---
id: 20260415-T0-15
title: "Self-Calibrating：测试时判别式蒸馏校准LLM"
title_en: "Self-Calibrating Language Models via Test-Time Distillation"
url: https://ai.daily.yangsir.net/daily/20260415-T0-15
issue_date: 2026-04-15
publish_date: 2026-04-14T04:00:00.000Z
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2604.09624
---

# Self-Calibrating：测试时判别式蒸馏校准LLM

论文提出一种无需标记数据的LLM自校准方法，通过测试时判别式蒸馏解决模型过度自信问题。该方法在分布偏移下表现稳定，使模型能够准确评估回答的可信度。

## English Version

**Self-Calibrating Language Models via Test-Time Distillation**

A paper introduces a self-calibration method for LLMs that works without labeled data. Using test-time discriminative distillation, it addresses overconfidence issues and maintains stability under distribution shifts, enabling accurate answer credibility assessment.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2604.09624)

**详情页**：https://ai.daily.yangsir.net/daily/20260415-T0-15

---

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