arXiv cs.CL· Jeremias Ferrao, Niclas M\"uller-Hof, Iustin S\^irbu, Traian Rebedea, Yftah Ziser·· 12 小时前AI 评分31
让安全分类器显式建模用户意图：AIMS 数据集与意图感知训练提升 LLM 安全分类
Paved with True Intents: Intent-Aware Training Improves LLM Safety Classification Across Training Regimes
AI 导读
研究者提出 AIMS——一个含 1,724 条困难安全提示、每条配意图描述与危害标签的人工标注数据集，主张安全分类器应把用户意图作为提示词与最终标签之间的显式信号。
正文 · 原文
Abstract:We argue that safety classifiers should model user intent as an explicit signal between the prompt and the final label. To study this, we introduce AIMS, a human-annotated dataset of 1,724 difficult safety prompts, each paired with an intent description and harm label. We use AIMS to evaluate intent-aware training across supervised fine-tuning, preference learning, reasoning distillation, and reinforcement learning. Despite its size, AIMS enables competitive safety classifiers across training regimes: DPO from model-generated intent errors improves over SFT, and intent-conditioned distillation outperforms reasoning-only distillation in most teacher-student pairs. Most notably, directly rewarding intent faithfulness with GRPO yields the strongest average performance across five external safety benchmarks, while our intent-aware models form the inference latency-F1 Pareto frontier. These results show that faithful intent modeling is a compact, high-quality supervision signal for more robust safety classifiers.
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2606.27210 [cs.CL] |
| (or arXiv:2606.27210v2 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2606.27210 arXiv-issued DOI via DataCite |
Submission history
From: Yftah Ziser [view email]
[v1]
Thu, 25 Jun 2026 16:03:57 UTC (154 KB)
[v2]
Fri, 16 Oct 2026 16:09:43 UTC (155 KB)
来源：arXiv cs.CL · arxiv.org