As large language models (LLMs) see wider real-world use, understanding and mitigating their unsafe behaviors is critical. Interpretation techniques can reveal causes of unsafe outputs and guide safety, but such connections with safety are often overlooked in prior surveys. We present the first survey that bridges this gap, introducing a unified framework that connects safety-focused interpretation methods, the safety enhancements they inform, and the tools that operationalize them. Our novel taxonomy, organized by LLM workflow stages, summarizes nearly 70 works at their intersections. We conclude with open challenges and future directions. This timely survey helps researchers and practitioners navigate key advancements for safer, more interpretable LLMs.

Large language models (LLMs) undergo extensive safety training to maximize both helpfulness and harmlessness in their responses. However, various jailbreak attacks jeopardize model safety, allowing malicious actors to bypass safety guidelines. Existing defense methods primarily focus on aligning the model’s output towards less harmful responses through post-processing or input perturbation. Consequently, these approaches are prone to general performance degradation and lack the ability to defend against a wide variety of attacks. In this paper, we propose goal-conditioned direct preference optimization (GC-DPO), which is trained to prioritize the system prompt over the user prompt through goal-conditioning, and thus enables a good balance between safety and performance. Empirically, we show that our approach significantly reduces the average Attack Success Rate (ASR) on a wide variety of jailbreak attacks. In particular, GC-DPO achieves a reduction of 67.1% to 5.0% in ASR for Vicuna-7B, a state-of-the-art result, without compromising the model’s general performance.

Recent studies have determined that the learned token embeddings of large-scale neural language models are degenerated to be anisotropic with a narrow-cone shape. This phenomenon, called the representation degeneration problem, facilitates an increase in the overall similarity between token embeddings that negatively affect the performance of the models. Although the existing methods that address the degeneration problem based on observations of the phenomenon triggered by the problem improves the performance of the text generation, the training dynamics of token embeddings behind the degeneration problem are still not explored. In this study, we analyze the training dynamics of the token embeddings focusing on rare token embedding. We demonstrate that the specific part of the gradient for rare token embeddings is the key cause of the degeneration problem for all tokens during training stage. Based on the analysis, we propose a novel method called, adaptive gradient gating(AGG). AGG addresses the degeneration problem by gating the specific part of the gradient for rare token embeddings. Experimental results from language modeling, word similarity, and machine translation tasks quantitatively and qualitatively verify the effectiveness of AGG.

In this paper, we describe the neural machine translation (NMT) system submitted by the Kangwon National University and HYUNDAI (KNU-HYUNDAI) team to the translation tasks of the 6th workshop on Asian Translation (WAT 2019). We participated in all tasks of ASPEC and JPC2, which included those of Chinese-Japanese, English-Japanese, and Korean->Japanese. We submitted our transformer-based NMT system with built using the following methods: a) relative positioning method for pairwise relationships between the input elements, b) back-translation and multi-source translation for data augmentation, c) right-to-left (r2l)-reranking model robust against error propagation in autoregressive architectures such as decoders, and d) checkpoint ensemble models, which selected the top three models with the best validation bilingual evaluation understudy (BLEU) . We have reported the translation results on the two aforementioned tasks. We performed well in both the tasks and were ranked first in terms of the BLEU scores in all the JPC2 subtasks we participated in.