SammanfattningI  denna  studie  utforskas  kapaciteterna  och  begränsningarna  hos  avancerade  stora språkmodeller (SSM), med särskilt fokus på ChatGPT-4 och Google Bard. Studien inleds med att ge en historisk bakgrund till artificiell intelligens och hur denna utveckling har lett fram till skapandet av dessa modeller. Därefter genomförs en kritisk analys av deras prestanda i språkbehandling och problemlösning. Genom att evaluera deras effektivitet i hanteringen av nyhetsinnehåll och sociala medier, samt i utförandet av kreativa uppgifter som pussel, belyses deras förmåga inom språklig bearbetning samt de utmaningar de möter i att förstå nyanser och utöva kreativt tänkande.I denna studie framkom det att SSM har en avancerad förmåga att förstå och reagera på komplexa språkstrukturer. Denna förmåga är dock inte utan begränsningar, speciellt när det kommer till uppgifter som kräver en noggrann bedömning för att skilja mellan sanning och osanning. Denna observation lyfter fram en kritisk aspekt av SSM:ernas nuvarande kapacitet, de är effektiva inom många områden, men möter fortfarande utmaningar i att hantera de finare nyanserna i mänskligt språk och tänkande. Studiens resultat betonar även vikten av mänsklig tillsyn vid användning av artificiell intelligens (AI), vilket pekar på behovet av att ha realistiska förväntningar på AI:s kapacitet och betonar vidare betydelsen av en ansvarsfull utveckling  av  AI,  där  en  noggrann  uppmärksamhet  kring etiska  aspekter  är  central.  En kombination av mänsklig intelligens och AI föreslås som en lösning för att hantera komplexa utmaningar, vilket bidrar till en fördjupad förståelse av avancerade språkmodellers dynamik och deras roll inom AI:s bredare utveckling och tillämpning.

<p dir="ltr">State-level policy decisions significantly influence various aspects of our daily lives, such as access to healthcare and education. Despite their importance, there is a limited understanding of how these policies and decisions are formulated within the legislative process. This dissertation aims to bridge that gap by utilizing data-driven methods and the latest advancements in machine learning (ML) and natural language processing (NLP). By leveraging data-driven approaches, we can achieve a more objective and comprehensive understanding of policy formation. The incorporation of ML and NLP techniques aids in processing and interpreting large volumes of complex legislative texts, uncovering patterns and insights that might be overlooked through manual analysis. In this dissertation, we pose new analytical questions about the state legislative process and address them in three stages:</p><p><br></p><p dir="ltr">First, we aim to understand the language of political agreement and disagreement in legislative texts. We introduce a novel NLP/ML task: predicting significant conflicts among legislators and sharp divisions in their votes on state bills, influenced by factors such as gender, rural-urban divides, and ideological differences. To achieve this, we construct a comprehensive dataset from multiple sources, linking state bills with legislators’ information, geographical data about their districts, and details about donations and donors. We then develop a shared relational and textual deep learning model that captures the interactions between the bill’s text and the legislative context in which it is presented. Our experiments demonstrate that incorporating this context enhances prediction accuracy compared to strong text-based models.</p><p><br></p><p dir="ltr">Second, we analyze the impact of legislation on relevant stakeholders, such as teachers in education bills. We introduce this as a new prediction task within our framework to better understand the state legislative process. To address this task, we enhance our modeling and expand our dataset using various techniques, including crowd-sourcing, to generate labeled data. This approach also helps us decode legislators’ decision-making processes and voting patterns. Consequently, we refine our model to predict the winners and losers of bills, using this information to more accurately forecast the legislative body’s vote breakdown based on demographic and ideological criteria.</p><p><br></p><p dir="ltr">Third, we enhance our analysis and modeling of state-level bills and policies using two techniques: We normalize the inconsistent, verbose, and complex language of state policies by leveraging Generative Large Language Models (LLMs). Additionally, we evaluate the policies within a broader network context by expanding the number of US states analyzed from 3 to 50 and incorporating new data sources, such as interest groups’ ratings of legislators and public information on legislators’ positions on various issues.</p><p><br></p><p dir="ltr">By following these steps in this dissertation, we aim to better understand the legislative processes that shape state-level policies and their far-reaching effects on society.</p>

Pre-trained and large language models (PLMs, LLMs) have had a transformative impact on the artificial intelligence (AI) for software engineering (SE) research field. Through large-scale pre-training on terabytes of natural and programming language data, these models excel in generative coding tasks such as program repair and code generation. Existing approaches to align the model's behaviour with specific tasks propose using parameter-free methods like prompting or fine-tuning to improve their effectiveness. Nevertheless, it remains unclear how to align code PLMs and LLMs to more complex scenarios that extend beyond task effectiveness. We focus on model alignment in three overlooked scenarios for code generation, each addressing a specific objective: optimizing fine-tuning costs, aligning models with new data while retaining previous knowledge, and aligning with user coding preferences or non-functional requirements. We explore these scenarios in three articles, which constitute the main contributions of this thesis. In the first article, we conduct an empirical study on parameter-efficient fine-tuning techniques (PEFTs) for code LLMs in resource-constraint settings. Our study reveals the superiority of PEFTs over few-shot learning, showing that PEFTs like LoRA and QLoRA allow fine-tuning LLMs with up to 33 billion parameters on a single 24GB GPU without compromising task effectiveness. In the second article, we examine the behaviour of code PLMs in a continual fine-tuning setting, where the model acquires new knowledge from sequential domain-specific datasets. Each dataset introduces new data about third-party libraries not seen during pre-training or previous fine-tuning. We demonstrate that sequential fine-tuning leads to catastrophic forgetting and implement replay- and regularization-based continual learning approaches, showcasing their superiority in balancing task effectiveness and knowledge retention. In our third article, we introduce CodeUltraFeedback and CODAL-Bench, a novel dataset and benchmark for aligning code LLMs to user coding preferences or non-functional requirements. Our experiments reveal that tuning LLMs with reinforcement learning techniques like direct preference optimization (DPO) using CodeUltraFeedback results in better-aligned LLMs to coding preferences and substantial improvement in the functional correctness of LLM-generated code. / Les modèles de langue pré-entraînés et de grande taille (PLMs, LLMs) ont eu un impact transformateur sur le domaine de la recherche en intelligence artificielle (IA) pour l’ingénierie logicielle (SE). Grâce à un pré-entraînement à grande échelle sur des téraoctets de données en langage naturel et de programmation, ces modèles excellent dans les tâches de codage génératif telles que la réparation de programmes et la génération de code. Les approches existantes pour aligner le comportement du modèle avec des tâches spécifiques proposent l’utilisation de méthodes non paramétriques telles que le prompting ou le fine-tuning pour améliorer leur efficacité. Néanmoins, il reste incertain comment aligner les PLMs et LLMs de code sur des scénarios plus complexes qui nécessitent plus que garantir l’efficacité du modèle sur des tâches cibles. Nous nous concentrons sur l’alignement des modèles dans trois scénarios négligés pour la génération de code, chacun abordant un objectif spécifique: optimiser les coûts de fine-tuning, aligner les modèles avec de nouvelles données dans le temps tout en conservant les connaissances antérieures, et aligner les modèles sur les préférences de codage des utilisateurs ou exigences non fonctionnelles. Nous explorons ces scénarios dans trois articles, qui constituent les principales contributions de cette thèse. Dans le premier article, nous réalisons une étude empirique sur les techniques de finetuning efficaces en paramètres (PEFTs) pour les LLMs de code dans des environnements à ressources limitées. Notre étude révèle la supériorité des PEFTs par rapport au few-shot learning, montrant que des PEFTs comme LoRA et QLoRA permettent de fine-tuner des LLMs jusqu’à 33 milliards de paramètres sur un seul GPU de 24Go sans compromettre l’efficacité sur les tâches. Dans le deuxième article, nous examinons le comportement des PLMs de code dans un contexte de fine-tuning continu, où le modèle acquiert de nouvelles connaissances à partir de jeux de données séquentiels. Chaque jeu de données introduit de nouvelles informations sur des bibliothèques tierces non vues lors de la phase de préentraînement ou dans les jeux de données de fine-tuning précédents. Nous démontrons que le fine-tuning séquentiel conduit à de l’oubli catastrophique et mettons en œuvre des approches d’apprentissage continu basées sur le replay et la régularisation, et montrons leur supériorité pour balancer l’efficacité du modèle et la rétention des connaissances. Dans notre troisième article, nous introduisons CodeUltraFeedback et CODAL-Bench, un nouveau jeu de données et un banc d’essai pour aligner les LLMs de code sur les préférences de codage des utilisateurs ou exigences non fonctionnelles. Nos expériences révèlent que le tuning des LLMs avec des techniques d’apprentissage par renforcement comme l’optimisation directe des préférences (DPO) utilisant CodeUltraFeedback résulte en des LLMs mieux alignés sur les préférences de codage et une amélioration substantielle de l’exactitude fonctionnelle des codes générés.

This master thesis explores the transformative potential of Large Language Models (LLMs) in enhancing business processes across various industries, with a specific focus on Management Innovation. As organizations face the pressures of digitalization, LLMs emerge as powerful tools that can revolutionize traditional business workflows through enhanced decision-making, automation of routine tasks, and improved operational efficiency. The research investigates the integration of LLMs within four key business domains: Human Resources, Tender Management, Consultancy, and Compliance. It highlights how LLMs facilitate Management Innovation by enabling new forms of workflow automation, data analysis, and compliance management, thus driving substantial improvements in efficiency and innovation. Employing a mixed-method approach, the study combines an extensive literature review with surveys and interviews with industry professionals to evaluate the impact and practical applications of LLMs. The findings reveal that LLMs not only offer significant operational benefits but also pose challenges related to data security, integration complexities, and privacy concerns. This thesis significantly contributes to the academic and practical understanding of LLMs, proposing a framework for their strategic adoption to foster Management Innovation. It underscores the need for businesses to align LLM integration with both technological capabilities and strategic business objectives, paving the way for a new era of management practices shaped by advanced technologies. / Denna masteruppsats utforskar den transformativa potentialen hos Stora Språkmodeller (LLMs) i att förbättra affärsprocesser över olika industrier, med särskilt fokus på Management Innovation. När organisationer möter digitaliseringens press, framträder LLMs som kraftfulla verktyg som kan revolutionera traditionella affärsarbetsflöden genom förbättrat beslutsfattande, automatisering av rutinuppgifter och förbättrad operationell effektivitet. Forskningen undersöker integrationen av LLMs inom fyra centrala affärsområden: Human Resources, Anbudshantering, Konsultverksamhet och Regelefterlevnad. Den belyser hur LLMs underlättar Management Innovation genom att möjliggöra nya former av arbetsflödesautomatisering, dataanalys och efterlevnadshantering, vilket driver påtagliga förbättringar i effektivitet och innovation. Genom att använda en blandad metodansats kombinerar studien en omfattande litteraturöversikt med enkäter och intervjuer med branschproffs för att utvärdera påverkan och praktiska tillämpningar av LLMs. Resultaten visar att LLMs inte bara erbjuder betydande operationella fördelar utan även medför utmaningar relaterade till datasäkerhet, integrationskomplexitet och integritetsfrågor. Denna uppsats bidrar avsevärt till den akademiska och praktiska förståelsen av LLMs, och föreslår en ram för deras strategiska antagande för att främja Management Innovation. Den understryker behovet för företag att anpassa LLM-integrationen med både teknologiska kapabiliteter och strategiska affärsmål, vilket banar väg för en ny era av ledningspraxis formad av avancerade teknologier.

The integration of Large Language Models (LLMs) such as GPT-4 and Gemini into healthcare, particularly for elderly care, represents a significant opportunity in the use of artificial intelligence in medical settings. This thesis investigates the capabilities of these models to understand and respond to the healthcare needs of older adults effectively. A framework was developed to evaluate their performance, consisting of specifically designed medical scenarios that simulate real-life interactions, prompting strategies to elicit responses and a comprehensive user evaluation to assess technical performance and contextual understanding.  The analysis reveals that while LLMs such as GPT-4 and Gemini exhibit high levels of technical proficiency, their contextual performance shows considerable variability, especially in personalization and handling complex, empathy-driven interactions. In simpler tasks, these models demonstrate appropriate responsiveness, but they struggle with more complex scenarios that require deep medical reasoning and personalized communication.  Despite these challenges, the research highlights the potential of LLMs to significantly enhance healthcare delivery for older adults by providing timely and relevant medical information. However, to realize a truly effective implementation, further development is necessary to improve the models’ ability to engage in meaningful dialogue and understand the nuanced needs of an aging population.  The findings underscore the necessity of actively involving older adults in the development of AI technologies, ensuring that these models are tailored to their specific needs. This includes focusing on enhancing the contextual and demographic awareness of AI systems. Future efforts should focus on enhancing these models by incorporating user feedback from the older population and applying user-centered design principles to improve accessibility and usability. Such improvements will better support the diverse needs of aging populations in healthcare settings, enhancing care delivery for both patients and doctors while maintaining the essential human touch in medical interactions. / Integrationen av stora språkmodeller (LLMs) såsom GPT-4 och Gemini inom sjukvården, särskilt inom äldrevård, representerar betydande möjligheter i användningen av artificiell intelligens i medicinska sammanhang. Denna avhandling undersöker dessa modellers förmåga att förstå och effektivt svara på äldres vårdbehov. För att utvärdera deras prestanda utvecklades ett ramverk bestående av specifikt utformade medicinska situationer som simulerar verkliga interaktioner, strategier för att framkalla relevanta svar från modellerna och en omfattande användarutvärdering för att bedöma både teknisk prestanda och kontextuell förståelse.  Analysen visar att även om LLMs såsom GPT-4 och Gemini visar på hög teknisk prestationsförmåga, är dess kontextuella förmåga mer begränsad, särskilt när det gäller personalisering och hantering av komplexa, empatidrivna interaktioner. Vid enklare uppgifter visar dessa modeller på en lämplig responsivitet, men de utmanas vid mer komplexa scenarier som kräver djup medicinsk resonemang och personlig kommunikation.  Trots dessa utmaningar belyser denna forskning potentialen hos LLMs att väsentligt förbättra vårdleveransen för äldre genom att tillhandahålla aktuell och relevant medicinsk information. Däremot krävs ytterligare utveckling för att verkligen möjliggöra en effektiv implementering, vilket inkluderar att förbättra modellernas förmåga att delta i en meningsfull dialog och förstå de nyanserade behoven hos äldre patienter.  Resultaten från denna avhandling understryker nödvändigheten av att aktivt involvera äldre individer i utvecklingen av AI-teknologier, för att säkerställa att dessa modeller är skräddarsydda för deras specifika behov. Detta inkluderar ett fokus på att förbättra den kontextuella och demografiska medvetenheten hos AI-system. Framtida insatser bör inriktas på att förbättra dessa modeller genom att integrera användarfeedback från äldre populationer och tillämpa principer för användarcentrerad design för att förbättra tillgänglighet och användbarhet. Sådana förbättringar kommer att bättre stödja de mångsidiga behoven hos äldre i vårdsammanhang, förbättra vårdleveransen för både patienter och läkare samtidigt som den väsentliga mänskliga kontakten i medicinska interaktioner bibehålls.

This essay critically examines the limitations of artificial intelligence (AI) in achieving human-like understanding and intelligence. Despite significant advancements in AI, such as the development of sophisticated machine learning algorithms and neural networks, current systems fall short in comprehending the cognitive depth and flexibility inherent in human intelligence. Through an exploration of historical and contemporary arguments, including Searle's Chinese Room thought experiment and Dennett's Frame Problem, this essay highlights the inherent differences between human cognition and AI. Central to this analysis is the role of metaphorical thinking and embodied cognition, as articulated by Lakoff and Johnson, which are fundamental to human understanding but absent in AI. Proponents of AGI, like Kurzweil and Bostrom, argue for the potential of AI to surpass human intelligence through recursive self-improvement and technological integration. However, this essay contends that these approaches do not address the core issues of experiential knowledge and contextual awareness. By integrating insights from contemporary scholars like Bender, Koller, Buckner, Thorstad, and Hoffmann, the essay ultimately concludes that AI, while a powerful computational framework, is fundamentally incapaple of replicating the true intelligence and understanding unique to humans.

In the rapidly advancing field of artificial intelligence, Large Language Models (LLMs) like GPT-3, GPT-4, and Gemini have revolutionized sectors by automating complex tasks. Despite their advancements, LLMs and more noticeably smaller language models (SLMs) still face challenges, such as generating unfounded content "hallucinations." This project aims to enhance SLMs for broader accessibility without extensive computational infrastructure. By supervised fine-tuning of smaller models with new datasets, SQUAD-ei and SQUAD-GPT, the resulting model, KERMIT-7B, achieved superior performance in TYDIQA-GoldP, demonstrating improved information extraction while retaining generative quality. / Inom det snabbt växande området artificiell intelligens har stora språkmodeller (LLM) som GPT-3, GPT-4 och Gemini revolutionerat sektorer genom att automatisera komplexa uppgifter. Trots sina framsteg stårdessa modeller, framför allt mindre språkmodeller (SLMs) fortfarande inför utmaningar, till exempel attgenerera ogrundat innehåll "hallucinationer". Denna studie syftar till att förbättra SLMs för bredare till-gänglighet utan krävande infrastruktur. Genom supervised fine-tuning av mindre modeller med nya data-set, SQUAD-ei och SQUAD-GPT, uppnådde den resulterande modellen, KERMIT-7B, överlägsen pre-standa i TYDIQA-GoldP, vilket visar förbättrad informationsutvinning samtidigt som den generativa kva-liteten bibehålls.

The goal of the thesis was to create several variants of agent systems based on LLM, analyze and compare their results in problems requiring reasoning. Due to time constraints, the research had to be limited to one data set, namely GSM8K. The solution to each problem in it is a single number, which is easy to evaluate. One of the main tasks was to create an agent system based on large language models with a small number of parameters by their standards. To select a candidate for an agent, large language models with open source code starting from 7 billion parameters were considered. As a result of the observations made, it was decided to use llama3 8b. In order for the agent to better understand its place in the multi-agent system, as well as to improve interaction between agents, the agents were assigned the following roles: a student agent, a teacher agent, an agent changing the wording of the text (while maintaining the meaning), an agent checking the final answer, and an agent changing the level of abstraction of the task. As a result of the final qualifying work, twelve different agent systems were created in two versions: with and without the use of thought chains. Most of the systems did not prove to be better than one large language model, but two of them were still able to distinguish themselves: a system of two student agents with different approaches to solving and a teacher agent, as well as a system consisting of two agents changing the level of abstraction of the task and a student agent. / Целью ВКР являлось создание нескольких вариантов систем агентов на основе LLM, анализ и сравнение их результатов в задачах, требующих рассуждений. За счет временных ограничений, исследования пришлось ограничить в рамках одного набора данных, а именно GSM8K. Решение каждой задачи в нем являются единственным числом, что легко оценивать. Одной из основных задач было создание системы агентов на основе больших языковых моделей с небольшим по их меркам количеством параметров. Для выбора кандидата в агенты рассматривались большие языковые модели с кодом в открытом доступе размером начиная от 7 миллиардов параметров. В результате проделанных наблюдений было решено использовать llama3 8b. Чтобы агент лучше понимал своё место в мультиагентной системе, а также для улучшения взаимодействия между агентами, агентам назначались следующие роли: агент-ученик, агент-учитель, агент, меняющий формулировку текста (сохраняя при этом смысл), агент, проверяющий итоговый ответ, а также агент, меняющий уровень абстракции задачи. В результате ВКР было создано двенадцать различных систем агентов в двух версиях: с применением цепей мыслей и без. Большая часть систем не показала себя лучше одной большой языковой модели, однако две из них всё же смогли отличиться: система из двух агентов-учеников с разными подходами к решению и агента учителя, а также система, состоящая из двух агентов, меняющих уровень абстракции задачи, и агента ученика.

<p dir="ltr">We now live in a world where we can reach people directly through social media, without relying on traditional media such as television and radio. The landscape of social media is highly distributed, as users generate and consume a variety of content. On the other hand, social media platforms collect vast amounts of data and create very specific profiles of different users through targeted advertising. Various interest groups, politicians, advertisers, and stakeholders utilize these platforms to target potential users to advance their interests by adapting their messaging.  A significant challenge lies in understanding this messaging and how it changes depending on the targeted user groups. Another challenge arises when we do not know who the users are and what their motivations are for engaging with content. The initial phase of our research focuses on comprehensively understanding users and their underlying motivations, whether practitioner-based or promotional. Gaining this understanding is crucial in reshaping our perspective on the content disseminated by these users. Step beyond that, assuming the identification of the involved parties, this study aims to characterize the messaging and explore how it adapts based on various targeted demographic groups. This thesis addresses these challenges by developing computational approaches and frameworks for (1) characterizing user types and their motivations, (2) analyzing the messaging based on topics relevant to the users and their responses to it, and (3) delving into the deeper understanding of the themes and arguments involved in the messaging.</p>