Abstract

Training and evaluating large language models (LLMs) on deductive reasoning tasks has attracted much attention in recent times. There have been various attempts at training and evaluating LLMs to perform deductive reasoning tasks. Some among these have shown interesting results which suggest that LLMs behave like humans by displaying content effect, that is, they reason better on reasoning tasks containing rules that align with our everyday beliefs and reason poorly when the rules are beliefviolating. Others, using chain-of-thought prompting, a technique to make LLMs generate intermediate steps before making them arrive at the final conclusion, suggest that LLMs may emulate human-like reasoning thought processes by generating intermediate reasoning steps. On the other hand, there are studies which conclude that language models do not yet demonstrate reliable deductive reasoning, since their performance is shown to decrease upon introduction of perturbations, such as synonym substitution, and attribute their reasoning to artefacts from the training data. In this study, we look at these three distinct attempts at evaluating LLMs on deductive competence tasks, each employing different criteria such as content-based reasoning (Dasgupta et al., 2023 [1]), chain-of-thought prompting (Wei et. al., 2023 [2]), and introduction of perturbations (Yuan et al., 2023 [3]). In order to make sense of these claims concerning genuine reasoning, we introduce a framework developed by Diane Proudfoot using externalist criteria for machine cognition (Proudfoot 2004 [4]), which is based on Wittgenstein’s argument that deductive reasoning involves rule-following which is normative in nature. Based on the notion that deductive competence involves the following of normative rules, the framework proposes the use of the Wittgensteinian distinction between rule-following and quasi rule-following as a method to distinguish genuine deductive competence from quasi-competence. The criteria distinguishing between rule-following and quasi rule-following shall be adapted to analyse whether these LLMs can be said to reason genuinely or are merely imitating reasoning-like behaviour. We propose this use of Proudfoot’s criteria for rule-following as a framework to distinguish genuine deductive competence from quasi deductive competence. In doing so, we also draw attention to the limitations and implications of Proudfoot’s claims regarding machine cognition through the introduction of a thought experiment. This thought experiment enables us to think through Proudfoot’s argument, according to which it is due to pragmatic considerations–and not in principle–that LLMs are unlikely to possess genuine reasoning.

Abstract

Many real-world systems can be modeled as dynamic graphs, where nodes and edges evolve with time. Graph Neural Networks (GNNs) excel at modeling relational data. Temporal GNNs capture the dynamics of time-changing data very well. Still, their reliability in risk-sensitive domains where errors in fraud detection, legal judgments, or medical diagnoses carry severe consequences remains limited. Traditional GNNs lack mechanisms to quantify uncertainty or abstain from low-confidence predictions, particularly in dynamic systems where temporal evolution and class imbalance amplify ambiguity. This thesis addresses these gaps by introducing strategic abstention mechanisms into graph learning, helping models to prioritize high-confidence decisions while rejecting uncertain predictions. We unify this approach across dynamic and static graphs, advancing reliability in high-stakes applications through uncertainty-aware frameworks. For the first time, our approach integrates a reject option strategy within the framework of GNNs for continuous-time dynamic graphs. This allows the model to strategically abstain from making predictions when uncertainty is high and confidence is low, minimizing the risk of critical misclassification and enhancing reliability. We propose a coverage-based abstention prediction model to implement the reject option that maximizes predictions within specified coverage. It improves prediction scores for link prediction and node classification tasks. Temporal GNNs deal with skewed datasets for the next state prediction or node classification. In cases of class imbalance, our method can be tuned to provide a higher weight to the minority class. Exhaustive experiments are presented on four datasets for dynamic link prediction and two for dynamic node classification tasks. This demonstrates our approach’s effectiveness in improving reliability and area under the curve (AUC)/average precision (AP) scores for predictions in dynamic graph scenarios. The results highlight our model’s ability to efficiently handle trade-offs between prediction confidence and coverage, making it a dependable solution for applications requiring high precision in dynamic and uncertain environments. Beyond temporal graphs, we extend the concept of classification with the reject option to static graph settings. We reformulate legal judgment prediction (LJP) as node classification on citation networks (ILDC dataset), integrating cost- and coverage-based abstention. Our models (NCwR-Cost/NCwR-Cov) improve accuracy by rejecting uncertain cases, ensuring reliability in legal decision-making. SHAPbased explanations reveal case-specific abstention rationale, enhancing transparency. Further validation on medical datasets (thyroid diagnosis, diabetes prediction) confirms cross-domain validation, with abstention mechanisms reducing misclassification risks in ambiguous cases. This work provides deployable solutions for applications where reliability is non-negotiable.

Abstract

The human brain undergoes continuous structural changes throughout the lifespan, driven by a complex interplay of aging processes, environmental influences, and disease-related mechanisms. Patterns of structural change—particularly atrophy associated with tissue loss and shrinkage—emerge gradually over time and are observable using medical imaging techniques. While these changes are shaped by common biological mechanisms, they are also highly individualized, influenced by factors such as lifestyle, and neurological conditions like Alzheimer’s Disease (AD), Parkinson’s disease, tumors, and stroke. Understanding the progression of these changes—both at the individual level and across populations—is critical for advancing our knowledge of healthy aging and the dynamics of neurodegenerative disease. To study how brain structure evolves over time, researchers rely on longitudinal neuroimaging: repeated imaging of the same individuals at multiple timepoints. Unlike cross-sectional imaging, which captures a single snapshot per subject, longitudinal scans provide a temporal sequence that enables direct observation of anatomical trajectories. These sequences allow for the measurement of rates of change, identification of early biomarkers, and modeling of disease progression in a subject-specific manner. However, acquiring complete longitudinal datasets in practice remains challenging. Subject dropout, missed clinical visits, and protocol variability often result in missing scans, interrupting the temporal continuity required for accurate modeling. These gaps limit the effectiveness of methods that rely on temporally complete inputs and can bias downstream analyses. Imputing the missing scan to complete the subject’s imaging timeline is therefore a critical step toward enabling robust longitudinal modeling and improving our understanding of neurodegenerative processes. This thesis addresses the challenges of modeling and analyzing longitudinal brain changes by developing anatomically grounded methods for data imputation, latent space disentanglement, and downstream trajectory analysis. We first introduce SynBADD, a deformation-based framework that synthesizes missing brain scans by predicting physiologically plausible stationary velocity fields (SVFs)—parametric fields that encode smooth, invertible deformations over space and time—rather than directly generating full image intensities. By operating in the deformation space, SynBADD preserves anatomical coherence and spatial fidelity while mitigating the artifacts typically associated with intensity-based synthesis approaches. Building on this foundation, we propose DIVA, a metadata-informed variational autoencoder designed to learn a temporally disentangled latent space for modeling brain morphological changes. DIVA is trained on synthetically augmented Stationary Velocity Fields (SVFs), enriching intra-subject variation and improving the model’s capacity to generalize across limited real-world samples. Age, disease label, and temporal information are explicitly disentangled through conditional bottleneck supervision, allowing the learned latent representations to reflect meaningful clinical and chronological factors. This disentanglement enhances temporal predictability, supports more accurate subject-specific trajectory modeling, and enables the use of powerful transformer-based architectures for latent space interpolation and prediction. We further extend our analysis in an “Analysis by Synthesis” framework, investigating how metadatainformed conditioning affects generation quality and how different temporal reasoning strategies (past-only, future-only, and bidirectional) impact anatomical plausibility. We perform trajectory-based analyses of generated scans, evaluating how well imputed data aligns with true subject-specific anatomical trends across key regions of interest (ROIs) such as the hippocampus and parahippocampus. Additionally, we assess subgroup differences by age, sex, and disease status, and evaluate downstream task performance with and without synthetic imputation. Through comprehensive evaluations on the Alzheimer’s Disease Neuroimaging Initiative (ADNI) longitudinal dataset, our approaches achieve substantial improvements over state-of-the-art baselines in both image fidelity and clinically relevant anatomical accuracy. Beyond technical advancements, this work provides new insights into the modeling of individualized brain aging patterns and opens pathways for data augmentation in clinical studies where longitudinal completeness is often unattainable.

Abstract

Traffic flow prediction is a critical component of intelligent transportation systems, with direct implications for congestion mitigation, safety enhancement, and efficient travel planning. Given the complex spatio-temporal nature of traffic data, a wide range of hybrid deep learning models—such as CNN-LSTM, ConvLSTM, and Temporal Convolutional Networks (TCNs)— have been explored to capture spatial, temporal, and periodic dependencies effectively. In this work, we perform a comprehensive comparison of deep learning models with and without periodicity to evaluate their prediction accuracies. To address the challenge of hyperparameter tuning in these models, we propose a Genetic Algorithm (GA)-based optimization framework for CNN-LSTM, ConvLSTM, and a novel GA-TCN architecture, all of which demonstrate improved performance on benchmark datasets. Furthermore, we introduce a purely temporal deep learning model named Grid LSTM-based Attention Modelling for Traffic Flow Prediction (GLSTM-A), which leverages a combination of Grid LSTM for long-term dependencies, a standard LSTM for recent trends, and a custom attention mechanism to automatically prioritize significant temporal features. GLSTM-A exhibits superior prediction accuracy and memory efficiency compared to existing temporal models such as TCN, LSTM, and Bi-LSTM. Finally, to tackle the challenge posed by limited long-term traffic datasets, we develop TrafficFlowGAN—a GAN-based time-series data generation framework that effectively captures temporal patterns through a joint supervised and adversarial learning process. The synthetic data generated by TrafficFlowGAN closely resembles real-world traffic patterns, thereby enhancing the robustness and accuracy of downstream prediction models. Extensive experimental evaluations and ablation studies validate the efficiency of the proposed models across various traffic forecasting scenarios.

Abstract

The rapid adoption of Large Language Models (LLMs) as tools for annotation and evaluation tasks in Natural Language Processing (NLP) has led to important questions about their reliability, robustness, and alignment with human expectations. This thesis investigates these concerns across three connected lines of inquiry: • the indistinguishability of LLM-generated annotations from human annotations, • human alignment and reliability of LLMs as subjective judges of language quality, and • the susceptibility of LLM-judges to misinformation attacks in evaluation settings. In the first part of this thesis, we study whether annotations made by LLMs can be reliably distinguished from those created by humans. Previous research has claimed that LLMs behave as ’human-like annotators’, motivating our work to rigorously test this hypothesis. We frame this as a classification task: Given a dataset of annotations, can a model detect their origin? Surprisingly, our findings indicate that even state-of-the-art classifiers achieve near-chance performance, with accuracy not exceeding 51%. These results offer strong empirical evidence that LLMs produce annotations that are indistinguishable from human annotators in most settings, validating the ’human-like’ claim from a discriminative modeling point of view. Building on this, we explore whether LLMs can function as reliable evaluators: a more nuanced role that goes beyond labeling and involves subjective scoring of text quality (e.g., summaries, open-ended responses). In this phase, we test how well LLM-generated judgments align with human evaluations and how robust they are to subtle changes in prompt phrasing. We introduce controlled prompt perturbations and evaluate the consistency of LLM scores and textual justifications. The results show a significant lack of robustness: Small, often misleading perturbations lead to large changes in judgment outputs. These inconsistencies expose limitations in using LLMs as trustworthy evaluators, especially in high-stakes or subjective settings. The final part of this thesis presents a preliminary exploration into the robustness of LLM-based evaluators when exposed to misinformation, marking a conceptual shift from controlled perturbations to real-world-inspired adversarial scenarios. We introduce a novel framework that systematically injects misinformation into prompts, grounded in a taxonomy of ten misinformation types. This study probes how these manipulations affect LLM judgments, analyzing both score alignment and justification consistency. Contrary to our initial expectations, the results suggest that LLM-judges demonstrate a surprising degree of resilience. Scores and justifications remain largely consistent across many misinformation types, even when factual content is altered. However, this robustness appears to be superficial, with the LLM-Judge often failing to explicitly recognize or reflect awareness of being misled in its justifications. Rather than detecting and correcting misinformation, they tend to preserve their prior judgments, raising concerns about unconscious robustness without epistemic awareness. These early findings open new avenues for future research. Comparative assessment, misinformation detection, and deeper semantic analysis may reveal subtler vulnerabilities. More broadly, this work lays the groundwork for building misinformation-aware evaluation systems and motivates the development of LLMs that not only evaluate well but do so with fact-sensitive reasoning. Collectively, this thesis explores the evolving role of LLMs as annotation and evaluation agents. While they show promise as ’human-like’ annotators, their behavior as evaluators under adversarial and misinformation-rich contexts highlights both strengths and blind spots. By offering early insights into their robustness and limitations, we provide a foundation for future work aimed at building trustworthy, interpretable, and context-aware LLM judges.

Abstract

Large Language Models (LLMs) are powerful AI tools with broad abilities, but they still struggle with tasks that require deep understanding and complex, non-standard reasoning. This thesis examines two major challenges: how well LLMs understand and interpret cross-cultural communication, and how to improve their advanced problem-solving skills like lateral thinking and context-aware support. It evaluates how well current LLMs handle these tasks and explores two main strategies to improve their performance: advanced prompting and reinforcement learning-based fine-tuning. The first part of the research focuses on cultural understanding. A user study of book reviews shows that most (83%) contain Culture-Specific Items (CSIs), often making cross-cultural understanding difficult. We test different LLMs—including GPT-4o and smaller open models like Aya and Gemma—on how well they can detect and categorize these CSIs. The results show mixed strengths across models (e.g., Gemma-2 does well with social references, Aya with customs), but all models show a noticeable Western bias compared to human evaluations. To support future work, we also provide standardized datasets. The second part explores LLM performance in complex reasoning tasks such as puzzles that require creative and unconventional thinking. We find that advanced prompting—using both static and dynamic examples, along with model-generated reasoning steps—greatly improves results with Gemini 1.0 Pro over basic methods. However, models still do not reach human-level performance, suggesting that prompting is a promising way to improve complex reasoning, even if it is not yet perfect. Finally, the thesis focuses on generating context-aware assistance for solving challenging math problems, particularly those at the Olympiad level. We propose a framework for producing diverse, multistage synthetic hints tailored to different student needs. These hints cover a range of instructional types, including: beginning steps to help students start; relevant theorems and definitions to build conceptual grounding; equivalent examples for analogical learning; next steps and summaries of remaining steps based on student progress; identification of unused information from the problem; and correctionoriented hints to help students recognize and fix mistakes. Using this framework, we fine-tuned Small Language Models (SLMs) through several reinforcement learning approaches—Supervised Fine-Tuning (SFT), Kahneman-Tversky Optimization (KTO), and Odds Ratio Policy Optimization (ORPO)—to train models to generate helpful, targeted mathematical hints. This not only addresses a key gap in automated pedagogical support for difficult problems but also tackles an important research question: how effectively can these reinforcement learning methods align a smaller model with significantly fewer parameters to synthetic data generated by a much larger LLM? Overall, we work on multi-dimensional evaluation of LLMs in cultural understanding and advanced problem-solving. It highlights current limitations, reveals existing biases, and demonstrates that prompting and fine-tuning together can build more capable, adaptive, and culturally aware AI systems.

Abstract

Re-identification (ReID) is a critical challenge in computer vision, predominantly studied in the context of pedestrians and vehicles. However, robust object-instance ReID, which has significant implications for tasks such as autonomous exploration, long-term perception, and scene understanding, remains underexplored. In this work, we address this gap by proposing a novel dual-path object-instance re-identification transformer architecture that integrates multimodal RGB and depth information. By leveraging depth data, we demonstrate improvements in ReID across scenes that are cluttered or have varying illumination conditions. Additionally, we develop a ReID-based localization framework that enables accurate camera localization and pose identification across different viewpoints. We validate our methods using two custom-built RGB-D datasets, as well as multiple sequences from the open-source TUM RGB-D datasets. Our approach demonstrates significant improvements in both object instance ReID (mAP of 75.18) and localization accuracy (success rate of 83% on TUM-RGBD), highlighting the essential role of object ReID in advancing robotic perception. Our models, frameworks, and datasets have been made publicly available.

Abstract

Flapping-wing robots (FWRs) have garnered significant attention in bio-inspired robotics due to their ability to replicate the highly efficient and manoeuvrable flight of birds and insects. Unlike traditional aerial vehicles, FWRs generate lift and thrust through unsteady aerodynamics, allowing for improved agility and safer operation in confined or complex environments. While early research focused on fundamental aerodynamic modelling and control, recent advancements have explored hybrid propulsion systems, morphing wing designs, and lightweight, high-efficiency actuation mechanisms to enhance endurance and payload. However, most existing research follows a trade-off between range and payload capacity, limiting adaptability for mission-specific tasks. This thesis presents an innovative approach to addressing this limitation by developing a modular flapping-wing robot (FWR) with a swappable powertrain module. The proposed system enables dynamic switching between payloadfocused and endurance-optimized flight profiles by designing interchangeable motor-gearbox configurations and varying flapping kinematics. The research investigates the effects of flapping frequency, stroke amplitude, and tandem auxiliary propulsion on flight performance, leveraging simulation tools like PteraSoftware to refine aerodynamic efficiency. Two powertrain modules are fabricated and tested—one optimized for high-frequency flapping to support payload capacities and another with a tandem propeller to reduce flapping effort and extend range. The flight data collected over multiple experimental flight tests validate the adaptability of this modular design, demonstrating its effectiveness in altering mission capabilities without redesigning the entire platform. Beyond enhancing versatility in FWRs, this research contributes to the broader field by addressing key challenges in adaptive aerial robotics. The findings have direct applications in search-and-rescue missions, environmental monitoring, and silent surveillance, where traditional UAVs face operational constraints. Furthermore, the modularity principle introduced here opens pathways for integrating emerging actuation technologies, such as dielectric elastomer actuators (DEA), for future iterations of highly efficient, autonomous flapping-wing robots. Through this thesis, we aim to push the boundaries of biomimetic flight, offering a scalable and reconfigurable approach to mission-specific aerial robotics.

Abstract

The rapid adoption of electric vehicles (EVs) has accelerated the need for scalable, cost-effective charging infrastructure, particularly in urban environments. This thesis presents a curb-side EVcharging solution that is low-cost, plug-and-play, and interoperable with existing smart-city platforms while requiring no continuous Internet connectivity at each charger. This thesis presents an integrated approach to EV charging, combining Wireless Smart Ubiquitous Network (Wi-SUN) technology with Bluetooth Low Energy (BLE) for enhanced connectivity and cost-effective deployment. The proposed architecture employs Wi-SUN, a long-range, mesh-based sub-GHz communication protocol, as a robust and scalable backbone for charger-to-cloud connectivity, thus reducing operational costs by eliminating reliance on cellular networks. BLE is utilized for efficient short-range interactions, enabling user authentication, session management, and rapid connectivity between the user’s smartphone and the EV charger. The charging system adheres to Level-2 AC charging standards (220-240 V) and follows the Open Charge Point Protocol (OCPP) to ensure interoperability, while also complying with oneM2M middleware standards for seamless integration into broader smart city frameworks. The developed system, encapsulated in an IP65-rated polycarbonate enclosure, integrates a Raspberry Pi as the central processor, Arduino Nano 33 IoT for BLE communication, Wi-SUN transceivers, a solid-state relay, and a energy-metering unit. The field evaluation conducted within the Smart City Living Lab at IIIT Hyderabad validated key performance metrics, demonstrating a reliable communication range of approximately 370 meters between two consecutive Wi-SUN nodes, BLE connection establishment in under 300 milliseconds, and an end-to-end authentication time of less than a second. Furthermore, a proof-of-concept emergency vehicle detection feature underscores the versatility and extensibility of the developed architecture. Overall, this integrated Wi-SUN and BLE-based streetlight-mounted EV charger provides a practical, scalable solution to support the widespread adoption of electric vehicles in urban centers, paving the way for advanced smart-city applications.

Abstract

r exaggerate content to drive user engagement. While they may boost clicks, such headlines and articles compromise reader trust and contribute to the spread of misinformation. Tackling clickbait thus represents not just a technical challenge but also a crucial step toward improving the quality and transparency of online information. To address this, recent research has introduced the task of clickbait spoiling, where we attempt to ”spoil” the headline by generating the key information the article withholds. Our work contributes to this growing area by firstly engaging with the clickbait spoiling task in English using modern Large Language Models (LLMs) and secondly, in Hindi, where we extend this task into a vastly underexplored linguistic space. The spoiling task is especially meaningful because it does not attempt to censor or suppress content instead it just simply makes it more honest, giving users the context they need upfront to make informed choices about what they click. From the evaluations on the English dataset it is revealed that LLMs, particularly when prompted with a few-shot setup, can be highly effective at generating relevant and informative spoilers. This demonstrates the potential of prompting-based systems to perform sophisticated reasoning and summarization tasks, even in cases where training data may be limited. Building on these insights, we make a significant contribution by shifting focus to Hindi, a language that has been critically underrepresented in this field of NLP research despite being spoken by hundreds of millions. We curate the first-of-its-kind Hindi clickbait dataset, using automatically generated annotations from multiple powerful LLMs. This dataset will not only address a huge gap in multilingual clickbait research but also lay a foundation for future work in Hindi misinformation and media literacy. We also observed notable inconsistencies in the behavior of certain large language models. Hermes-3, in particular, exhibited significant divergence in both its clickbait classification judgments and its assigned ratings. It was unable to understand explicit instructions and hence in light of this unreliability, we made a principled decision to take it off the annotator lineup. This screams for quality control whenever using LLMs for such tasks. We then extended the spoiling task to Hindi using this dataset by introducing a selection process based on agreement across models, combined with manual curation, resulting in a dataset that captures the diverse forms of Hindi clickbait. We also classify spoiler types and evaluate LLM performance in zero-shot and few-shot settings. Our findings reveal clear patterns: model performance varies widely, prompting strategies significantly affect outcomes, and certain models (like Deepseek-R1) demonstrate robust generalization even in zero-shot settings.