文章背景与核心概要
本文介绍了“指令遵循函数调用”(Instruction-Followed Function Calling,简称 IFFC)这一全新框架,旨在提升智能体大语言模型(LLM)的可靠性与效率。作者通过研究证明,当把大语言模型视为标准的指令遵循任务而非专用的工具调用场景时,其在函数调用方面的表现会更佳。
通过将函数调用逻辑从主 LLM 中解耦,并委派给更小、更专注的模型,IFFC 框架实现了更高的准确率(特别是在偏向推理的模型中),同时在激进的量化(Quantization)条件下依然保持了强健的性能。该方法为边缘计算环境中实现可靠且资源高效的函数调用提供了一种可扩展的解决方案。
Small Reasoning Models are Instruction Followers in Function Calling
arXiv: 2608.22472
Date: August 23, 2026
Authors: Yalda Taheri, Mohammad Hassan Heydari, Erfan Naaman, Afsaneh Fatemi
arXiv: 2608.22472
Date: August 23, 2026
Authors: Yalda Taheri, Mohammad Hassan Heydari, Erfan Naaman, Afsaneh Fatemi
Summary
本论文引入了指令遵循函数调用(IFFC),这是一个能够提升智能体大语言模型(LLM)可靠性与效率的新型框架。作者证明,当大语言模型被作为标准的指令遵循任务而非专用的工具调用上下文来处理时,其在函数调用上的表现更佳。通过将函数调用逻辑从主 LLM 中解耦并委派给更小、更专注的模型,IFFC 框架实现了卓越的准确率——尤其是在面向推理的模型中——同时在激进的量化下保持了强健的性能。这种方法为边缘计算环境中的可靠、资源高效的函数调用提供了一个可扩展的解决方案。
This paper introduces Instruction-Followed Function Calling (IFFC), a novel framework that improves the reliability and efficiency of agentic large language models (LLMs). The authors demonstrate that LLMs perform better at function calling when treated as standard instruction-following tasks rather than specialized tool-calling contexts. By decoupling function-calling logic from the primary LLM and delegating it to a smaller, dedicated model, the IFFC framework achieves superior accuracy—particularly in reasoning-oriented models—while maintaining robust performance under aggressive quantization. This approach offers a scalable solution for reliable, resource-efficient function calling in edge-computing environments.
Key Contributions
- 范式转变: 证明了指令遵循模型在工具执行方面极为高效,从而对原生函数调用(NFC)架构的必要性提出了挑战。
- IFFC 框架: 一种解耦架构,将函数调用逻辑委派给更小的专业化模型。
- 性能提升: 在基准测试中持续优于原生函数调用(NFC)和基于提示词的函数调用(PFC)。
- 边缘端就绪: 证明了 IFFC 方法在量化过程中对精度下降具有极高的抵抗力,非常适合端侧部署。
- Paradigm Shift: Challenges the necessity of native function-calling (NFC) architectures by proving that instruction-following models are highly effective at tool execution.
- IFFC Framework: A decoupled architecture that delegates function-calling logic to smaller, specialized models.
- Performance Gains: Consistently outperforms both native function calling (NFC) and prompt-based function calling (PFC) baselines.
- Edge-Ready: Demonstrates that the IFFC approach is highly resistant to accuracy degradation during quantization, making it ideal for on-device deployment.
Access & Resources
- 查看 PDF: 下载论文
- HTML 版本: 实验性 HTML
- TeX 源码: 源代码
- 许可协议: 知识共享署名 4.0 国际
- View PDF: Download Paper
- HTML Version: Experimental HTML
- TeX Source: Source Code
- License: Creative Commons Attribution 4.0 International

Metadata
| 字段 (Field) | 详情 (Details) |
|---|---|
| 学科 (Subjects) | 人工智能 (cs.AI);计算与语言 (cs.CL) |
| DOI | https://doi.org/10.48550/arXiv.2608.22472 |
| 提交时间 (Submission) | 2026年8月23日 |
Field Details Subjects Artificial Intelligence (cs.AI); Computation and Language (cs.CL) DOI https://doi.org/10.48550/arXiv.2608.22472 Submission 23 Aug 2026