---
id: 20261025-T0-05
title: "FusionEval：模型融合别只看平均分，长链推理可能大幅掉点"
title_en: "FusionEval: Model Fusion Can Mask Big Losses in Long-Chain Reasoning"
url: https://ai.daily.yangsir.net/daily/20261025-T0-05
issue_date: 2026-10-25
publish_date: 2026-10-24T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2610.30162
---

# FusionEval：模型融合别只看平均分，长链推理可能大幅掉点

arXiv新论文FusionEval提出针对大模型模型融合的评测基准。现有多数评测对长链思维推理和工具使用的覆盖有限，而融合模型往往平均分亮眼，却在特定能力上出现明显损失。该基准通过更细粒度的任务设计揭示融合过程中的能力取舍，帮助研究者和工程团队判断一个融合模型在哪些场景真正可用，而不是被总体指标误导。

## English Version

**FusionEval: Model Fusion Can Mask Big Losses in Long-Chain Reasoning**

A new arXiv paper, FusionEval, introduces a benchmark for evaluating model fusion in large language models. Existing evaluations offer limited coverage of long chain-of-thought reasoning and tool use, and fused models often post strong averages while suffering substantial losses in specific capabilities. FusionEval uses finer-grained task design to expose these trade-offs during fusion, helping researchers and engineering teams determine where a fused model is genuinely usable instead of being misled by aggregate scores.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2610.30162)

**详情页**：https://ai.daily.yangsir.net/daily/20261025-T0-05

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*