# Google 论文：程序图让 LLM 智能体长任务更稳

- 来源：Rohan Paul (@rohanpaul_ai)
- 发布时间：2026-09-21 17:56
- AIHOT 分数：48
- AIHOT 链接：https://aihot.news/items/cmub36exi0gynrolniruo6h7e
- 原文链接：https://x.com/rohanpaul_ai/status/2101973711612190756

## AI 摘要

Google 新论文提出 Procedural Graphs，把 LLM 智能体的工作流从聊天历史中移出，改为可编辑的程序图，并让另一个 LLM 对比成功与失败案例来修改图，只有不损害留出任务的修改才会被保留。

## 正文

New Google paper shows LLM agents handle long tasks better when their workflow lives in an editable procedure graph that learns from execution, instead of being buried in chat history.

The problem is: as agents run longer, they can forget where they are, repeat tools, or do steps in the wrong order.

Procedural Graphs give the agent a small map of what can happen next, while still letting the LLM reason freely.

After runs finish, another LLM compares successes and failures and edits the map, but an edit is kept only if it does not hurt held-out tasks.

Across 24 model-and-benchmark combinations, the method ranked 1st or tied 1st in 21.

It also repaired a bad human-designed workflow: on MultiChallenge, success went from 58.93% with the flawed graph to 92.86% after refinement.

This guidance costs extra tokens, so it makes most sense when long workflows are the bottleneck.

Overall, it recommends move critical procedures out of chat history and into an explicit, editable workflow that improves from execution.

– arxiv. org/abs/2609.09153

Title: "Procedural Graphs: Self-Evolving Execution Structures for LLM Agents"