将大语言模型用作智能体：成本几何？

arXiv cs.LG 论文
llm-agents agent-harness compute-cost resource-theory arxiv theory

摘要

这篇 arXiv 论文提出了语言模型智能体机器（LAM），这是一个受资源约束的框架，用于量化 LLM 智能体框架（上下文、记忆、工具、验证）的计算成本，建立了下界与可靠性结论，并在 GPT-6 Astra 上进行了验证。

arXiv:2610.02488v1 Announce Type: new Abstract: Language-model agents increasingly rely on harnesses that manage bounded context, persistent memory, tools, verification, and repeated execution, yet existing notions of model capability do not quantify the computational resources these mechanisms consume. We introduce the Language Model Agent Machine (LAM), a resource-bounded abstraction that fixes the underlying semantic model while explicitly charging harness-level resources. We establish four classes of results. Communication: LAM execution is instancewise equivalent to red--blue pebbling under simultaneous call--transfer budgets, transferring classical I/O lower bounds to context--memory traffic. Access: memory interfaces induce asymptotic separations, including a $\Theta(n)$ gap between random and non-speculative sequential access on pointer chasing. Recomputation: bit-reversal DAGs require $\Theta(n^2/(C+S)+n)$ model calls with context capacity $C$ and persistent-memory capacity $S$, quantifying when stored intermediate state avoids repeated semantic computation. Reliability: we derive tight stage-local sampling bounds, exact imperfect-verification costs, and a Young--Daly-type checkpoint law with a closed-form optimal verification interval. Controlled and held-out experiments on GPT-6 Astra test communication and reliability predictions, including checkpoint optima, policy selection under programmatic checking, and tradeoffs among call granularity, logical input traffic, and reliability on chained MATH tasks. Together, these results provide a resource theory for the computational cost of language-model agent harnesses.
查看原文
查看缓存全文

缓存时间: 2026/10/05 10:05 

# Harnessing LLMs as Agents: What Does It Cost?
Source: [https://arxiv.org/abs/2610.02488](https://arxiv.org/abs/2610.02488)
[View PDF](https://arxiv.org/pdf/2610.02488)

> Abstract:Language\-model agents increasingly rely on harnesses that manage bounded context, persistent memory, tools, verification, and repeated execution, yet existing notions of model capability do not quantify the computational resources these mechanisms consume\. We introduce the Language Model Agent Machine \(LAM\), a resource\-bounded abstraction that fixes the underlying semantic model while explicitly charging harness\-level resources\. We establish four classes of results\. Communication: LAM execution is instancewise equivalent to red\-\-blue pebbling under simultaneous call\-\-transfer budgets, transferring classical I/O lower bounds to context\-\-memory traffic\. Access: memory interfaces induce asymptotic separations, including a $\\Theta\(n\)$ gap between random and non\-speculative sequential access on pointer chasing\. Recomputation: bit\-reversal DAGs require $\\Theta\(n^2/\(C\+S\)\+n\)$ model calls with context capacity $C$ and persistent\-memory capacity $S$, quantifying when stored intermediate state avoids repeated semantic computation\. Reliability: we derive tight stage\-local sampling bounds, exact imperfect\-verification costs, and a Young\-\-Daly\-type checkpoint law with a closed\-form optimal verification interval\. Controlled and held\-out experiments on GPT\-6 Astra test communication and reliability predictions, including checkpoint optima, policy selection under programmatic checking, and tradeoffs among call granularity, logical input traffic, and reliability on chained MATH tasks\. Together, these results provide a resource theory for the computational cost of language\-model agent harnesses\.

## Submission history

From: Zelin Zhao \[[view email](https://arxiv.org/show-email/fc60db2d/2610.02488)\] **\[v1\]**Thu, 1 Oct 2026 21:05:49 UTC \(1,820 KB\)

相似文章

LLM Agents Factory: Retrieval of Domain-Specific LLM Agents

arXiv cs.CL

The paper presents LLM Agents Factory, a retrieval-based framework that constructs domain-specific LLM agents from a base of over 20K predefined agent profiles, offering a cost-efficient and controllable alternative to dynamic agent generation. Experiments show accuracy comparable to AutoGen with a 120B backbone at substantially lower inference cost.

一个 Token 值多少算力？基于 Mixture-of-Agents 测量大模型的每 Token 充分算力

arXiv cs.AI

本文利用由 Qwen、OLMo 和 R1-distilled 系列共 15 个模型组成的 Mixture-of-Agents 评审面板，测量了大模型的每 Token「充分算力」。研究发现，一个 0.5B 的小模型即可复现参考模型 92%–95% 的输出 Token，而成本最高的 10% Token 却占用了 64%–80% 的 FLOPs。由此得到的每 Token 算力分布图将模型路由延迟从 7.59 秒降低至 5.12 秒，并使推测式草稿生成所需的草稿 Token 数比固定窗口基线减少 32.6%。