20260706:新增一些文章
This commit is contained in:
45
concepts/rubric-aggregation.md
Normal file
45
concepts/rubric-aggregation.md
Normal file
@@ -0,0 +1,45 @@
|
||||
---
|
||||
title: "Rubric Aggregation"
|
||||
created: 2026-06-27
|
||||
updated: 2026-06-27
|
||||
type: concept
|
||||
tags:
|
||||
- rubric
|
||||
- evaluation
|
||||
- scoring
|
||||
sources:
|
||||
- "rubrics-survey-2026"
|
||||
---
|
||||
|
||||
# Rubric Aggregation
|
||||
|
||||
## 定义
|
||||
|
||||
Rubric aggregation 指将多个 rubric item 的逐项评分**合并为单一总分**的方法。这是 rubric 评估流水线中 form item scores → overall score 的关键步骤。
|
||||
|
||||
## 三类聚合策略
|
||||
|
||||
### 1. 直接平均/求和 (Direct Averaging/Summation)
|
||||
所有 rubric items 等权重,直接加总或平均:
|
||||
- S_avg = (1/k) Σ cⱼ
|
||||
- S_sum = Σ cⱼ
|
||||
- 最简单透明,适用于所有 items 同等重要时
|
||||
|
||||
### 2. 加权求和 (Weighted Summation)
|
||||
不同 items 赋予不同权重:
|
||||
- S_R = Σ wⱼcⱼ / Σ wⱼ
|
||||
- 允许关键维度(如 safety、task completion)权重更高
|
||||
- 最灵活,但是权重设计本身成为新的设计挑战
|
||||
|
||||
### 3. 隐式聚合 (Implicit Aggregation)
|
||||
将完整 rubric + 模型输出直接交给 judge model,让它隐式输出总分:
|
||||
- S_imp = f_ϕ(x, y, R)
|
||||
- 推理时更简单、成本更低,但不再能观察到每个 item 的贡献
|
||||
|
||||
### 本文主要聚焦显式聚合(加权求和/直接平均),因为可检查、可分析。
|
||||
|
||||
## 参考
|
||||
|
||||
- [[rubrics-for-llms|Rubrics for LLMs]]
|
||||
- [[rubrics-survey-2026|Rubrics Survey (2026)]]
|
||||
- [[rubric-driven-evaluation]]
|
||||
Reference in New Issue
Block a user