---
id: 20260609-T0-10
title: "IDPR框架：让LLM在需要时才深度思考"
title_en: "IDPR Framework Makes LLMs Think Deeply Only When Needed"
url: https://ai.daily.yangsir.net/daily/20260609-T0-10
issue_date: 2026-06-09
publish_date: 2026-06-08T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2606.06745
---

# IDPR框架：让LLM在需要时才深度思考

arXiv论文提出IDPR框架，解决LLM推理效率问题。现有方法对所有输入都调用慢速推理，计算成本高且不必要。新框架根据响应条件动态决定是否启用深度思考，在保证质量的同时降低开销。研究显示，该方法在复杂推理任务中效果接近全程深度思考，但计算效率提升显著。将优化资源分配，让大模型更聪明地工作。

## English Version

**IDPR Framework Makes LLMs Think Deeply Only When Needed**

IDPR framework on arXiv addresses LLM reasoning inefficiency. Current methods use slow deliberation for all inputs, incurring high unnecessary costs. The new framework dynamically enables deep thinking based on response conditions, maintaining quality while reducing overhead. Research shows it achieves near-full deliberation performance on complex tasks with significant efficiency gains, optimizing resource allocation for smarter LLM operation.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2606.06745)

**详情页**：https://ai.daily.yangsir.net/daily/20260609-T0-10

---

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