# 读对了，做错了：诊断代理循环中 System One 模型的决策失败

> 原文链接：https://robotics.hellobike.com/papers/2026-10-13-2610-18579
> 频道：papers | 更新日期：2026-10-13 | 标签：arxiv, cs.RO
> 核心摘要：本文用独立事实探针诊断 System One 模型代理决策错误，并在多类代理决策状态基准上验证。

# 读对了，做错了：诊断代理循环中 System One 模型的决策失败

::: v-pre
> 原文链接：[Reading Right, Acting Wrong: Diagnosing Decision Failures of System One Models in Agent Loops](https://arxiv.org/abs/2610.18579)（arXiv 2610.18579 · 发布日期：2026-10-13）

**Reading Right, Acting Wrong: Diagnosing Decision Failures of System One Models in Agent Loops**

本文用独立事实探针诊断 System One 模型代理决策错误，并在多类代理决策状态基准上验证。

## 核心要点

- System One 模型快速低成本，但选错动作时难判断是误读状态还是读对仍行动错。
- 方法为每个下一动作问题配对同状态事实探针，并让模型独立回答这些探针。
- 在十七个快速系统和一个推理模型上，多数错误动作出现在正确读取之后。
- 当被探查事实足以决定动作时，模型读全事实仍可能未能组合；用其答案经代码组合可修复。

**验证范围与局限**：验证范围为冻结的代理决策状态基准，来源包括仿真环境、客服策略、工具使用、网页导航和编码。

## 英文摘要 (Abstract)

System One models choose an agent's next action by scoring a supplied set of options in a single pass. This makes them fast and cheap, but when they choose wrongly there is nothing to inspect: did the model misread what the agent has seen, or read it correctly and still act wrongly? We answer behaviorally. On a frozen benchmark of agent decision states from simulated environments, customer-service policies, tool use, web navigation and coding, we pair every next-action question with factual probes about the same state, each answered independently. Across seventeen fast systems and a reasoning model, the answer is consistent: most wrong actions follow correct reading, and the more accurate a system, the more of its errors take this form. These errors are not noise; they gather on the same states across systems of every kind. Where the probed facts provably determine the action, models that read every fact still fail to combine them, and their own answers, composed by code, repair the failure. On real states, what is missing is often not a fact but the next sub-goal.

作者：Yubo Li, Ramayya Krishnan, Rema Padman
:::

## 原始参考来源 (Sources)

- [arXiv 2610.18579：Reading Right, Acting Wrong: Diagnosing Decision Failures of System One Models in Agent Loops](https://arxiv.org/abs/2610.18579)