---
id: 20260816-T0-05
title: "LLM指令遵循存在相变：约束太多时性能断崖下跌"
title_en: "LLMs Show Phase Transitions in Constraint Satisfaction: Too Many Instructions Breaks Performance"
url: https://ai.daily.yangsir.net/daily/20260816-T0-05
issue_date: 2026-08-16
publish_date: 2026-08-15T04:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2608.12426
---

# LLM指令遵循存在相变：约束太多时性能断崖下跌

arXiv新研究发现，大语言模型在处理多个显式约束（如推理结构、安全边界、输出格式）时，性能会出现相变式断崖下跌。每个单约束处理良好，但组合约束数量增加时，模型遵循能力急剧恶化。这一发现对依赖复合指令的实际部署有直接影响，提示开发者在设计提示词时需控制约束数量。

## English Version

**LLMs Show Phase Transitions in Constraint Satisfaction: Too Many Instructions Breaks Performance**

A new arXiv study finds LLMs exhibit phase transitions when handling multiple explicit constraints—reasoning structure, safety boundaries, output schemas. Single constraints are handled well, but as compositional constraints increase, adherence sharply degrades. This has direct implications for deployments relying on compound instructions, suggesting prompt designs should limit constraint counts.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2608.12426)

**详情页**：https://ai.daily.yangsir.net/daily/20260816-T0-05

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*