training
AI 摘要
Together AI 在 ICML 2025 论文中提出 Mixture-of-Agents Alignment (MoAA) 方法，通过将多个开源大模型的集体智能蒸馏到单个小模型中，使 Llama-3.1-8B 和 Gemma-2-9B 等模型性能接近甚至超越 10 倍大小的模型。该方法包括 MoAA-SFT 和 MoAA-DPO 两个阶段，使用 MoA 生成高质量 SFT 数据和偏好数据，成本比 GPT-4o 低 15%。团队已发布论文、SFT 数据及四个模型的权重。
正文节选
TL;DR: In our new ICML 2025 paper, we proposed Mixture-of-Agents Alignment (MoAA) as an effective post-training approach that harnesses the collective intelligence of open-source large language models (LLMs). Building on our previous work where Mixture-of-Agents ensembles outperformed GPT-4o, MoAA now achieves this collaborative advantage in a single, efficient model. We released our paper, SFT data along with the model weights for Llama-3.1-8B-Instructt-MoAA-SFT, Llama-3.1-8B-Instructt-MoAA-DPO