The growing awareness of safety concerns in large language models (LLMs) has sparked considerable interest in the evaluation of safety. This study investigates an under-explored issue about the evaluation of LLMs, namely the substantial discrepancy in performance between multiple-choice questions and open-ended questions. Inspired by research on jailbreak attack patterns, we argue this is caused by mismatched generalization. That is, LLM only remembers the answer style for open-ended safety questions, which makes it unable to solve other forms of safety tests. We refer to this phenomenon as fake alignment and construct a comparative benchmark to empirically verify its existence in LLMs. We introduce a Fake alIgNment Evaluation (FINE) framework and two novel metrics——Consistency Score (CS) and Consistent Safety Score (CSS), which jointly assess two complementary forms of evaluation to quantify fake alignment and obtain corrected performance estimation. Applying FINE to 14 widely-used LLMs reveals several models with purported safety are poorly aligned in practice. Subsequently, we found that multiple-choice format data can also be used as high-quality contrast distillation-based fine-tuning data, which can strongly improve the alignment consistency of LLMs with minimal fine-tuning overhead. For data and code, see https://github.com/AIFlames/Fake-Alignment.

The task of adopting a model with good performance to a target domain that is different from the source domain used for training has received considerable attention in sentiment analysis. Most existing approaches mainly focus on learning representations that are domain-invariant in both the source and target domains. Few of them pay attention to domain-specific information, which should also be informative. In this work, we propose a method to simultaneously extract domain specific and invariant representations and train a classifier on each of the representation, respectively. And we introduce a few target domain labeled data for learning domain-specific information. To effectively utilize the target domain labeled data, we train the domain invariant representation based classifier with both the source and target domain labeled data and train the domain-specific representation based classifier with only the target domain labeled data. These two classifiers then boost each other in a co-training style. Extensive sentiment analysis experiments demonstrated that the proposed method could achieve better performance than state-of-the-art methods.