FusionEval：大语言模型模型融合基准
原标题：FusionEval: Benchmarking Model Fusion for Large Language Models
AI 摘要
研究团队提出 FusionEval 基准，用于系统评测大语言模型融合方法。该基准基于 Qwen3 和 Llama 构建数学、代码、科学推理、指令遵循和工具使用五个领域的独立专家池，统一比较免训练与需训练两类融合方法在八个公开基准上的能力保留率、运行时间和峰值显存。实验发现免训练方法能保留推理能力但显著损害工具使用，多教师在线策略蒸馏（MOPD）能力更均衡但计算成本更高，低成本同时保留推理、指令遵循与工具使用仍是挑战。
正文节选
FusionEval: Benchmarking Model Fusion for Large Language Models Abstract Model fusion aims to integrate the capabilities of multi-expert models into a single model. Existing evaluations offer limited coverage of reasoning with long chains of thought and tool use, while strong average performance can mask substantial losses in individual expert capabilities. We introduce FusionEval, a benchmark that provides controlled expert pools, integrates fusion methods with and without additional training,