20260514:增加新内容
This commit is contained in:
37
concepts/long-context-understanding.md
Normal file
37
concepts/long-context-understanding.md
Normal file
@@ -0,0 +1,37 @@
|
||||
---
|
||||
title: 长上下文理解 (Long-Context Understanding)
|
||||
created: 2026-05-01
|
||||
updated: 2026-05-01
|
||||
type: concept
|
||||
tags: [llm, architecture, benchmark]
|
||||
sources: [papers/hunyuan-team-cl-bench-life.md]
|
||||
---
|
||||
|
||||
# 长上下文理解 (Long-Context Understanding)
|
||||
|
||||
> 语言模型在超长输入序列(10K–1M+ tokens)中检索信息和进行推理的能力。与 [[real-life-context-learning]] 相关但不等价。
|
||||
|
||||
## 定义
|
||||
|
||||
长上下文理解考察模型在以下方面的表现:
|
||||
- **信息检索**:能否在长文本中的任意位置找到特定事实(Needle-in-a-Haystack)
|
||||
- **多跳推理**:能否组合分散在不同位置的信息
|
||||
- **位置鲁棒性**:性能是否随目标信息位置变化(如 [[lost-in-the-middle]])
|
||||
|
||||
## 与真实生活上下文学习的解耦
|
||||
|
||||
CL-bench Life 的重要发现:**长上下文能力 ≠ 真实生活上下文学习能力**:
|
||||
|
||||
- CL-bench Life 的上下文长度(5.4K–170.8K)在大多数前沿模型窗口内
|
||||
- 任务解决率与上下文长度**无强相关性**
|
||||
- 混乱上下文的推理质量是独立于上下文长度的瓶颈
|
||||
|
||||
## 相关概念
|
||||
- [[context-learning]] — 通用上下文学习
|
||||
- [[real-life-context-learning]] — 真实生活上下文学习
|
||||
- [[lost-in-the-middle]] — 中间信息丢失
|
||||
- [[million-token-context]] — 百万 Token 上下文
|
||||
|
||||
---
|
||||
|
||||
*Last Updated: 2026-05-01*
|
||||
Reference in New Issue
Block a user