arXiv cs.AI· Jiale Dai, Liuxian Ma, Hongcan Deng, Xiaoke Niu, Wenjing Zhang, Zhaoxiang Liu, Shiguo Lian, Guojie Song·· 13 小时前AI 评分31
LaVGuard：面向上下文价值漂移的选择性解码方法
LaVGuard: Selective Decoding for Contextual Value Drift
AI 导读
LaVGuard 将上下文价值漂移建模为任务相关的解码问题，通过分离授权十维价值画像与可见前缀的冻结价值估计，用画像相对密度门控决定何时评估候选续写，并以改变位置与累积 KL 预算约束 logit 调整。在 5% FPR 下，画像条件化将边界 TPR 从 .58 提升至 .72；在相同 .12 平均干预比例下，LaVGuard 的风险安全得分为 .88，位置匹配编辑为 .72。
正文 · 原文
Abstract:A content-safe response can still promote priorities that conflict with its task: an assistant may demand compliance when asked to preserve autonomy. We study contextual value drift as a task-relative decoding problem with an explicit allowance for change. LaVGuard separates an authorized ten-dimensional profile from a frozen estimate of the values expressed in the visible prefix. A profile-relative density gate selects when to evaluate candidate continuations; their signed progress determines a logit tilt constrained by both changed-position and cumulative-KL budgets. We establish local surrogate improvement, pathwise budget feasibility, and a bound on response-distribution change. Profile conditioning raises boundary TPR at 5% FPR from .58 to .72. At the same .12 mean intervention ratio, the selection study reports risk-safety scores of .88 for LaVGuard and .72 for position-matched edits. The resulting interface makes task priorities actionable while keeping the extent of intervention explicit and bounded.
| Comments: | 37 pages, 5 figures, 16 tables, including appendices |
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.29135 [cs.AI] |
| (or arXiv:2610.29135v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.29135 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Jiale Dai [view email]
[v1]
Wed, 21 Oct 2026 03:14:15 UTC (400 KB)
来源：arXiv cs.AI · arxiv.org