Current methods for content safety in Large Language Models (LLMs), such as Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF), often rely on multi-stage training pipelines and lack fine-grained, post-deployment controllability. To address these limitations, we propose a unified co-training framework that efficiently integrates multiple safety behaviors: positive (lawful/prosocial), negative (unfiltered/risk-prone) and rejective (refusal-oriented/conservative) within a single SFT stage. Notably, each behavior is dynamically activated via a simple system-level instruction, or magic token, enabling stealthy and efficient behavioral switching at inference time. This flexibility supports diverse deployment scenarios, such as positive for safe user interaction, negative for internal red-teaming, and rejective for context-aware refusals triggered by upstream moderation signals. This co-training strategy induces a distinct Safety Alignment Margin in the output space, characterized by well-separated response distributions corresponding to each safety mode. The existence of this margin provides empirical evidence for the model's safety robustness and enables unprecedented fine-grained control. Experiments show that our method matches the safety alignment quality of SFT+DPO, with our 8B model notably surpassing DeepSeek-R1 (671B) in safety performance, while significantly reducing both training complexity and deployment costs. This work presents a scalable, efficient, and highly controllable solution for LLM content safety.

Zhewen Tan, Wenhan Yu, Jianfeng Si, Tongxin Liu, Kaiqi Guan, Huiyan Jin, Jiawen Tao, Xiaokun Yuan, Xiangzheng Zhang, Duohe Ma, Tong Yang, Lin Sun. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Identifying aspect-based opinions has been studied extensively in recent years. However, existing work primarily focused on adjective, adverb, and noun expressions. Clearly, verb expressions can imply opinions too. We found that in many domains verb expressions can be even more important to applications because they often describe major issues of products or services. These issues enable brands and businesses to directly improve their products or services. To the best of our knowledge, this problem has not received much attention in the literature. In this paper, we make an attempt to solve this problem. Our proposed method first extracts verb expressions from reviews and then employs Markov Networks to model rich linguistic features and long distance relationships to identify negative issue expressions. Since our training data is obtained from titles of reviews whose labels are automatically inferred from review ratings, our approach is applicable to any domain without manual involvement. Experimental results using real-life review datasets show that our approach outperforms strong baselines.

In this paper we first exploit cash-tags ("$" followed by stocks' ticker symbols) in Twitter to build a stock network, where nodes are stocks connected by edges when two stocks co-occur frequently in tweets.We then employ a labeled topic model to jointly model both the tweets and the network structure to assign each node and each edge a topic respectively.This Semantic Stock Network (SSN) summarizes discussion topics about stocks and stock relations.We further show that social sentiment about stock (node) topics and stock relationship (edge) topics are predictive of each stock's market.For prediction, we propose to regress the topic-sentiment time-series and the stock's price time series.Experimental results demonstrate that topic sentiments from close neighbors are able to help improve the prediction of a stock markedly.