跳到主要内容Skip to main content

← 返回AI 论文← Back to AI Papers

Agent 与工具Agents & tools

Toolformer: Language Models Can Teach Themselves to Use ToolsToolformer: Language Models Can Teach Themselves to Use Tools

问题:语言模型会写会读,却在算术、事实查询这类基础功能上不如小工具;能不能让模型自己学会调 API。
证据位置:摘要给出自监督训练目标:选哪个 API、何时调、传什么参数、如何把结果并入后续 token 预测;每个 API 只要少量示例;工具含计算器、问答、两个搜索引擎、翻译与日历。
作者自陈局限:工具集固定且少;自监督过滤依赖工具输出质量;实验模型与规模有限;没有讨论开放工具生态下的权限、安全与错误传播。
它没主张的:不是通用工具使用框架;不是不需要人工审核;不是所有 API 都能自学。
Problem: language models write and read well but underperform small tools on arithmetic and factual lookup; can they learn to call APIs themselves?
Where the evidence sits: the abstract describes a self-supervised objective that decides which API to call, when, with what arguments, and how to incorporate the result into future token prediction; a handful of demonstrations per API; tools include calculator, QA, two search engines, translation and calendar.
What the authors say about limits: the tool set is fixed and small; self-supervised filtering depends on tool output quality; model scale is limited; permissions, safety and error propagation in an open tool ecosystem are not discussed.
What it does not claim: not a general tool-use framework; not human-review-free; not every API can be self-taught.

关键参数Key specs

证据位置 / evidence
摘要:自监督 API 调用与少量示例
工具数 / tools
6 类
复现状态 / reproduction
未复现(只读摘要)

官方原文怎么说What the source says

LMs can teach themselves to use external tools via simple APIs
1 https://arxiv.org/abs/2302.04761
decide which APIs to call, when to call them, what arguments to pass, and how to best incorporate the results
2 https://arxiv.org/abs/2302.04761
requiring nothing more than a handful of demonstrations for each API
3 https://arxiv.org/abs/2302.04761

资料与许可Source and licence

资料时间Source checked
2026-11-19
许可 / 类型Licence / kind
arxiv-preprint
许可说明Licence evidence
https://arxiv.org/abs/2302.04761
类目Category
Agent 与工具Agents & tools

论文里的数字怎么读How to read the numbers

先分清测什么、怎么算、和谁比、有没有数据污染。拿不准就回原文看实验设置。Before citing a score, separate what it measures, how it is scored, what it compares against and whether contamination is controlled. If unsure, go back to the experiment setup in the paper. 看评测基准See benchmark cards · 看数据集卡See dataset cards

去官网看原页Open the official page →

← 返回AI 论文← Back to AI Papers