20260706:新增一些文章

This commit is contained in:
2026-07-06 10:14:02 +08:00
parent 6021dea160
commit 24b006225b
194 changed files with 8512 additions and 91 deletions

48
concepts/span-kto.md Normal file
View 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 PointEMA 估计)
$$z_{ref} \leftarrow \alpha \cdot z_{ref} + (1-\alpha) \cdot \bar{r}_{batch}$$
### Span 级偏好损失
positive span$-\lambda_w \cdot \sigma(\beta \cdot a_k)$ 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 benchmark 上全面超越 SFT RW-SFT
- Aone-bench+13.3ppSFT 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]]