The deployment of Large Language Models (LLMs) in content generation raises significant safety concerns, particularly regarding the transparency and interpretability of content evaluations. Current methods, primarily focused on binary safety classifications, lack mechanisms for detailed critique, limiting their utility for model improvement and user trust. To address these limitations, we introduce SAFETY-J, a bilingual generative safety evaluator for English and Chinese with critique-based judgment. SAFETY-J utilizes a robust training dataset that includes diverse dialogues and augmented query-response pairs to assess safety across various scenarios comprehensively. We establish an automated meta-evaluation benchmark that objectively assesses the quality of critiques with minimal human intervention, facilitating scalable and continuous improvement. Additionally, SAFETY-J employs an iterative preference learning technique to dynamically refine safety assessments based on meta-evaluations and critiques. Our evaluations demonstrate that SAFETY-J provides more nuanced and accurate safety evaluations, thereby enhancing both critique quality and predictive reliability in complex content scenarios. To facilitate further research and application, we open-source SAFETY-J's training protocols, datasets, and code at https://github.com/GAIR-NLP/Safety-J.

We introduce EXAONE 3.0 instruction-tuned language model, the first open model in the family of Large Language Models (LLMs) developed by LG AI Research. Among different model sizes, we publicly release the 7.8B instruction-tuned model to promote open research and innovations. Through extensive evaluations across a wide range of public and in-house benchmarks, EXAONE 3.0 demonstrates highly competitive real-world performance with instruction-following capability against other state-of-the-art open models of similar size. Our comparative analysis shows that EXAONE 3.0 excels particularly in Korean, while achieving compelling performance across general tasks and complex reasoning. With its strong real-world effectiveness and bilingual proficiency, we hope that EXAONE keeps contributing to advancements in Expert AI. Our EXAONE 3.0 instruction-tuned model is available at https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct

The appreciation and utilisation of risk and uncertainty can play a key role in helping to solve some of the many ethical issues that are posed by AI. Understanding the uncertainties can allow algorithms to make better decisions by providing interrogatable avenues to check the correctness of outputs. Allowing algorithms to deal with variability and ambiguity with their inputs means they do not need to force people into uncomfortable classifications. Provenance enables algorithms to know what they know preventing possible harms. Additionally, uncertainty about provenance highlights the trustworthiness of algorithms. It is essential to compute with what we know rather than make assumptions that may be unjustified or untenable. This paper provides a perspective on the need for the importance of risk and uncertainty in the development of ethical AI, especially in high-risk scenarios. It argues that the handling of uncertainty, especially epistemic uncertainty, is critical to ensuring that algorithms do not cause harm and are trustworthy and ensure that the decisions that they make are humane.

Large Language Models (LLMs) have demonstrated strong reasoning and memorization capabilities via pretraining on massive textual corpora. However, training LLMs on human-written text entails significant risk of privacy and copyright violations, which demands an efficient machine unlearning framework to remove knowledge of sensitive data without retraining the model from scratch. While Gradient Ascent (GA) is widely used for unlearning by reducing the likelihood of generating unwanted information, the unboundedness of increasing the cross-entropy loss causes not only unstable optimization, but also catastrophic forgetting of knowledge that needs to be retained. We also discover its joint application under low-rank adaptation results in significantly suboptimal computational cost vs. generative performance trade-offs. In light of this limitation, we propose two novel techniques for robust and cost-efficient unlearning on LLMs. We first design an Inverted Hinge loss that suppresses unwanted tokens by increasing the probability of the next most likely token, thereby retaining fluency and structure in language generation. We also propose to initialize low-rank adapter weights based on Fisher-weighted low-rank approximation, which induces faster unlearning and better knowledge retention by allowing model updates to be focused on parameters that are important in generating textual data we wish to remove.

Conceptual Engineers want to make words better. However, they often underestimate how varied our usage of words is. In this paper, we take the first steps in exploring the contextual nuances of words by creating conceptual landscapes -- 2D surfaces representing the pragmatic usage of words -- that conceptual engineers can use to inform their projects. We use the spoken component of the British National Corpus and BERT to create contextualised word embeddings, and use Gaussian Mixture Models, a selection of metrics, and qualitative analysis to visualise and numerically represent lexical landscapes. Such an approach has not yet been used in the conceptual engineering literature and provides a detailed examination of how different words manifest in various contexts that is potentially useful to conceptual engineering projects. Our findings highlight the inherent complexity of conceptual engineering, revealing that each word exhibits a unique and intricate landscape. Conceptual Engineers cannot, therefore, use a one-size-fits-all approach when improving words -- a task that may be practically intractable at scale.

As large language models (LLMs) continue to scale, their enhanced performance often proves insufficient for solving domain-specific tasks. Systematically analyzing their failures and effectively enhancing their performance remain significant challenges. This paper introduces the Re-TASK framework, a novel theoretical model that Revisits LLM Tasks from cApability, Skill, Knowledge perspectives, guided by the principles of Bloom's Taxonomy and Knowledge Space Theory. The Re-TASK framework provides a systematic methodology to deepen our understanding, evaluation, and enhancement of LLMs for domain-specific tasks. It explores the interplay among an LLM's capabilities, the knowledge it processes, and the skills it applies, elucidating how these elements are interconnected and impact task performance. Our application of the Re-TASK framework reveals that many failures in domain-specific tasks can be attributed to insufficient knowledge or inadequate skill adaptation. With this insight, we propose structured strategies for enhancing LLMs through targeted knowledge injection and skill adaptation. Specifically, we identify key capability items associated with tasks and employ a deliberately designed prompting strategy to enhance task performance, thereby reducing the need for extensive fine-tuning. Alternatively, we fine-tune the LLM using capability-specific instructions, further validating the efficacy of our framework. Experimental results confirm the framework's effectiveness, demonstrating substantial improvements in both the performance and applicability of LLMs.

Minimizing negative impacts of Artificial Intelligent (AI) systems on human societies without human supervision requires them to be able to align with human values. However, most current work only addresses this issue from a technical point of view, e.g., improving current methods relying on reinforcement learning from human feedback, neglecting what it means and is required for alignment to occur. Here, we propose to distinguish strong and weak value alignment. Strong alignment requires cognitive abilities (either human-like or different from humans) such as understanding and reasoning about agents' intentions and their ability to causally produce desired effects. We argue that this is required for AI systems like large language models (LLMs) to be able to recognize situations presenting a risk that human values may be flouted. To illustrate this distinction, we present a series of prompts showing ChatGPT's, Gemini's and Copilot's failures to recognize some of these situations. We moreover analyze word embeddings to show that the nearest neighbors of some human values in LLMs differ from humans' semantic representations. We then propose a new thought experiment that we call "the Chinese room with a word transition dictionary", in extension of John Searle's famous proposal. We finally mention current promising research directions towards a weak alignment, which could produce statistically satisfying answers in a number of common situations, however so far without ensuring any truth value.

With the growing demand for offline PDF chatbots in automotive industrial production environments, optimizing the deployment of large language models (LLMs) in local, low-performance settings has become increasingly important. This study focuses on enhancing Retrieval-Augmented Generation (RAG) techniques for processing complex automotive industry documents using locally deployed Ollama models. Based on the Langchain framework, we propose a multi-dimensional optimization approach for Ollama's local RAG implementation. Our method addresses key challenges in automotive document processing, including multi-column layouts and technical specifications. We introduce improvements in PDF processing, retrieval mechanisms, and context compression, tailored to the unique characteristics of automotive industry documents. Additionally, we design custom classes supporting embedding pipelines and an agent supporting self-RAG based on LangGraph best practices. To evaluate our approach, we constructed a proprietary dataset comprising typical automotive industry documents, including technical reports and corporate regulations. We compared our optimized RAG model and self-RAG agent against a naive RAG baseline across three datasets: our automotive industry dataset, QReCC, and CoQA. Results demonstrate significant improvements in context precision, context recall, answer relevancy, and faithfulness, with particularly notable performance on the automotive industry dataset. Our optimization scheme provides an effective solution for deploying local RAG systems in the automotive sector, addressing the specific needs of PDF chatbots in industrial production environments. This research has important implications for advancing information processing and intelligent production in the automotive industry.

Large Language Models (LLMs) are known for their remarkable ability to generate synthesized 'knowledge', such as text documents, music, images, etc. However, there is a huge gap between LLM's and human capabilities for understanding abstract concepts and reasoning. We discuss these issues in a larger philosophical context of human knowledge acquisition and the Turing test. In addition, we illustrate the limitations of LLMs by analyzing GPT-4 responses to questions ranging from science and math to common sense reasoning. These examples show that GPT-4 can often imitate human reasoning, even though it lacks understanding. However, LLM responses are synthesized from a large LLM model trained on all available data. In contrast, human understanding is based on a small number of abstract concepts. Based on this distinction, we discuss the impact of LLMs on acquisition of human knowledge and education.

Hate speech is a pressing issue in modern society, with significant effects both online and offline. Recent research in hate speech detection has primarily centered on text-based media, largely overlooking multimodal content such as videos. Existing studies on hateful video datasets have predominantly focused on English content within a Western context and have been limited to binary labels (hateful or non-hateful), lacking detailed contextual information. This study presents MultiHateClip1 , an novel multilingual dataset created through hate lexicons and human annotation. It aims to enhance the detection of hateful videos on platforms such as YouTube and Bilibili, including content in both English and Chinese languages. Comprising 2,000 videos annotated for hatefulness, offensiveness, and normalcy, this dataset provides a cross-cultural perspective on gender-based hate speech. Through a detailed examination of human annotation results, we discuss the differences between Chinese and English hateful videos and underscore the importance of different modalities in hateful and offensive video analysis. Evaluations of state-of-the-art video classification models, such as VLM, GPT-4V and Qwen-VL, on MultiHateClip highlight the existing challenges in accurately distinguishing between hateful and offensive content and the urgent need for models that are both multimodally and culturally nuanced. MultiHateClip represents a foundational advance in enhancing hateful video detection by underscoring the necessity of a multimodal and culturally sensitive approach in combating online hate speech.