跳转至

文章背景与核心概要

为大语言模型（LLM）配备外部工具能够解锁强大的智能体（Agent）能力，但这往往伴随着巨大的计算和财务成本。此外，工具的使用并非百利而无一害——冗余或低效用的调用甚至会主动降低任务性能。本文引入了一个基于决策理论的规范化框架，旨在优化LLM何时以及如何使用工具。

通过从规范视角（推断真实的必要性与效用）和描述视角（分析模型自我感知的行为）对模型进行分析，作者揭示了当前LLM工具调用策略中存在的基础性错位。为解决这一问题，他们提出训练直接衍生自模型隐藏状态的轻量级隐式需求估计器（Latent Need Estimators, LNEs）。该方法在显著优于模型自我报告的同时，还改善了不同模型规模和工具类型下的预算工具分配效率。

To Call or Not to Call: A Framework to Assess and Optimize LLM Tool Calling

arXiv ID: 2605.00737 [cs.AI]
Authors: Qinyuan Wu, Soumi Das, Mahsa Amani, Arijit Nag, Seungeon Lee, Krishna P. Gummadi, Abhilasha Ravichander, Muhammad Bilal Zafar
Submitted: May 1, 2026 (Last revised August 6, 2026)
Status: Preprint, under review

arXiv ID: 2605.00737 [cs.AI]
Authors: Qinyuan Wu, Soumi Das, Mahsa Amani, Arijit Nag, Seungeon Lee, Krishna P. Gummadi, Abhilasha Ravichander, Muhammad Bilal Zafar
Submitted: May 1, 2026 (Last revised August 6, 2026)
Status: Preprint, under review

📌 Executive Summary

Augmenting Large Language Models (LLMs) with external tools unlocks powerful agentic capabilities, but it often comes with substantial computational and financial costs. Furthermore, tool use is not unconditionally beneficial—redundant or low-utility calls can actively degrade task performance.

This paper introduces a principled framework based on decision-making theory to optimize when and how LLMs utilize tools. By analyzing models through both a normative perspective (inferring true need and utility) and a descriptive perspective (analyzing the model's self-perceived behavior), the authors uncover a fundamental misalignment in current LLM tool-calling policies. To resolve this, they propose training lightweight Latent Need Estimators (LNEs) derived from model hidden states, which significantly outperform self-reports and improve budgeted tool allocation across various model scales and tool types.

📌 执行摘要

为大语言模型（LLM）配备外部工具能够解锁强大的智能体能力，但这往往伴随着巨大的计算和财务成本。此外，工具的使用并非百利而无一害——冗余或低效用的调用甚至会主动降低任务性能。

本文引入了一个基于决策理论的规范化框架，旨在优化LLM何时以及如何使用工具。通过从规范视角（推断真实的必要性与效用）和描述视角（分析模型自我感知的行为）对模型进行分析，作者揭示了当前LLM工具调用策略中存在的基础性错位。为了解决这一问题，他们提出训练直接衍生自模型隐藏状态的轻量级隐式需求估计器（Latent Need Estimators, LNEs）。该方法在性能上显著优于自我报告，并在各种模型规模和工具类型下改善了基于预算的工具分配。

🧭 Key Concepts & Framework

The study evaluates tool-use decisions along three primary factors: 1. Necessity: Is the external tool fundamentally required to complete the task? 2. Utility: Does invoking the tool positively impact the final output quality? 3. Affordability: Is the cost of the tool call justified under resource or budget constraints?

Evaluation Lenses

  • Normative Perspective: Infers the true need and utility required for optimal tool invocation.
  • Descriptive Perspective: Infers the self-perceived need and utility derived from observed model behaviors.

🧭 核心概念与框架

该研究从以下三个主要因素评估工具使用决策： 1. 必要性（Necessity）： 完成该任务是否根本上需要外部工具？ 2. 效用（Utility）： 调用工具是否对最终输出质量产生积极影响？ 3. 可承受性（Affordability）： 在资源或预算限制下，工具调用的成本是否合理？

评估视角

  • 规范视角（Normative Perspective）： 推断实现最佳工具调用所需的真实需求与效用。
  • 描述视角（Descriptive Perspective）： 从观察到的模型行为中推断出自我感知的需求与效用。

🧪 Empirical Findings

The researchers evaluated six open-source models and one proprietary OpenAI model across native and customized harnesses, two different tools, and six distinct tasks. Key takeaways include: * Misalignment: Models' perceived needs and utilities frequently misalign with their true ground-truth values, especially under tight budget constraints. * Adverse Effects: This misalignment frequently results in costly overuse and performance-degrading tool calls.

🧪 实证发现

研究人员在原生与定制测试框架、两种不同工具以及六项不同任务上，评估了六个开源模型和一个OpenAI专有模型。主要结论包括： * 错位： 模型感知到的需求和效用常常与其真实的基准真实值（ground-truth values）不符，尤其是在严苛的预算限制下。 * 负面影响： 这种错位往往导致代价高昂的过度使用以及损害性能的工具调用。

🛠️ Optimization: Latent Need Estimators (LNEs)

To overcome model self-reporting inaccuracies, the authors introduce Latent Need Estimators (LNEs): * Mechanism: Lightweight classifiers trained directly on the model's hidden states. * Performance: LNEs predict true task-based tool necessity more accurately than direct model self-reports. * Impact: Enables efficient, budgeted tool allocation across varying model scales and tool typologies.

🛠️ 优化方案：隐式需求估计器（LNEs）

为了克服模型自我报告的不准确性，作者引入了隐式需求估计器（Latent Need Estimators, LNEs）： * 机制： 直接在模型的隐藏状态上进行训练的轻量级分类器。 * 性能： LNEs预测基于真实任务的工具需求时，比直接的模型自我报告更加准确。 * 影响： 能够在不同的模型规模和工具类型之间实现高效、带预算约束的工具分配。

🔗 Resources & Artifacts

🔗 资源与工件

📝 License

This work is published under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

license icon

📝 许可协议

本作品根据 知识共享署名 4.0 国际许可协议 (CC BY 4.0) 发布。

license icon