arXiv cs.AI· Yunhao Feng, Yifan Ding, Yuxiang Xie, Zheng Li, Mingrui Lao, Zeyuan Wang, Yanming Guo·· 13 小时前AI 评分38
AdaGuard：支持用户自定义策略的自适应防护模型
AdaGuard: An Adaptive Guard Model with User-defined Policies
AI 导读
研究团队提出 AdaGuard 系列防护模型（0.6B、4B、8B），可在推理时按用户自定义策略评估语言模型智能体的行为轨迹。配套发布 AdaptiveSafety 数据集，含 10,939 条训练样本和 1,000 条测试样本，覆盖 1–100 条规则。4B 模型在 AdaptiveSafety 上二分类准确率达 89.30%，在 DynaBench 上为 71.82%。
正文 · 原文
Abstract:Guard models support the safe deployment of language model agents, but fixed risk taxonomies limit their ability to accommodate requirements that vary across applications and tasks. Under user-defined policies, detecting violations requires interpreting both the applicable rules and the agent's behavior, since identical actions can receive different judgments under different policies. To support learning this capability, we introduce AdaptiveSafety, a dataset of 10,939 training examples and 1,000 test examples covering policies with 1--100 rules. The dataset combines trajectories from multiple sources with policy and behavioral counterfactuals, pairing each example with an explanation and the complete set of violated rules. These counterfactuals expose changes that alter compliance, while structural augmentations provide supervision for consistency under rule reordering and identifier remapping. Building on this supervision, we propose SafePO, a reinforcement learning algorithm for refining violation identification while balancing explanatory reasoning and final verdicts. SafePO uses structured rewards to assess prediction correctness, retains group-relative advantages at the response level, and employs a separately trained value model to modulate token weights within explanation and verdict regions. Separate normalization controls their relative contribution to training despite differences in length. Through supervised initialization followed by SafePO, we develop AdaGuard, a family of 0.6B, 4B, and 8B guard models that assess agent trajectories under policies supplied at inference time. Our 4B model achieves binary accuracies of 89.30\% on AdaptiveSafety and 71.82\% on DynaBench. The project repository is available at this https URL
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2609.34241 [cs.AI] |
| (or arXiv:2609.34241v2 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2609.34241 arXiv-issued DOI via DataCite |
Submission history
From: Yunhao Feng [view email]
[v1]
Mon, 28 Sep 2026 03:48:12 UTC (484 KB)
[v2]
Wed, 21 Oct 2026 09:34:41 UTC (484 KB)
来源：arXiv cs.AI · arxiv.org