Domain-specific QA systems require not just generative fluency but high factual accuracy grounded in structured expert knowledge. While recent Retrieval-Augmented Generation (RAG) frameworks improve context recall, they struggle with integrating heterogeneous data and maintaining reasoning consistency. To address these challenges, we propose DO-RAG, a scalable and customizable hybrid QA framework that integrates multi-level knowledge graph construction with semantic vector retrieval. Our system employs a novel agentic chain-of-thought architecture to extract structured relationships from unstructured, multimodal documents, constructing dynamic knowledge graphs that enhance retrieval precision. At query time, DO-RAG fuses graph and vector retrieval results to generate context-aware responses, followed by hallucination mitigation via grounded refinement. Experimental evaluations in the database and electrical domains show near-perfect recall and over 94% answer relevancy, with DO-RAG outperforming baseline frameworks by up to 33.38%. By combining traceability, adaptability, and performance efficiency, DO-RAG offers a reliable foundation for multi-domain, high-precision QA at scale.

AI is transforming education by enabling powerful tools that enhance learning experiences. Among recent advancements, large language models (LLMs) hold particular promise for revolutionizing how learners interact with educational content. In this work, we investigate the potential of LLMs to support standardized test preparation by focusing on English Standardized Tests (ESTs). Specifically, we assess their ability to generate accurate and contextually appropriate solutions across a diverse set of EST question types. We introduce ESTBOOK, a comprehensive benchmark designed to evaluate the capabilities of LLMs in solving EST questions. ESTBOOK aggregates five widely recognized tests, encompassing 29 question types and over 10,576 questions across multiple modalities, including text, images, audio, tables, and mathematical symbols. Using ESTBOOK, we systematically evaluate both the accuracy and inference efficiency of LLMs. Additionally, we propose a breakdown analysis framework that decomposes complex EST questions into task-specific solution steps. This framework allows us to isolate and assess LLM performance at each stage of the reasoning process. Evaluation findings offer insights into the capability of LLMs in educational contexts and point toward targeted strategies for improving their reliability as intelligent tutoring systems.

Recent advances in transformer architectures have revolutionised natural language processing, but their application to healthcare domains presents unique challenges. Patient timelines are characterised by irregular sampling, variable temporal dependencies, and complex contextual relationships that differ substantially from traditional language tasks. This paper introduces \METHOD~(Modular Efficient Transformer for Health Outcome Discovery), a novel transformer architecture specifically designed to address the challenges of clinical sequence modelling in electronic health records. \METHOD~integrates three key innovations: (1) a patient-aware attention mechanism that prevents information leakage whilst enabling efficient batch processing; (2) an adaptive sliding window attention scheme that captures multi-scale temporal dependencies; and (3) a U-Net inspired architecture with dynamic skip connections for effective long sequence processing. Evaluations on the MIMIC-IV database demonstrate that \METHOD~consistently outperforms the state-of-the-art \ETHOS~model, particularly in predicting high-severity cases that require urgent clinical intervention. \METHOD~exhibits stable performance across varying inference lengths, a crucial feature for clinical deployment where patient histories vary significantly in length. Analysis of learned embeddings reveals that \METHOD~better preserves clinical hierarchies and relationships between medical concepts. These results suggest that \METHOD~represents a significant advancement in transformer architectures optimised for healthcare applications, providing more accurate and clinically relevant predictions whilst maintaining computational efficiency.

While large language models (LLMs) have demonstrated remarkable capabilities in understanding human languages, this study explores how they translate this understanding into social exchange contexts that capture certain essences of real world human interactions. Three LLMs - ChatGPT-4, Claude, and Bard - were placed in economic trust games where players balance self-interest with trust and reciprocity, making decisions that reveal their social preferences and interactive reasoning abilities. Our study shows that LLMs deviate from pure self-interest and exhibit trust and reciprocity even without being prompted to adopt a specific persona. In the simplest one-shot interaction, LLMs emulated how human players place trust at the beginning of such a game. Larger human-machine divergences emerged in scenarios involving trust repayment or multi-round interactions, where decisions were influenced by both social preferences and interactive reasoning. LLMs responses varied significantly when prompted to adopt personas like selfish or unselfish players, with the impact outweighing differences between models or game types. Response of ChatGPT-4, in an unselfish or neutral persona, resembled the highest trust and reciprocity, surpassing humans, Claude, and Bard. Claude and Bard displayed trust and reciprocity levels that sometimes exceeded and sometimes fell below human choices. When given selfish personas, all LLMs showed lower trust and reciprocity than humans. Interactive reasoning to the actions of counterparts or changing game mechanics appeared to be random rather than stable, reproducible characteristics in the response of LLMs, though some improvements were observed when ChatGPT-4 responded in selfish or unselfish personas.

Large language models (LLMs) excel at generating contextually relevant content. However, tailoring these outputs to individual users for effective personalization is a significant challenge. While rich user-specific information often exists as pre-existing user representations, such as embeddings learned from preferences or behaviors, current methods to leverage these for LLM personalization typically require costly fine-tuning or token-heavy prompting. We propose Embedding-to-Prefix (E2P), a parameter-efficient method that injects pre-computed context embeddings into an LLM's hidden representation space through a learned projection to a single soft token prefix. This enables effective personalization while keeping the backbone model frozen and avoiding expensive adaptation techniques. We evaluate E2P across two public datasets and in a production setting: dialogue personalization on Persona-Chat, contextual headline generation on PENS, and large-scale personalization for music and podcast consumption. Results show that E2P preserves contextual signals and achieves strong performance with minimal computational overhead, offering a scalable, efficient solution for contextualizing generative AI systems.

In medical practices across the United States, physicians have begun implementing generative artificial intelligence (AI) tools to perform the function of scribes in order to reduce the burden of documenting clinical encounters. Despite their widespread use, no established methods exist to gauge the quality of AI scribes. To address this gap, we developed a blinded study comparing the relative performance of large language model (LLM) generated clinical notes with those from field experts based on audio-recorded clinical encounters. Quantitative metrics from the Physician Documentation Quality Instrument (PDQI9) provided a framework to measure note quality, which we adapted to assess relative performance of AI generated notes. Clinical experts spanning 5 medical specialties used the PDQI9 tool to evaluate specialist-drafted Gold notes and LLM authored Ambient notes. Two evaluators from each specialty scored notes drafted from a total of 97 patient visits. We found uniformly high inter rater agreement (RWG greater than 0.7) between evaluators in general medicine, orthopedics, and obstetrics and gynecology, and moderate (RWG 0.5 to 0.7) to high inter rater agreement in pediatrics and cardiology. We found a modest yet significant difference in the overall note quality, wherein Gold notes achieved a score of 4.25 out of 5 and Ambient notes scored 4.20 out of 5 (p = 0.04). Our findings support the use of the PDQI9 instrument as a practical method to gauge the quality of LLM authored notes, as compared to human-authored notes.

Massive and diverse web data are increasingly vital for government disaster response, as demonstrated by the 2022 floods in New South Wales (NSW), Australia. This study examines how X (formerly Twitter) and public inquiry submissions provide insights into public behaviour during crises. We analyse more than 55,000 flood-related tweets and 1,450 submissions to identify behavioural patterns during extreme weather events. While social media posts are short and fragmented, inquiry submissions are detailed, multi-page documents offering structured insights. Our methodology integrates Latent Dirichlet Allocation (LDA) for topic modelling with Large Language Models (LLMs) to enhance semantic understanding. LDA reveals distinct opinions and geographic patterns, while LLMs improve filtering by identifying flood-relevant tweets using public submissions as a reference. This Relevance Index method reduces noise and prioritizes actionable content, improving situ-ational awareness for emergency responders. By combining these complementary data streams, our approach introduces a novel AI-driven method to refine crisis-related social media content, improve real-time disaster response, and inform long-term resilience planning.

Prompt engineering has emerged as a critical component in optimizing large language models (LLMs) for domain-specific tasks. However, the role of prompt specificity, especially in domains like STEM (physics, chemistry, biology, computer science and mathematics), medicine, and law, remains underexplored. This thesis addresses the problem of whether increasing the specificity of vocabulary in prompts improves LLM performance in domain-specific question-answering and reasoning tasks. We developed a synonymization framework to systematically substitute nouns, verbs, and adjectives with varying specificity levels, measuring the impact on four LLMs: Llama-3.1-70B-Instruct, Granite-13B-Instruct-V2, Flan-T5-XL, and Mistral-Large 2, across datasets in STEM, law, and medicine. Our results reveal that while generally increasing the specificity of prompts does not have a significant impact, there appears to be a specificity range, across all considered models, where the LLM performs the best. Identifying this optimal specificity range offers a key insight for prompt design, suggesting that manipulating prompts within this range could maximize LLM performance and lead to more efficient applications in specialized domains.

Voice Conversion research in recent times has increasingly focused on improving the zero-shot capabilities of existing methods. Despite remarkable advancements, current architectures still tend to struggle in zero-shot cross-lingual settings. They are also often unable to generalize for speakers of unseen languages and accents. In this paper, we adopt a simple yet effective approach that combines discrete speech representations from self-supervised models with a non-autoregressive Diffusion-Transformer based conditional flow matching speech decoder. We show that this architecture allows us to train a voice-conversion model in a purely textless, self-supervised fashion. Our technique works without requiring multiple encoders to disentangle speech features. Our model also manages to excel in zero-shot cross-lingual settings even for unseen languages. For Demo: https://ez-vc.github.io/EZ-VC-Demo/

Large reasoning models (LRMs) have demonstrated strong performance on complex reasoning tasks, but often suffer from overthinking, generating redundant content regardless of task difficulty. Inspired by the dual process theory in cognitive science, we propose Adaptive Cognition Policy Optimization (ACPO), a reinforcement learning framework that enables LRMs to achieve efficient reasoning through adaptive cognitive allocation and dynamic system switch. ACPO incorporates two key components: (1) introducing system-aware reasoning tokens to explicitly represent the thinking modes thereby making the model's cognitive process transparent, and (2) integrating online difficulty estimation and token length budget to guide adaptive system switch and reasoning during reinforcement learning. To this end, we propose a two-stage training strategy. The first stage begins with supervised fine-tuning to cold start the model, enabling it to generate reasoning paths with explicit thinking modes. In the second stage, we apply ACPO to further enhance adaptive system switch for difficulty-aware reasoning. Experimental results demonstrate that ACPO effectively reduces redundant reasoning while adaptively adjusting cognitive allocation based on task complexity, achieving efficient hybrid reasoning.