跳到正文
研踪PaperTrail
← 研踪 PaperTrail

arXiv:2609.30780 · v2

Shielded Trajectory Sampling for Personalized Relational Safety in Conversational AI

首次提交 2026-09-25 12:04
版本更新 2026-11-17 20:13
雷达观测 2026-11-18 11:33
arXiv ↗PDF ↗

摘要

Anthropomorphic conversational systems can provide companionship and emotional support, but repeated, seemingly harmless interactions can encourage excessive dependence and reduced human contact. We formulate safety in user-AI relationships as a sequential decision problem, using conversational cues to track the relationship and selecting responses according to their anticipated effects over time. A conditional Generative Flow Network learns from sampled response plans to select replies that preserve the largest estimated mass of safe and useful continuations. Model-based checks screen plans during training and responses during interaction, requiring at least one sequence of future response strategies that meets a prescribed safety bound under every candidate user model, independently of the learned network. Strategies are selected directly from learned first-step probabilities, avoiding online plan sampling. We evaluate the approach in simulation against reactive safeguards, filters with and without memory, and alternative planners, using exact computation to validate the learned approximation. We also evaluate relational cue extraction from real user messages in WildChat. To our knowledge, this is the first approach to select conversational replies by anticipating their effects on the user-AI relationship while imposing an explicit model-based safety bound on future response sequences. Results show close agreement with exact planning, lower decision costs at longer horizons, and improved safety-utility trade-offs when protective responses are less effective. The approach balances immediate support with the longer-term effects of repeated interaction. The source code is available at https://github.com/benedettapicano/ANTHROPOMORPHIC_SAFETY_TRAJ.

AI 论文解读

基于摘要 · 非全文审阅

此版本尚未生成 AI 解读。解读由后台任务生成并按论文版本保存,登录后共享已有结果。

版本差异

本站首次收录为 v2,暂无更早版本可对照。

观测到的版本

相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。

v2 · 2026-11-17 20:13 · 最新观测版本

Shielded Trajectory Sampling for Personalized Relational Safety in Conversational AI

Anthropomorphic conversational systems can provide companionship and emotional support, but repeated, seemingly harmless interactions can encourage excessive dependence and reduced human contact. We formulate safety in user-AI relationships as a sequential decision problem, using conversational cues to track the relationship and selecting responses according to their anticipated effects over time. A conditional Generative Flow Network learns from sampled response plans to select replies that preserve the largest estimated mass of safe and useful continuations. Model-based checks screen plans during training and responses during interaction, requiring at least one sequence of future response strategies that meets a prescribed safety bound under every candidate user model, independently of the learned network. Strategies are selected directly from learned first-step probabilities, avoiding online plan sampling. We evaluate the approach in simulation against reactive safeguards, filters with and without memory, and alternative planners, using exact computation to validate the learned approximation. We also evaluate relational cue extraction from real user messages in WildChat. To our knowledge, this is the first approach to select conversational replies by anticipating their effects on the user-AI relationship while imposing an explicit model-based safety bound on future response sequences. Results show close agreement with exact planning, lower decision costs at longer horizons, and improved safety-utility trade-offs when protective responses are less effective. The approach balances immediate support with the longer-term effects of repeated interaction. The source code is available at https://github.com/benedettapicano/ANTHROPOMORPHIC_SAFETY_TRAJ.

查看此版本解读 →查看此版本原文 ↗