20260706:新增一些文章
This commit is contained in:
48
concepts/span-kto.md
Normal file
48
concepts/span-kto.md
Normal file
@@ -0,0 +1,48 @@
|
||||
---
|
||||
title: "Span-KTO"
|
||||
created: 2026-07-02
|
||||
updated: 2026-07-02
|
||||
type: concept
|
||||
tags: [preference-learning, kto, user-feedback, training, alignment]
|
||||
sources:
|
||||
- "[[verification-horizon-no-silver-bullet]]"
|
||||
---
|
||||
|
||||
# Span-KTO
|
||||
|
||||
**Span 级 KTO**(Kahneman-Tversky Optimization),将 KTO 偏好学习从 response 级别扩展到 span 级别,每个 span 对应一个完整用户请求的 agent 回复。
|
||||
|
||||
## 与标准 KTO 的关系
|
||||
|
||||
- **KTO**(Ethayarajh et al., 2024):将前景理论引入 LLM 对齐,用 policy-reference log-likelihood ratio 作为隐式奖励,无需成对偏好数据
|
||||
- **Step-level KTO**:扩展到步骤级,捕获更细粒度反馈
|
||||
- **Span-KTO**:将奖励判断单元定义为 **人类标注 polarity 划分的连续 span**
|
||||
|
||||
## 核心公式
|
||||
|
||||
### Span 隐式奖励
|
||||
|
||||
$$r_\theta(x, S_k) = \sum_{t=s_k}^{e_k} [\log \pi_\theta(y_t|x, y_{<t}) - \log \pi_{ref}(y_t|x, y_{<t})]$$
|
||||
|
||||
### Reference Point(EMA 估计)
|
||||
|
||||
$$z_{ref} \leftarrow \alpha \cdot z_{ref} + (1-\alpha) \cdot \bar{r}_{batch}$$
|
||||
|
||||
### Span 级偏好损失
|
||||
|
||||
对 positive span:$-\lambda_w \cdot \sigma(\beta \cdot a_k)$,对 negative span:$-\lambda_l \cdot \sigma(-\beta \cdot a_k)$,其中 $a_k = r_\theta(x, S_k) - z_{ref}$
|
||||
|
||||
### Neutral Token 正则化
|
||||
|
||||
Neutral token 不参与偏好学习,但保留标准 CE loss 作为正则化。
|
||||
|
||||
## 实验效果
|
||||
|
||||
- 5 个 benchmark 上全面超越 SFT 和 RW-SFT
|
||||
- Aone-bench:+13.3pp(SFT 14.8% → Span-KTO 28.1%)
|
||||
- **关键发现**:不仅"解决更多问题",更重要的是"失败时表现更合理"——Inefficiency 改善 +34.5%,Communication 改善 +26.5%
|
||||
|
||||
## 参考
|
||||
|
||||
- [[verification-horizon-no-silver-bullet|论文原文]]
|
||||
- [[human-implicit-reward-signals|人类隐式奖励信号]]
|
||||
Reference in New Issue
Block a user