20260514:增加新内容
This commit is contained in:
25
concepts/confidence-correctness-alignment.md
Normal file
25
concepts/confidence-correctness-alignment.md
Normal file
@@ -0,0 +1,25 @@
|
||||
---
|
||||
title: 置信度-正确性对齐 (Confidence-Correctness Alignment)
|
||||
created: 2025-04-15
|
||||
updated: 2026-05-01
|
||||
type: concept
|
||||
tags: []
|
||||
sources: []
|
||||
---
|
||||
|
||||
# 置信度-正确性对齐 (Confidence-Correctness Alignment)
|
||||
|
||||
**模型初始置信度与实际正确性的一致程度**,是决定 [[intrinsic-rewards-sharpening|内在 URLVR]] 成败的关键变量。
|
||||
|
||||
## 二分命运
|
||||
|
||||
- **对齐**: 锐化 → 放大正确推理路径 → 性能提升
|
||||
- **错位**: 锐化 → 系统性放大错误 → 模型崩溃
|
||||
|
||||
对齐强度可通过 [[model-collapse-step|MCS]] 量化。
|
||||
|
||||
## 相关概念
|
||||
|
||||
- [[intrinsic-rewards-sharpening]] — Sharpening 机制
|
||||
- [[model-collapse-step]] — 对齐强度的度量
|
||||
- [[he-urlvr-sharpening-2026]] — 综述参考
|
||||
Reference in New Issue
Block a user