LaVGuard: Selective Decoding for Contextual Value Drift

Key points

• A content-safe response can still promote priorities that conflict with its task: an assistant may demand compliance when asked to preserve autonomy.1
• LaVGuard separates an authorized ten-dimensional profile from a frozen estimate of the values expressed in the visible prefix.1
• 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.1
• We establish local surrogate improvement, pathwise budget feasibility, and a bound on response-distribution change.1

Sources (1)

• [1]LaVGuard: Selective Decoding for Contextual Value Drift
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 21, 03:14 AM

  We study contextual value drift as a task-relative decoding problem with an explicit allowance for change.

  A content-safe response can still promote priorities that conflict with its task: an assistant may demand compliance when asked to preserve autonomy.

Extractive summary: sentences quoted from the sources.