Reasoning models often generate very long reasoning traces, making inference computationally expensive. Existing approaches typically improve efficiency either through inference-time early-stopping mechanisms or by explicitly encouraging shorter reasoning during training, for example through reinforcement learning with length penalties. We show that substantial efficiency gains can instead emerge

---
**📖 中文解读**
以上内容由AI翻译自英文原文，可能存在不准确之处。建议阅读[原文](https://arxiv.org/abs/2609.31619v1)获取最准确的信息。

---
🔗 **原文链接**: [Learning to Stop without Learning to Stop: Self-Supervised C](https://arxiv.org/abs/2609.31619v1)
🏷️ **转载来源**: ArXiv cs.AI
> 本文由小九AI技术站翻译整理，内容版权归原作者所有。
👤 作者: Parsa Hosseini, Akasha Tigalappanavara, Sumit Nawathe, Chenrui Fan, Sourya Basu, Genta Indra Winata, Anirban Das, Soheil Feizi, Nima Chitsazan

---
🐾 **小九锐评**

这篇论文来自arXiv预印本，虽然还没有经过同行评审，但选题方向值得关注。
建议先读中文摘要判断是否相关，再看全文细节。
推理能力是LLM的下一个战场。这篇文章技术细节到位，适合有一定基础的同学细读。

你对这个话题有什么看法？欢迎在评论区讨论 💬

> _转载自 ArXiv cs.AI，内容版权归原作者所有_

---
⏱️ 2026-09-28 13:01