The growing prevalence of artificial intelligence (AI) in various applications underscores the need for agents that can successfully navigate and adapt to an ever-changing, open-ended world. A key challenge is ensuring these AI agents are robust, excelling not only in familiar settings observed during training but also effectively generalising to previously unseen and varied scenarios. In this thesis, we harness methodologies from open-endedness and multi-agent learning to train and evaluate robust AI agents capable of generalising to novel environments, out-of-distribution inputs, and interactions with other co-player agents. We begin by introducing MiniHack, a sandbox framework for creating diverse environments through procedural content generation. Based on the game of NetHack, MiniHack enables the construction of new tasks for reinforcement learning (RL) agents with a focus on generalisation. We then present Maestro, a novel approach for generating adversarial curricula that progressively enhance the robustness and generality of RL agents in two-player zero-sum games. We further probe robustness in multi-agent domains, utilising quality-diversity methods to systematically identify vulnerabilities in state-of-the-art, pre-trained RL policies within the complex video game football domain, characterised by intertwined cooperative and competitive dynamics. Finally, we extend our exploration of robustness to the domain of LLMs. Here, our focus is on diagnosing and enhancing the robustness of LLMs against adversarial prompts, employing evolutionary search to generate a diverse range of effective inputs that aim to elicit undesirable outputs from an LLM. This work collectively paves the way for future advancements in AI robustness, enabling the development of agents that not only adapt to an ever-evolving world but also thrive in the face of unforeseen challenges and interactions.

Language models are vulnerable to short adversarial suffixes that can reliably alter predictions. Previous works usually find such suffixes with gradient search or rule-based methods, but these are brittle and often tied to a single task or model. In this paper, a reinforcement learning framework is used where the suffix is treated as a policy and trained with Proximal Policy Optimization against a frozen model as a reward oracle. Rewards are shaped using calibrated cross-entropy, removing label bias and aggregating across surface forms to improve transferability. The proposed method is evaluated on five diverse NLP benchmark datasets, covering sentiment, natural language inference, paraphrase, and commonsense reasoning, using three distinct language models: Qwen2-1.5B Instruct, TinyLlama-1.1B Chat, and Phi-1.5. Results show that RL-trained suffixes consistently degrade accuracy and transfer more effectively across tasks and models than previous adversarial triggers of similar genres.

Most post-training backdoor detection methods rely on attacked models exhibiting extreme outlier detection statistics for the target class of an attack, compared to non-target classes. However, these approaches may fail: (1) when some (non-target) classes are easily discriminable from all others, in which case they may naturally achieve extreme detection statistics (e.g., decision confidence); and (2) when the backdoor is subtle, i.e., with its features weak relative to intrinsic class-discriminative features. A key observation is that the backdoor target class has contributions to its detection statistic from both the backdoor trigger and from its intrinsic features, whereas non-target classes only have contributions from their intrinsic features. To achieve more sensitive detectors, we thus propose to suppress intrinsic features while optimizing the detection statistic for a given class. For non-target classes, such suppression will drastically reduce the achievable statistic, whereas for the target class the (significant) contribution from the backdoor trigger remains. In practice, we formulate a constrained optimization problem, leveraging a small set of clean examples from a given class, and optimizing the detection statistic while orthogonalizing with respect to the class's intrinsic features. We dub this plug-and-play approach Class Subspace Orthogonalization (CSO) and assess it against challenging mixed-label and adaptive attacks.

Despite advances in generative large language models (LLMs), practical application of specialized conversational AI agents remains constrained by computation costs, latency requirements, and the need for precise domain-specific relevance measures. While existing embedding models address the first two constraints, they underperform on information retrieval in specialized domains like finance. This paper introduces a scalable pipeline that trains specialized models from an unlabeled corpus using a general purpose retrieval embedding model as foundation. Our method yields an average of 27.7% improvement in MRR$\texttt{@}$5, 44.6% improvement in mean DCG$\texttt{@}$5 across 14 financial filing types measured over 21,800 query-document pairs, and improved NDCG on 3 of 4 document classes in FinanceBench. We adapt retrieval embeddings (bi-encoder) for RAG, not LLM generators, using LLM-judged relevance to distill domain knowledge into a compact retriever. There are prior works which pair synthetically generated queries with real passages to directly fine-tune the retrieval model. Our pipeline differs from these by introducing interaction between student and teacher models that interleaves retrieval-based mining of hard positive/negative examples from the unlabeled corpus with iterative retraining of the student model's weights using these examples. Each retrieval iteration uses the refined student model to mine the corpus for progressively harder training examples for the subsequent training iteration. The methodology provides a cost-effective solution to bridging the gap between general-purpose models and specialized domains without requiring labor-intensive human annotation.

We investigate the short-context dominance hypothesis: that for most sequences, a small local prefix suffices to predict their next tokens. Using large language models as statistical oracles, we measure the minimum context length (MCL) needed to reproduce accurate full-context predictions across datasets with sequences of varying lengths. For sequences with 1-7k tokens from long-context documents, we consistently find that 75-80% require only the last 96 tokens at most. Given the dominance of short-context tokens, we then ask whether it is possible to detect challenging long-context sequences for which a short local prefix does not suffice for prediction. We introduce a practical proxy to MCL, called Distributionally Aware MCL (DaMCL), that does not require knowledge of the actual next-token and is compatible with sampling strategies beyond greedy decoding. Our experiments validate that simple thresholding of the metric defining DaMCL achieves high performance in detecting long vs. short context sequences. Finally, to counter the bias that short-context dominance induces in LLM output distributions, we develop an intuitive decoding algorithm that leverages our detector to identify and boost tokens that are long-range-relevant. Across Q&A tasks and model architectures, we confirm that mitigating the bias improves performance.

We propose an interpretable deep competing risks model called the Deep Kernel Aalen-Johansen (DKAJ) estimator, which generalizes the classical Aalen-Johansen nonparametric estimate of cumulative incidence functions (CIFs). Each data point (e.g., patient) is represented as a weighted combination of clusters. If a data point has nonzero weight only for one cluster, then its predicted CIFs correspond to those of the classical Aalen-Johansen estimator restricted to data points from that cluster. These weights come from an automatically learned kernel function that measures how similar any two data points are. On four standard competing risks datasets, we show that DKAJ is competitive with state-of-the-art baselines while being able to provide visualizations to assist model interpretation.

Pretrained Large Language Models (LLMs) have achieved remarkable success across diverse domains, with education and research emerging as particularly impactful areas. Among current state-of-the-art LLMs, ChatGPT and DeepSeek exhibit strong capabilities in mathematics, science, medicine, literature, and programming. In this study, we present a comprehensive evaluation of these two LLMs through background technology analysis, empirical experiments, and a real-world user survey. The evaluation explores trade-offs among model accuracy, computational efficiency, and user experience in educational and research affairs. We benchmarked these LLMs performance in text generation, programming, and specialized problem-solving. Experimental results show that ChatGPT excels in general language understanding and text generation, while DeepSeek demonstrates superior performance in programming tasks due to its efficiency-focused design. Moreover, both models deliver medically accurate diagnostic outputs and effectively solve complex mathematical problems. Complementing these quantitative findings, a survey of students, educators, and researchers highlights the practical benefits and limitations of these models, offering deeper insights into their role in advancing education and research.

Clinical decision-making in oncology requires predicting dynamic disease evolution, a task current static AI predictors cannot perform. While world models (WMs) offer a paradigm for generative prediction, existing medical applications remain limited. Existing methods often rely on stochastic diffusion models, focusing on visual reconstruction rather than causal, physiological transitions. Furthermore, in medical domain, models like MeWM typically ignore patient-specific temporal and clinical contexts and lack a feedback mechanism to link predictions to treatment decisions. To address these gaps, we introduce CLARITY, a medical world model that forecasts disease evolution directly within a structured latent space. It explicitly integrates time intervals (temporal context) and patient-specific data (clinical context) to model treatment-conditioned progression as a smooth, interpretable trajectory, and thus generate physiologically faithful, individualized treatment plans. Finally, CLARITY introduces a novel prediction-to-decision framework, translating latent rollouts into transparent, actionable recommendations. CLARITY demonstrates state-of-the-art performance in treatment planning. On the MU-Glioma-Post dataset, our approach outperforms recent MeWM by 12\%, and significantly surpasses all other medical-specific large language models.

Cartographic reasoning is the skill of interpreting geographic relationships by aligning legends, map scales, compass directions, map texts, and geometries across one or more map images. Although essential as a concrete cognitive capability and for critical tasks such as disaster response and urban planning, it remains largely unevaluated. Building on progress in chart and infographic understanding, recent large vision language model studies on map visual question-answering often treat maps as a special case of charts. In contrast, map VQA demands comprehension of layered symbology (e.g., symbols, geometries, and text labels) as well as spatial relations tied to orientation and distance that often span multiple maps and are not captured by chart-style evaluations. To address this gap, we introduce FRIEDA, a benchmark for testing complex open-ended cartographic reasoning in LVLMs. FRIEDA sources real map images from documents and reports in various domains and geographical areas. Following classifications in Geographic Information System (GIS) literature, FRIEDA targets all three categories of spatial relations: topological (border, equal, intersect, within), metric (distance), and directional (orientation). All questions require multi-step inference, and many require cross-map grounding and reasoning. We evaluate eleven state-of-the-art LVLMs under two settings: (1) the direct setting, where we provide the maps relevant to the question, and (2) the contextual setting, where the model may have to identify the maps relevant to the question before reasoning. Even the strongest models, Gemini-2.5-Pro and GPT-5-Think, achieve only 38.20% and 37.20% accuracy, respectively, far below human performance of 84.87%. These results reveal a persistent gap in multi-step cartographic reasoning, positioning FRIEDA as a rigorous benchmark to drive progress on spatial intelligence in LVLMs.

Gestures are an integral part of our daily interactions with the environment. Hand gesture recognition (HGR) is the process of interpreting human intent through various input modalities, such as visual data (images and videos) and bio-signals. Bio-signals are widely used in HGR due to their ability to be captured non-invasively via sensors placed on the arm. Among these, surface electromyography (sEMG), which measures the electrical activity of muscles, is the most extensively studied modality. However, less-explored alternatives such as inertial measurement units (IMUs) can provide complementary information on subtle muscle movements, which makes them valuable for gesture recognition. In this study, we investigate the potential of using IMU signals from different muscle groups to capture user intent. Our results demonstrate that IMU signals contain sufficient information to serve as the sole input sensor for static gesture recognition. Moreover, we compare different muscle groups and check the quality of pattern recognition on individual muscle groups. We further found that tendon-induced micro-movement captured by IMUs is a major contributor to static gesture recognition. We believe that leveraging muscle micro-movement information can enhance the usability of prosthetic arms for amputees. This approach also offers new possibilities for hand gesture recognition in fields such as robotics, teleoperation, sign language interpretation, and beyond.