Ensuring that large language models (LLMs) are both helpful and harmless is a critical challenge, as overly strict constraints can lead to excessive refusals, while permissive models risk generating harmful content. Existing approaches, such as reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO), attempt to balance these trade-offs but suffer from performance conflicts, limited controllability, and poor extendability. To address these issues, we propose Preference Vector, a novel framework inspired by task arithmetic. Instead of optimizing multiple preferences within a single objective, we train separate models on individual preferences, extract behavior shifts as preference vectors, and dynamically merge them at test time. This modular approach enables fine-grained, user-controllable preference adjustments and facilitates seamless integration of new preferences without retraining. Experiments show that our proposed Preference Vector framework improves helpfulness without excessive conservatism, allows smooth control over preference trade-offs, and supports scalable multi-preference alignment.

As modern software systems grow in complexity and scale, their ability to autonomously detect, diagnose, and recover from failures becomes increasingly vital. Drawing inspiration from biological healing--where the human body detects damage, signals the brain, and activates targeted recovery--this paper explores the concept of self-healing software driven by artificial intelligence. We propose a novel framework that mimics this biological model: system observability tools serve as sensory inputs, AI models function as the cognitive core for diagnosis and repair, and healing agents apply targeted code and test modifications. By combining log analysis, static code inspection, and AI-driven generation of patches or test updates, our approach aims to reduce downtime, accelerate debugging, and enhance software resilience. We evaluate the effectiveness of this model through case studies and simulations, comparing it against traditional manual debugging and recovery workflows. This work paves the way toward intelligent, adaptive, and self-reliant software systems capable of continuous healing, akin to living organisms.

Personalized food recommendation systems (Food-RecSys) critically underperform due to fragmented component understanding and the failure of conventional machine learning with vast, imbalanced food data. While Large Language Models (LLMs) offer promise, current generic Recommendation as Language Processing (RLP) strategies lack the necessary specialization for the food domain's complexity. This thesis tackles these deficiencies by first identifying and analyzing the essential components for effective Food-RecSys. We introduce two key innovations: a multimedia food logging platform for rich contextual data acquisition and the World Food Atlas, enabling unique geolocation-based food analysis previously unavailable. Building on this foundation, we pioneer the Food Recommendation as Language Processing (F-RLP) framework - a novel, integrated approach specifically architected for the food domain. F-RLP leverages LLMs in a tailored manner, overcoming the limitations of generic models and providing a robust infrastructure for effective, contextual, and truly personalized food recommendations.

Clinical trials are crucial for assessing new treatments; however, recruitment challenges - such as limited awareness, complex eligibility criteria, and referral barriers - hinder their success. With the growth of online platforms, patients increasingly turn to social media and health communities for support, research, and advocacy, expanding recruitment pools and established enrollment pathways. Recognizing this potential, we utilized TrialGPT, a framework that leverages a large language model (LLM) as its backbone, to match 50 online patient cases (collected from published case reports and a social media website) to clinical trials and evaluate performance against traditional keyword-based searches. Our results show that TrialGPT outperforms traditional methods by 46% in identifying eligible trials, with each patient, on average, being eligible for around 7 trials. Additionally, our outreach efforts to case authors and trial organizers regarding these patient-trial matches yielded highly positive feedback, which we present from both perspectives.

Multiword expressions, characterised by non-compositional meanings and syntactic irregularities, are an example of nuanced language. These expressions can be used literally or idiomatically, leading to significant changes in meaning. While large language models have demonstrated strong performance across many tasks, their ability to handle such linguistic subtleties remains uncertain. Therefore, this study evaluates how state-of-the-art language models process the ambiguity of potentially idiomatic multiword expressions, particularly in contexts that are less frequent, where models are less likely to rely on memorisation. By evaluating models across in Portuguese and Galician, in addition to English, and using a novel code-switched dataset and a novel task, we find that large language models, despite their strengths, struggle with nuanced language. In particular, we find that the latest models, including GPT-4, fail to outperform the xlm-roBERTa-base baselines in both detection and semantic tasks, with especially poor performance on the novel tasks we introduce, despite its similarity to existing tasks. Overall, our results demonstrate that multiword expressions, especially those which are ambiguous, continue to be a challenge to models.

Online fake news moderation now faces a new challenge brought by the malicious use of large language models (LLMs) in fake news production. Though existing works have shown LLM-generated fake news is hard to detect from an individual aspect, it remains underexplored how its large-scale release will impact the news ecosystem. In this study, we develop a simulation pipeline and a dataset with ~56k generated news of diverse types to investigate the effects of LLM-generated fake news within neural news recommendation systems. Our findings expose a truth decay phenomenon, where real news is gradually losing its advantageous position in news ranking against fake news as LLM-generated news is involved in news recommendation. We further provide an explanation about why truth decay occurs from a familiarity perspective and show the positive correlation between perplexity and news ranking. Finally, we discuss the threats of LLM-generated fake news and provide possible countermeasures. We urge stakeholders to address this emerging challenge to preserve the integrity of news ecosystems.

Educational question generation (EQG) is a crucial component of intelligent educational systems, significantly aiding self-assessment, active learning, and personalized education. While EQG systems have emerged, existing datasets typically rely on predefined, carefully edited texts, failing to represent real-world classroom content, including lecture speech with a set of complementary slides. To bridge this gap, we collect a dataset of educational questions based on lectures from real-world classrooms. On this realistic dataset, we find that current methods for EQG struggle with accurately generating questions from educational videos, particularly in aligning with specific timestamps and target answers. Common challenges include selecting informative contexts from extensive transcripts and ensuring generated questions meaningfully incorporate the target answer. To address the challenges, we introduce a novel framework utilizing large language models for dynamically selecting and rewriting contexts based on target timestamps and answers. First, our framework selects contexts from both lecture transcripts and video keyframes based on answer relevance and temporal proximity. Then, we integrate the contexts selected from both modalities and rewrite them into answer-containing knowledge statements, to enhance the logical connection between the contexts and the desired answer. This approach significantly improves the quality and relevance of the generated questions. Our dataset and code are released in https://github.com/mengxiayu/COSER.

The trend towards large language models (LLMs) for guardrailing against undesired behaviors is increasing and has shown promise for censoring user inputs. However, increased latency, memory consumption, hosting expenses and non-structured outputs can make their use prohibitive. In this work, we show that task-specific data generation can lead to fine-tuned classifiers that significantly outperform current state of the art (SoTA) while being orders of magnitude smaller. Secondly, we show that using a single model, \texttt{MultiTaskGuard}, that is pretrained on a large synthetically generated dataset with unique task instructions further improves generalization. Thirdly, our most performant models, \texttt{UniGuard}, are found using our proposed search-based model merging approach that finds an optimal set of parameters to combine single-policy models and multi-policy guardrail models. % On 7 public datasets and 4 guardrail benchmarks we created, our efficient guardrail classifiers improve over the best performing SoTA publicly available LLMs and 3$^{\text{rd}}$ party guardrail APIs in detecting unsafe and safe behaviors by an average F1 score improvement of \textbf{29.92} points over Aegis-LlamaGuard and \textbf{21.62} over \texttt{gpt-4o}, respectively. Lastly, our guardrail synthetic data generation process that uses custom task-specific guardrail poli

Mixture-of-Experts (MoE) have emerged as a powerful architecture for large language models (LLMs), enabling efficient scaling of model capacity while maintaining manageable computational costs. The key advantage lies in their ability to route different tokens to different ``expert'' networks within the model, enabling specialization and efficient handling of diverse input. However, the vulnerabilities of MoE-based LLMs still have barely been studied, and the potential for backdoor attacks in this context remains largely unexplored. This paper presents the first backdoor attack against MoE-based LLMs where the attackers poison ``dormant experts'' (i.e., underutilized experts) and activate them by optimizing routing triggers, thereby gaining control over the model's output. We first rigorously prove the existence of a few ``dominating experts'' in MoE models, whose outputs can determine the overall MoE's output. We also show that dormant experts can serve as dominating experts to manipulate model predictions. Accordingly, our attack, namely BadMoE, exploits the unique architecture of MoE models by 1) identifying dormant experts unrelated to the target task, 2) constructing a routing-aware loss to optimize the activation triggers of these experts, and 3) promoting dormant experts to dominating roles via poisoned training data. Extensive experiments show that BadMoE successfully enforces malicious prediction on attackers' target tasks while preserving overall model utility, making it a more potent and stealthy attack than existing methods.

High-resolution image (HRI) understanding aims to process images with a large number of pixels, such as pathological images and agricultural aerial images, both of which can exceed 1 million pixels. Vision Large Language Models (VLMs) can allegedly handle HRIs, however, there is a lack of a comprehensive benchmark for VLMs to evaluate HRI understanding. To address this gap, we introduce HRScene, a novel unified benchmark for HRI understanding with rich scenes. HRScene incorporates 25 real-world datasets and 2 synthetic diagnostic datasets with resolutions ranging from 1,024 $\times$ 1,024 to 35,503 $\times$ 26,627. HRScene is collected and re-annotated by 10 graduate-level annotators, covering 25 scenarios, ranging from microscopic to radiology images, street views, long-range pictures, and telescope images. It includes HRIs of real-world objects, scanned documents, and composite multi-image. The two diagnostic evaluation datasets are synthesized by combining the target image with the gold answer and distracting images in different orders, assessing how well models utilize regions in HRI. We conduct extensive experiments involving 28 VLMs, including Gemini 2.0 Flash and GPT-4o. Experiments on HRScene show that current VLMs achieve an average accuracy of around 50% on real-world tasks, revealing significant gaps in HRI understanding. Results on synthetic datasets reveal that VLMs struggle to effectively utilize HRI regions, showing significant Regional Divergence and lost-in-middle, shedding light on future research.