Large Language Models Research Articles

Generative AI

Generative AI is developing at an exponential pace, showing immense potential for reshaping teaching and learning methodologies. Designing high-quality examinations demands careful consideration, such as incorporating diverse question types to assess different levels of understanding and ensuring alignment with learning outcomes. However, many university educators encounter challenges in designing effective examinations due to insufficient training. This paper outlines our work-in-progress in leveraging generative AI, specifically large language models, as a copilot in examination design. In this study, we are evaluating large language models’ performance in assessing examination questions based on four key quality criteria. Preliminary findings indicate that ChatGPT 3.5 achieved comparable performance to the state-of-the-art in classifying questions by cognitive complexity, as defined by Bloom’s taxonomy. Additionally, ChatGPT 3.5 demonstrated promising initial results across other criteria. In conclusion, generative AI shows substantial potential to serve as a valuable tool in assisting university educators in enhancing the overall quality of examination design.

Possible Productivity Effects On Software Engineers by Advanced Artificial Intelligence

Within the past few years, starting from the greater public use of AI from the recent “AI Boom,” ChatGPT or AI-Language Model equivalents have been making their way into software and other computer science-related work environments for developers and software engineers to use without significant financial cost. In this paper, we often mention the word “productivity,” so it is important to know how we measure this: we measure productivity in lines of code (LOC) to gauge the raw amount of coding done, bug resolutions done by developers to measure the reviewing of code, and customer satisfaction to measure the quality of the code, then combine all of these into “overall productivity.” In this paper, we will examine the effects on productivity that these Large Language Models (LLMs) have had on software engineering or other similar jobs.

Beyond Binary Dialogues: Research and Development of a Linguistically Nuanced Conversation Design for Social Robots in Group–Robot Interactions

In this paper, we detail the technical development of a conversation design that is sensitive to group dynamics and adaptable, taking into account the subtleties of linguistic variations between dyadic (i.e., one human and one agent) and group interactions in human–robot interaction (HRI) using the German language as a case study. The paper details the implementation of robust person and group detection with YOLOv5m and the expansion of knowledge databases using large language models (LLMs) to create adaptive multi-party interactions (MPIs) (i.e., group–robot interactions (GRIs)). We describe the use of LLMs to generate training data for socially interactive agents including social robots, as well as a self-developed synthesis tool, knowledge expander, to accurately map the diverse needs of different users in public spaces. We also outline the integration of a LLM as a fallback for open-ended questions not covered by our knowledge database, ensuring it can effectively respond to both individuals and groups within the MPI framework.

Using Large Language Models to Enhance Genre-Based English for Academic Purposes (EAP) Writing: A Path to Sustainable Development

In the context of the evolving digital era, this study introduces a conceptual model that integrates Education for Sustainable Development (ESD) into English for Academic Purposes (EAP) writing, leveraging Critical Digital Literacies (CDL). It aims to develop EFL doctoral students' competencies necessary for sustainable development, addressing the gap in current genre-based EAP instruction that often neglects comprehensive ESD competencies such as systems thinking and self-awareness. Utilizing innovative tools like ChatGPT within a CDL framework, the model focuses on both explicit teaching objectives related to genre knowledge and implicit ones tied to ESD. Specifically, the study outlines eight scenarios where GPT technology aids in achieving these pedagogical goals. By implementing this model, we aim not only to enrich the EAP writing curriculum but also to subtly shift educational practices towards embracing the values of sustainability and digitalization.

Unboxing Occupational Bias: Debiasing LLMs with U.S. Labor Data

Large Language Models (LLMs) are prone to inheriting and amplifying societal biases embedded within their training data, potentially reinforcing harmful stereotypes related to gender, occupation, and other sensitive categories. This issue becomes particularly problematic as biased LLMs can have far-reaching consequences, leading to unfair practices and exacerbating social inequalities across various domains, such as recruitment, online content moderation, or even the criminal justice system. Although prior research has focused on detecting bias in LLMs using specialized datasets designed to highlight intrinsic biases, there has been a notable lack of investigation into how these findings correlate with authoritative datasets, such as those from the U.S. National Bureau of Labor Statistics (NBLS). To address this gap, we conduct empirical research that evaluates LLMs in a “bias-out-of-the-box” setting, analyzing how the generated outputs compare with the distributions found in NBLS data. Furthermore, we propose a straightforward yet effective debiasing mechanism that directly incorporates NBLS instances to mitigate bias within LLMs. Our study spans seven different LLMs, including instructable, base, and mixture-of-expert models, and reveals significant levels of bias that are often overlooked by existing bias detection techniques. Importantly, our debiasing method, which does not rely on external datasets, demonstrates a substantial reduction in bias scores, highlighting the efficacy of our approach in creating fairer and more reliable LLMs.

Proposed Explainable Interference Control Technique in 6G Networks Using Large Language Models (LLMs)

After the advent of 5th generation (5G) and 6th generation (6G) cellular networks, the complexity of managing real-time signal interference has increased in dense and dynamic environments. Traditional interference techniques, such as frequency reuse and allocation, while effective, lack robust adaptability and transparency needed to reduce interference in advanced communication networks. This paper introduces a novel approach that fuses large language models (LLMs) and Explainable Artificial Intelligence (XAI) to mitigate interference and enhance interference management in the mathematical foundations of 6G networks. The proposed approach provides accurate interference predictions, which the LLM balances with its complex architecture, necessary to meet the demands of beyond 5G and 6G networks, along with interpretable explanations to ensure transparency in decision-making. The proposed framework has been evaluated across various performance metrics. Interference latency consistently achieves lower rates of 0.95 s, compared to traditional techniques, which average around 1 s. Furthermore, the confidence score of the LLM shows a stable value of 0.87 throughout the system, compared to 0.85 in techniques without LLMs. Overall, the XAI-driven LLM demonstrates the potential of incorporating LLMs and XAI into wireless networks to improve resilience in next-generation networks. This proof of concept introduces a novel framework that offers new dimensions in wireless communication, particularly for interference management, prediction, and mitigation.

Wisdom of the silicon crowd: LLM ensemble prediction capabilities rival human crowd accuracy.

Human forecasting accuracy improves through the "wisdom of the crowd" effect, in which aggregated predictions tend to outperform individual ones. Past research suggests that individual large language models (LLMs) tend to underperform compared to human crowd aggregates. We simulate a wisdom of the crowd effect with LLMs. Specifically, we use an ensemble of 12 LLMs to make probabilistic predictions about 31 binary questions, comparing them with those made by 925 human forecasters in a 3-month tournament. We show that the LLM crowd outperforms a no-information benchmark and is statistically indistinguishable from the human crowd. We also observe human-like biases, such as the acquiescence bias. In another study, we find that LLM predictions (of GPT-4 and Claude 2) improve when exposed to the median human prediction, increasing accuracy by 17 to 28%. However, simply averaging human and machine forecasts yields more accurate results. Our findings suggest that LLM predictions can rival the human crowd's forecasting accuracy through simple aggregation.

Enhancement of the Performance of Large Language Models in Diabetes Education through Retrieval-Augmented Generation: Comparative Study.

Large language models (LLMs) demonstrated advanced performance in processing clinical information. However, commercially available LLMs lack specialized medical knowledge and remain susceptible to generating inaccurate information. Given the need for self-management in diabetes, patients commonly seek information online. We introduce the Retrieval-augmented Information System for Enhancement (RISE) framework and evaluate its performance in enhancing LLMs to provide accurate responses to diabetes-related inquiries. This study aimed to evaluate the potential of the RISE framework, an information retrieval and augmentation tool, to improve the LLM's performance to accurately and safely respond to diabetes-related inquiries. The RISE, an innovative retrieval augmentation framework, comprises 4 steps: rewriting query, information retrieval, summarization, and execution. Using a set of 43 common diabetes-related questions, we evaluated 3 base LLMs (GPT-4, Anthropic Claude 2, Google Bard) and their RISE-enhanced versions respectively. Assessments were conducted by clinicians for accuracy and comprehensiveness and by patients for understandability. The integration of RISE significantly improved the accuracy and comprehensiveness of responses from all 3 base LLMs. On average, the percentage of accurate responses increased by 12% (15/129) with RISE. Specifically, the rates of accurate responses increased by 7% (3/43) for GPT-4, 19% (8/43) for Claude 2, and 9% (4/43) for Google Bard. The framework also enhanced response comprehensiveness, with mean scores improving by 0.44 (SD 0.10). Understandability was also enhanced by 0.19 (SD 0.13) on average. Data collection was conducted from September 30, 2023 to February 5, 2024. The RISE significantly improves LLMs' performance in responding to diabetes-related inquiries, enhancing accuracy, comprehensiveness, and understandability. These improvements have crucial implications for RISE's future role in patient education and chronic illness self-management, which contributes to relieving medical resource pressures and raising public awareness of medical knowledge.

Assessment of Large Language Models in Cataract Care Information Provision: A Quantitative Comparison.

Cataracts are a significant cause of blindness. While individuals frequently turn to the Internet for medical advice, distinguishing reliable information can be challenging. Large language models (LLMs) have attracted attention for generating accurate, human-like responses that may be used for medical consultation. However, a comprehensive assessment of LLMs' accuracy within specific medical domains is still lacking. We compiled 46 commonly inquired questions related to cataract care, categorized into six domains. Each question was presented to the LLMs, and three consultant-level ophthalmologists independently assessed the accuracy of their responses on a three-point scale (poor, borderline, good) and their comprehensiveness on a five-point scale. A majority consensus approach established the final rating for each response. Responses rated as 'Poor' were prompted for self-correction and reassessed. For accuracy, ChatGPT-4o and Google Bard both achieved average sum scores of 8.7 (out of 9), followed by ChatGPT-3.5, Bing Chat, Llama 2, and Wenxin Yiyan. In consensus-based ratings, ChatGPT-4o outperformed Google Bard in the 'Good' rating. For completeness, ChatGPT-4o had the highest average sum score of 13.22 (out of 15), followed by Google Bard, ChatGPT-3.5, Llama 2, Bing Chat, and Wenxin Yiyan. Detailed performance data reveal nuanced differences in model capabilities. In the 'Prevention' domain, apart from Wenxin Yiyan, all other models were rated as 'Good'. All models showed improvement in self-correction. Bard and Bing improved 1/1 from 'Poor' to better, Llama improved 3/4, and Wenxin Yiyan improved 4/5. Our findings emphasize the potential of LLMs, particularly ChatGPT-4o, to deliver accurate and comprehensive responses to cataract-related queries, especially in prevention, indicating potential for medical consultation. Continuous efforts to enhance LLMs' accuracy through ongoing strategies and evaluations are essential.

Research on adversarial attack and defense of large language models

Abstract. Large language models (LLMs) have made excellent progress in text and image understanding and generation. However, with the wide range of applications of these models in various industries, the issue of their security, especially the defense against adversarial attacks, has become a focus of research. This study focuses on exploring the adversarial attacks faced by LLMs and their defense strategies, especially the design and optimization of defense mechanisms. Through literature review and case studies, this paper analyzes in detail the white-box and black-box attack patterns against LLMs, including model inversion, backdoor attacks, and token-based strategies. In response to these attacks, this paper proposes a series of defense strategies, including preventive measures such as data augmentation, adversarial training and model regularization, as well as real-time attack detection and response strategies such as anomaly detection and adversarial sample detection techniques. The core of this research is to improve the robustness and trustworthiness of LLMs, providing the necessary guarantees for their integration and sustainability in multiple industrial applications. In addition, this paper proposes future research directions, highlighting the importance of developing advanced defense systems, promoting interdisciplinary research and exploring new applications for LLMs. This research provides valuable insights into understanding and improving the security defense mechanisms of LLMs, which is essential for maintaining the security and user trust of these models.

Analysis of Accelerating Ideation Process with Large Language Model

Abstract. Large Language Models (LLMs) have rapidly advanced, demonstrating near-human intelligence in language comprehension, problem-solving, and autonomous decision-making, which can be leveraged in scientific discovery. Previous research of LLM for science often focuses on experiment execution part and the first stage of research, i.e., idea generation with large language models is lack of research. This paper fills this gap by surveying the use of LLM-based agent systems in research idea generation. This research first proposes three stages in research idea generation pipeline, pre-ideation (knowledge preparation), ideation (generation and iteration), post-ideation (evaluation), and provide detailed summary in different methods in each stage, then it summarizes different metrics and angles of evaluating generated research ideas, finally it discusses limitations and ethical concerns of existing works and suggest potential solutions and future directions. These results aim to provide a solid foundation for future research in improving LLM-based ideation systems and fostering responsible AI usage in scientific discovery.

Influence Reasoning Capabilities of Large Language Models in Social Environments

We ask whether state-of-the-art large language models can provide a viable alternative to human annotators for detecting and explaining behavioural influence online. Working with a large corpus of online interactions retrieved from the social media platform Mastodon, we cross-examine a dataset containing 11,000 LLM influence labels and explanations across nine state-of-the-art large language models from 312 scenarios. We use a range of resolution categories and four stages of shot prompting to further measure the importance of context to language model performance. We also consider the impact of model architecture, and how social media content and features from the explanation impact model labelling accuracy. Our experiment shows that whilst most large language models struggle to identify the correct framing of influence from an interaction, at lower label resolutions, models like Flan and GPT-4 Turbo perform with an accuracy of 70%-80%, demonstrating encouraging potential for future social influence identification and explanation, and contributing to our understanding of the general social reasoning capabilities of large language models.

Cause and Effect: Can Large Language Models Truly Understand Causality?

With the rise of Large Language Models (LLMs), it has become crucial to understand their capabilities and limitations in deciphering and explaining the complex web of causal relationships that language entails. Current methods use either explicit or implicit causal reasoning, yet there is a strong need for a unified approach combining both to tackle a wide array of causal relationships more effectively. This research proposes a novel architecture called Context-Aware Reasoning Enhancement with Counterfactual Analysis (CARE-CA) to enhance causal reasoning and explainability. The proposed framework incorporates an explicit causal detection module with ConceptNet and counterfactual statements, as well as implicit causal detection through LLMs. Our framework goes one step further with a layer of counterfactual explanations to accentuate LLMs' understanding of causality. The knowledge from ConceptNet enhances the performance of multiple causal reasoning tasks such as causal discovery, causal identification, and counterfactual reasoning. The counterfactual sentences add explicit knowledge of `not caused by' scenarios. By combining these powerful modules, our model aims to provide a deeper understanding of causal relationships, enabling enhanced interpretability. Evaluation of benchmark datasets shows improved performance across all metrics, such as accuracy, precision, recall, and F1 scores. We also present CausalNet, a novel dataset specifically curated to benchmark and enhance the causal reasoning capabilities of LLMs. This dataset is accompanied by code designed to facilitate further research in this domain.

Understanding the Daily Lives of Older Adults: Integrating Multi-modal Personal Health Tracking Data through Visualization and Large Language Models

Understanding the daily lives and routines of older adults is crucial to facilitate aging in place. Ubiquitous computing technologies like smartphones and wearables that are easy to deploy and scale, have become a popular method to collect comprehensive and longitudinal data for various demographics. Despite their popularity, several challenges persist when targeting the older adult population such as low compliance and hard to obtain feedback. In this work-in-progress paper, we present the design and development of a multi-modal sensing system that includes a phone, watch, and voice assistant. We are conducting an initial longitudinal study with one older adult participant over 30 days to explore how various types of data can be integrated through visualization techniques and large language models (LLMs). As a work-in-progress, we discussed our preliminary insights from the collected data, and conclude with a discussion of our future plans and directions for this research.

Knowledge Aware Automated Health Claims Processing with Medical Ontologies and Large Language Models

One of the major challenges in automating health insurance claims processing lies in the complexity involved in validating an incoming claim's medical diagnoses against its policy Underwriting (UW) exclusions. Termed UW Exclusion Detection, this process ensures claims are only paid out if their diagnoses are not medically associated with conditions excluded under the policy. Medical diagnoses in health insurance claims are typically represented by the International Classification of Disease (ICD) codes, established by the World Health Organization. For example, given a policy that excludes "all respiratory illness". A claim with the ICD code J45 (Asthma) will be subject to rejection as J45 is a respiratory-related diagnosis that falls within the scope of the policy's UW exclusion. The key challenge in automating this process lies in the wide range of available ICD codes. The ICD-10-CM coding scheme consists of over 40,000 codes, which often results in scenarios where codes encountered during inference are absent from the training data. These unseen ICD codes limit the effectiveness of data-driven approaches, which depend on the training data to discern medically relevant associations between UW exclusions and ICD codes. This underscores the need to supplement data-driven approaches with additional domain knowledge. We hypothesize that integrating implicit medical domain knowledge inherent in Large Language Models (LLMs) with explicit domain knowledge from medical ontologies, will enhance data-driven approaches for UW Exclusion Detection. Thoroughly validated on real-world health insurance claims data, our proposed approach proved effective in accurately establishing medically relevant associations between UW exclusions and ICD codes.

Investigating Open Source LLMs to Retrofit Competency Questions in Ontology Engineering

Competency Questions (CQs) are essential in ontology engineering; they express an ontology's functional requirements as natural language questions, offer crucial insights into an ontology's scope and are pivotal for various tasks, e.g. ontology reuse, testing, requirement specification, and pattern definition. Despite their importance, the practice of publishing CQs alongside ontological artefacts is not commonly adopted. We propose an approach based on Generative AI, specifically Large Language Models (LLMs) for retrofitting CQs from existing ontologies and we investigate how open LLMs (i.e. Llama-2-70b, Mistral 7B and Flan-T5-xl) perform in generating CQs for existing ontologies. We compare these results with our previous efforts using closed-source LLMs and we reflect on the results.

Generating Ontology-Learning Training-Data through Verbalization

Ontologies play an important role in the organization and representation of knowledge. However, in most cases, ontologies do not fully cover domain knowledge, resulting in a gap. This gap, often expressed as a lack of concepts, relations, or axioms, is usually filled by domain experts in a manual and tedious process. Utilizing large language models (LLMs) can ease this process; a fine-tuned LLM could receive as input up-to-date and reliable domain knowledge natural text and output a structured graph in OWL RDF/Turtle format, which is the standard format of ontologies. Thus, to fine-tune a model, text-owl sentence pairs that constitute such a dataset must be acquired. Unfortunately, such a dataset does not exist in the literature or within the open-source community. Therefore, this paper introduces our LLM-assisted verbalizer to create such a data set by converting OWL statements from existing ontologies into natural text. We evaluate the verbalizer on 322 classes from four different ontologies using two different LLMs, achieving precision and recall as high as 0.99 and 0.96, respectively.

Enhancing Fairness in LLM Evaluations: Unveiling and Mitigating Biases in Standard-Answer-Based Evaluations

Large Language Models (LLMs) are recognized for their effectiveness in comparing two answers. However, LLMs can still exhibit biases when comparing one answer to a standard answer, particularly in real-world scenarios like new employee orientations. This paper identifies positional and verbosity biases in LLM evaluators in such contexts. To mitigate these biases, we apply Chain of Thought prompting and Multi-Agent Debate strategies. Our research reveals that bias prevalence varies among different models, indicating the need for tailored approaches to ensure unbiased and constructive feedback.

A review of research on training responsible large language models

Abstract. In recent years, there has been a growing acceptance of large language models (LLM) as a mainstream method in the field of natural language processing. Consequently, numerous studies have been conducted on this topic. Training responsible Large Language Models have become a prominent subject of research in the past few years. This type of research mainly focuses on the examination of bias, morality and other aspects of LLM. There are certain similarities in the methodologies employed in those studies. This article presents a comprehensive overview of numerous recent investigations, analyzing and categorizing the methodologies employed in these studies, and offering a literature review. This review examines the three perspectives of LLM bias data set construction, bias detection and bias elimination It provides a comparative analysis of the advantages and disadvantages of different methods. After completing relevant evaluations, a comprehensive examination of the research on training responsible LLM is conducted and potential future research directions are proposed in this article.

Quality of Chatbot Information Related to Benign Prostatic Hyperplasia.

Large language model (LLM) chatbots, a form of artificial intelligence (AI) that excels at prompt-based interactions and mimics human conversation, have emerged as a tool for providing patients with information about urologic conditions. We aimed to examine the quality of information related to benign prostatic hyperplasia surgery from four chatbots and how they would respond to sample patient messages. We identified the top three queries in Google Trends related to "treatment for enlarged prostate." These were entered into ChatGPT (OpenAI), Bard (Google), Bing AI (Microsoft), and Doximity GPT (Doximity), both unprompted and prompted for specific criteria (optimized). The chatbot-provided answers to each query were evaluated for overall quality by three urologists using the DISCERN instrument. Readability was measured with the built-in Flesch-Kincaid reading level tool in Microsoft Word. To assess the ability of chatbots to answer patient questions, we prompted the chatbots with a clinical scenario related to holmium laser enucleation of the prostate, followed by 10 questions that the National Institutes of Health recommends patients ask before surgery. Accuracy and completeness of responses were graded with Likert scales. Without prompting, the quality of information was moderate across all chatbots but improved significantly with prompting (mean [SD], 3.3 [1.2] vs. 4.4 [0.7] out of 5; p < 0.001). When answering simulated patient messages, the chatbots were accurate (mean [SD], 5.6 [0.4] out of 6) and complete (mean [SD], 2.8 [0.3] out of 3). Additionally, 98% (39/40) had a median score of 5 or higher for accuracy, which corresponds to "nearly all correct." The readability was poor, with a mean (SD) Flesch-Kincaid reading level grade of 12.1 (1.3) (unprompted). LLM chatbots hold promise for patient education, but their effectiveness is limited by the need for careful prompting from the user and by responding at a reading level higher than that of most Americans (grade 8). Educating patients and physicians on optimal LLM interaction is crucial to unlock the full potential of chatbots.