The softmax activation function plays a crucial role in the success of large language models (LLMs), particularly in the self-attention mechanism of the widely adopted Transformer architecture. However, the underlying learning dynamics that contribute to the effectiveness of softmax remain largely unexplored. As a step towards better understanding, this paper provides a theoretical study of the optimization and generalization properties of two-layer softmax neural networks, providing theoretical insights into their superior performance as other activation functions, such as ReLU and exponential. Leveraging the Neural Tangent Kernel (NTK) framework, our analysis reveals that the normalization effect of the softmax function leads to a good perturbation property of the induced NTK matrix, resulting in a good convex region of the loss landscape. Consequently, softmax neural networks can learn the target function in the over-parametrization regime. To demonstrate the broad applicability of our theoretical findings, we apply them to the task of learning score estimation functions in diffusion models, a promising approach for generative modeling. Our analysis shows that gradient-based algorithms can learn the score function with a provable accuracy. Our work provides a deeper understanding of the effectiveness of softmax neural networks and their potential in various domains, paving the way for further advancements in natural language processing and beyond.

Despite superior reasoning prowess demonstrated by Large Language Models (LLMs) with Chain-of-Thought (CoT) prompting, a lack of understanding prevails around the internal mechanisms of the models that facilitate CoT generation. This work investigates the neural sub-structures within LLMs that manifest CoT reasoning from a mechanistic point of view. From an analysis of Llama-2 7B applied to multistep reasoning over fictional ontologies, we demonstrate that LLMs deploy multiple parallel pathways of answer generation for step-by-step reasoning. These parallel pathways provide sequential answers from the input question context as well as the generated CoT. We observe a functional rift in the middle layers of the LLM. Token representations in the initial half remain strongly biased towards the pretraining prior, with the in-context prior taking over in the later half. This internal phase shift manifests in different functional components: attention heads that write the answer token appear in the later half, attention heads that move information along ontological relationships appear in the initial half, and so on. To the best of our knowledge, this is the first attempt towards mechanistic investigation of CoT reasoning in LLMs.

Integrating large language models (LLMs) and knowledge graphs (KGs) holds great promise for revolutionizing intelligent education, but challenges remain in achieving personalization, interactivity, and explainability. We propose FOKE, a Forest Of Knowledge and Education framework that synergizes foundation models, knowledge graphs, and prompt engineering to address these challenges. FOKE introduces three key innovations: (1) a hierarchical knowledge forest for structured domain knowledge representation; (2) a multi-dimensional user profiling mechanism for comprehensive learner modeling; and (3) an interactive prompt engineering scheme for generating precise and tailored learning guidance. We showcase FOKE's application in programming education, homework assessment, and learning path planning, demonstrating its effectiveness and practicality. Additionally, we implement Scholar Hero, a real-world instantiation of FOKE. Our research highlights the potential of integrating foundation models, knowledge graphs, and prompt engineering to revolutionize intelligent education practices, ultimately benefiting learners worldwide. FOKE provides a principled and unified approach to harnessing cutting-edge AI technologies for personalized, interactive, and explainable educational services, paving the way for further research and development in this critical direction.

In recent years, Large Language Models (LLMs) have become integrated into our daily lives, serving as invaluable assistants in completing tasks. Widely embraced by users, the abuse of LLMs is inevitable, particularly in using them to generate text content for various purposes, leading to difficulties in distinguishing between text generated by LLMs and that written by humans. In this study, we present a dataset named ViDetect, comprising 6.800 samples of Vietnamese essay, with 3.400 samples authored by humans and the remainder generated by LLMs, serving the purpose of detecting text generated by AI. We conducted evaluations using state-of-the-art methods, including ViT5, BartPho, PhoBERT, mDeberta V3, and mBERT. These results contribute not only to the growing body of research on detecting text generated by AI but also demonstrate the adaptability and effectiveness of different methods in the Vietnamese language context. This research lays the foundation for future advancements in AI-generated text detection and provides valuable insights for researchers in the field of natural language processing.

The conversational recommendation system (CRS) has been criticized regarding its user experience in real-world scenarios, despite recent significant progress achieved in academia. Existing evaluation protocols for CRS may prioritize system-centric factors such as effectiveness and fluency in conversation while neglecting user-centric aspects. Thus, we propose a new and inclusive evaluation protocol, Concept, which integrates both system- and user-centric factors. We conceptualise three key characteristics in representing such factors and further divide them into six primary abilities. To implement Concept, we adopt a LLM-based user simulator and evaluator with scoring rubrics that are tailored for each primary ability. Our protocol, Concept, serves a dual purpose. First, it provides an overview of the pros and cons in current CRS models. Second, it pinpoints the problem of low usability in the "omnipotent" ChatGPT and offers a comprehensive reference guide for evaluating CRS, thereby setting the foundation for CRS improvement.

Robots are the future of every technology where every advanced technology eventually will be used to make robots which are more efficient. The major challenge today is to train the robots exactly and empathetically using knowledge representation. This paper gives you insights of how we can use unstructured knowledge representation and convert them to meaningful structured representation with the help of prompt engineering which can be eventually used in the robots to make help them understand how human brain can make wonders with the minimal data or objects can providing to them.

Large Language Models (LLMs) have demonstrated unparalleled effectiveness in various NLP tasks, and integrating LLMs with automatic speech recognition (ASR) is becoming a mainstream paradigm. Building upon this momentum, our research delves into an in-depth examination of this paradigm on a large open-source Chinese dataset. Specifically, our research aims to evaluate the impact of various configurations of speech encoders, LLMs, and projector modules in the context of the speech foundation encoder-LLM ASR paradigm. Furthermore, we introduce a three-stage training approach, expressly developed to enhance the model's ability to align auditory and textual information. The implementation of this approach, alongside the strategic integration of ASR components, enabled us to achieve the SOTA performance on the AISHELL-1, Test_Net, and Test_Meeting test sets. Our analysis presents an empirical foundation for future research in LLM-based ASR systems and offers insights into optimizing performance using Chinese datasets. We will publicly release all scripts used for data preparation, training, inference, and scoring, as well as pre-trained models and training logs to promote reproducible research.

Academic researchers and social media entities grappling with the identification of hate speech face significant challenges, primarily due to the vast scale of data and the dynamic nature of hate speech. Given the ethical and practical limitations of large predictive models like ChatGPT in directly addressing such sensitive issues, our research has explored alternative advanced transformer-based and generative AI technologies since 2019. Specifically, we developed a new data labeling technique and established a proof of concept targeting anti-Semitic hate speech, utilizing a variety of transformer models such as BERT (arXiv:1810.04805), DistillBERT (arXiv:1910.01108), RoBERTa (arXiv:1907.11692), and LLaMA-2 (arXiv:2307.09288), complemented by the LoRA fine-tuning approach (arXiv:2106.09685). This paper delineates and evaluates the comparative efficacy of these cutting-edge methods in tackling the intricacies of hate speech detection, highlighting the need for responsible and carefully managed AI applications within sensitive contexts.

Context. Nowadays, 83% of software developers use Large Language Models (LLMs) to generate code. LLMs recently became essential to increase the productivity of software developers and decrease the time and cost of software development. Developers ranging from novices to experts use LLM tools not only to detect and patch bugs, but also to integrate generated code into their software. However, as of today there is no objective assessment of the energy efficiency of the source code generated by LLM tools. Released in August 2023, Code Llama is one of the most recent LLM tools. Goal. In this paper, we present an empirical study that assesses the energy efficiency of Code Llama with respect to human-written source code. Method. We design an experiment involving three human-written benchmarks implemented in C++, JavaScript, and Python. We ask Code Llama to generate the code of the benchmarks using different prompts and temperatures. Therefore, we execute both implementations and profile their energy efficiency. Results. Our study shows that the energy efficiency of code generated by Code Llama is heavily-dependent on the chosen programming language and the specific code problem at hand. Also, human implementations tend to be more energy efficient overall, with generated JavaScript code outperforming its human counterpart. Moreover, explicitly asking Code Llama to generate energy-efficient code results in an equal or worse energy efficiency, as well as using different temperatures seems not to affect the energy efficiency of generated code. Conclusions. According to our results, code generated using Code Llama does not guarantee energy efficiency, even when prompted to do so. Therefore, software developers should evaluate the energy efficiency of generated code before integrating it into the software system under development.

Multi-modal large language models (MLLMs) have shown incredible capabilities in a variety of 2D vision and language tasks. We extend MLLMs' perceptual capabilities to ground and reason about images in 3-dimensional space. To that end, we first develop a large-scale pre-training dataset for 2D and 3D called LV3D by combining multiple existing 2D and 3D recognition datasets under a common task formulation: as multi-turn question-answering. Next, we introduce a new MLLM named Cube-LLM and pre-train it on LV3D. We show that pure data scaling makes a strong 3D perception capability without 3D specific architectural design or training objective. Cube-LLM exhibits intriguing properties similar to LLMs: (1) Cube-LLM can apply chain-of-thought prompting to improve 3D understanding from 2D context information. (2) Cube-LLM can follow complex and diverse instructions and adapt to versatile input and output formats. (3) Cube-LLM can be visually prompted such as 2D box or a set of candidate 3D boxes from specialists. Our experiments on outdoor benchmarks demonstrate that Cube-LLM significantly outperforms existing baselines by 21.3 points of AP-BEV on the Talk2Car dataset for 3D grounded reasoning and 17.7 points on the DriveLM dataset for complex reasoning about driving scenarios, respectively. Cube-LLM also shows competitive results in general MLLM benchmarks such as refCOCO for 2D grounding with (87.0) average score, as well as visual question answering benchmarks such as VQAv2, GQA, SQA, POPE, etc. for complex reasoning. Our project is available at https://janghyuncho.github.io/Cube-LLM.