# Microsoft 论文揭示 LLM Agent 在 16 步长程任务中从近乎满分跌至 0-33%

- 来源：Rohan Paul (@rohanpaul_ai)
- 发布时间：2026-09-08 07:45
- AIHOT 分数：68
- AIHOT 链接：https://aihot.virxact.com/items/cmtrwy5yy0273roft910r0rwt
- 原文链接：https://x.com/rohanpaul_ai/status/2097109113062998515

## AI 摘要

一篇 Microsoft 论文研究 LLM Agent 长程衰减，跨 9 个模型发现依赖步骤越多成功率越低，ToolQA 上短程近乎满分的模型到 16 步仅剩 0-33% 成功率。实验显示缩短上下文反而使衰减更糟，说明问题主要由步数而非上下文长度驱动；作者建议按真实工作流长度测试、度量每步可靠性并在错误扩散前加入检查点。

## 正文

New Microsoft paper. Long agent runs expose failures that short benchmarks miss. Agents can look reliable at 2 or 4 steps and fall apart by 16.

every agent step has some chance of going wrong, and those small errors compound as the workflow gets longer.

Across 9 models, success usually dropped as the number of dependent steps increased.

On ToolQA, models that were near-perfect on short runs fell to just 0-33% success by 16 steps.

Long context was not the main driver: shortening the context made the decline worse, so blindly trimming history is not a reliability fix.

For builders, the recommendation is straightforward: stop treating a benchmark pass rate as proof that an agent is production-ready.

Test agents at the workflow lengths you actually expect, measure per-step reliability, and add checks or checkpoints before a bad step poisons everything that follows.