This study investigates the feasibility and performance of federated learning (FL) for multi-label ICD code classification using clinical notes from the MIMIC-IV dataset. Unlike previous approaches that rely on centralized training or fine-tuned large language models, we propose a lightweight and scalable pipeline combining frozen text embeddings with simple multilayer perceptron (MLP) classifiers. This design offers a privacy-preserving and deployment-efficient alternative for clinical NLP applications, particularly suited to distributed healthcare settings. Extensive experiments across both centralized and federated configurations were conducted, testing six publicly available embedding models from Massive Text Embedding Benchmark leaderboard and three MLP classifier architectures under two medical coding (ICD-9 and ICD-10). Additionally, ablation studies over ten random stratified splits assess performance stability. Results show that embedding quality substantially outweighs classifier complexity in determining predictive performance, and that federated learning can closely match centralized results in idealized conditions. While the models are orders of magnitude smaller than state-of-the-art architectures and achieved competitive micro and macro F1 scores, limitations remain including the lack of end-to-end training and the simplified FL assumptions. Nevertheless, this work demonstrates a viable way toward scalable, privacy-conscious medical coding systems and offers a step toward for future research into federated, domain-adaptive clinical AI.

The UK has pursued a distinctive path in AI regulation: less cautious than the EU but more willing to address risks than the US, and has emerged as a global leader in coordinating AI safety efforts. Impressive developments from companies like London-based DeepMind began to spark concerns in the UK about catastrophic risks from around 2012, although regulatory discussion at the time focussed on bias and discrimination. By 2022, these discussions had evolved into a "pro-innovation" strategy, in which the government directed existing regulators to take a light-touch approach, governing AI at point of use, but avoided regulating the technology or infrastructure directly. ChatGPT arrived in late 2022, galvanising concerns that this approach may be insufficient. The UK responded by establishing an AI Safety Institute to monitor risks and hosting the first international AI Safety Summit in 2023, but - unlike the EU - refrained from regulating frontier AI development in addition to its use. A new government was elected in 2024 which promised to address this gap, but at the time of writing is yet to do so. What should the UK do next? The government faces competing objectives: harnessing AI for economic growth and better public services while mitigating risk. In light of these, we propose establishing a flexible, principles-based regulator to oversee the most advanced AI development, defensive measures against risks from AI-enabled biological design tools, and argue that more technical work is needed to understand how to respond to AI-generated misinformation. We argue for updated legal frameworks on copyright, discrimination, and AI agents, and that regulators will have a limited but important role if AI substantially disrupts labour markets. If the UK gets AI regulation right, it could demonstrate how democratic societies can harness AI's benefits while managing its risks.

Memory storage for Large Language models (LLMs) is becoming an increasingly active area of research, particularly for enabling personalization across long conversations. W e propose Pref-LSTM, a dynamic and lightweight framework that combines a BERT-based classifier with a LSTM memory module that generates memory embedding which then is soft-prompt injected into a frozen LLM. W e synthetically curate a dataset of preference and non-preference conversation turns to train our BERT-based classifier . Although our LSTM-based memory encoder did not yield strong results, we find that the BERT-based classifier performs reliably in identifying explicit and implicit user preferences. Our research demonstrates the viability of using preference filtering with LSTM gating principals as an efficient path towards scalable user preference modeling, without extensive overhead and fine-tuning.

Background: Clinical documentation represents a significant burden for healthcare providers, with physicians spending up to 2 hours daily on administrative tasks. Recent advances in large language models (LLMs) offer promising solutions, but privacy concerns and computational requirements limit their adoption in healthcare settings. Objective: To develop and evaluate a privacy-preserving, on-device medical transcription system using a fine-tuned Llama 3.2 1B model capable of generating structured medical notes from medical transcriptions while maintaining complete data sovereignty entirely in the browser. Methods: We fine-tuned a Llama 3.2 1B model using Parameter-Efficient Fine-Tuning (PEFT) with LoRA on 1,500 synthetic medical transcription-to-structured note pairs. The model was evaluated against the base Llama 3.2 1B on two datasets: 100 endocrinology transcripts and 140 modified ACI benchmark cases. Evaluation employed both statistical metrics (ROUGE, BERTScore, BLEURT) and LLM-as-judge assessments across multiple clinical quality dimensions. Results: The fine-tuned OnDevice model demonstrated substantial improvements over the base model. On the ACI benchmark, ROUGE-1 scores increased from 0.346 to 0.496, while BERTScore F1 improved from 0.832 to 0.866. Clinical quality assessments showed marked reduction in major hallucinations (from 85 to 35 cases) and enhanced factual correctness (2.81 to 3.54 on 5-point scale). Similar improvements were observed on the internal evaluation dataset, with composite scores increasing from 3.13 to 4.43 (+41.5%). Conclusions: Fine-tuning compact LLMs for medical transcription yields clinically meaningful improvements while enabling complete on-device browser deployment. This approach addresses key barriers to AI adoption in healthcare: privacy preservation, cost reduction, and accessibility for resource-constrained environments.

Multimodal Large Language Models (MLLMs) have achieved remarkable progress in multimodal reasoning. However, they often excessively rely on textual information during the later stages of inference, neglecting the crucial integration of visual input. Current methods typically address this by explicitly injecting visual information to guide the reasoning process. In this work, through an analysis of MLLM attention patterns, we made an intriguing observation: with appropriate guidance, MLLMs can spontaneously re-focus their attention on visual inputs during the later stages of reasoning, even without explicit visual information injection. This spontaneous shift in focus suggests that MLLMs are intrinsically capable of performing visual fusion reasoning. Building on this insight, we introduce Look-Back, an implicit approach designed to guide MLLMs to ``look back" at visual information in a self-directed manner during reasoning. Look-Back empowers the model to autonomously determine when, where, and how to re-focus on visual inputs, eliminating the need for explicit model-structure constraints or additional input. We demonstrate that Look-Back significantly enhances the model's reasoning and perception capabilities, as evidenced by extensive empirical evaluations on multiple multimodal benchmarks.

Generative AI systems have rapidly advanced, with multimodal input capabilities enabling reasoning beyond text-based tasks. In education, these advancements could influence assessment design and question answering, presenting both opportunities and challenges. To investigate these effects, we introduce a high-quality dataset of 201 university-level STEM questions, manually annotated with features such as image type, role, problem complexity, and question format. Our study analyzes how these features affect generative AI performance compared to students. We evaluate four model families with five prompting strategies, comparing results to the average of 546 student responses per question. Although the best model correctly answers on average 58.5 % of the questions using majority vote aggregation, human participants consistently outperform AI on questions involving visual components. Interestingly, human performance remains stable across question features but varies by subject, whereas AI performance is susceptible to both subject matter and question features. Finally, we provide actionable insights for educators, demonstrating how question design can enhance academic integrity by leveraging features that challenge current AI systems without increasing the cognitive burden for students.

We evaluate language models' ability to subvert monitoring protocols via collusion. More specifically, we have two instances of a model design prompts for a policy (P) and a monitor (M) in a programming task setting. The models collaboratively aim for M to classify all backdoored programs in an auditing dataset as harmful, but nevertheless classify a backdoored program produced by P as harmless. The models are isolated from each other, requiring them to independently arrive at compatible subversion strategies. We find that while Claude 3.7 Sonnet has low success rate due to poor convergence, it sometimes successfully colludes on non-obvious signals.

Recent advances in large language models (LLMs) have led to increasingly sophisticated safety protocols and features designed to prevent harmful, unethical, or unauthorized outputs. However, these guardrails remain susceptible to novel and creative forms of adversarial prompting, including manually generated test cases. In this work, we present two new test cases in mental health for (i) suicide and (ii) self-harm, using multi-step, prompt-level jailbreaking and bypass built-in content and safety filters. We show that user intent is disregarded, leading to the generation of detailed harmful content and instructions that could cause real-world harm. We conduct an empirical evaluation across six widely available LLMs, demonstrating the generalizability and reliability of the bypass. We assess these findings and the multilayered ethical tensions that they present for their implications on prompt-response filtering and context- and task-specific model development. We recommend a more comprehensive and systematic approach to AI safety and ethics while emphasizing the need for continuous adversarial testing in safety-critical AI deployments. We also argue that while certain clearly defined safety measures and guardrails can and must be implemented in LLMs, ensuring robust and comprehensive safety across all use cases and domains remains extremely challenging given the current technical maturity of general-purpose LLMs.

Competency Questions (CQs) are pivotal in knowledge engineering, guiding the design, validation, and testing of ontologies. A number of diverse formulation approaches have been proposed in the literature, ranging from completely manual to Large Language Model (LLM) driven ones. However, attempts to characterise the outputs of these approaches and their systematic comparison are scarce. This paper presents an empirical comparative evaluation of three distinct CQ formulation approaches: manual formulation by ontology engineers, instantiation of CQ patterns, and generation using state of the art LLMs. We generate CQs using each approach from a set of requirements for cultural heritage, and assess them across different dimensions: degree of acceptability, ambiguity, relevance, readability and complexity. Our contribution is twofold: (i) the first multi-annotator dataset of CQs generated from the same source using different methods; and (ii) a systematic comparison of the characteristics of the CQs resulting from each approach. Our study shows that different CQ generation approaches have different characteristics and that LLMs can be used as a way to initially elicit CQs, however these are sensitive to the model used to generate CQs and they generally require a further refinement step before they can be used to model requirements.

The enhancement of mathematical capabilities in large language models (LLMs) fosters new developments in mathematics education within primary and secondary schools, particularly as they relate to intelligent tutoring systems. However, LLMs require substantial computational resources, resulting in significant costs in educational contexts. To mitigate this drawback, this paper investigates the feasibility of compressing LLMs for solving math word problems (MWPs). We compress the embedded vectors encoded by BERT and distill a considerably smaller student model. Our findings indicate that the student model can maintain nearly 90% of the performance of the teacher model while utilizing only 1/12 of its parameters. In addition to achieving high accuracy, the model exhibits strong generalizability, as the compressed vectors perform well across all tasks related to MWPs, and the distillation process is not task-specific. The success of this distillation demonstrates that the underlying principles are generic and not limited to a specific task. We further explore the reasons behind the compressibility of embedded vectors, revealing that part-of-speech information, rather than entity recognition, is crucial for MWPs, which may significantly contribute to their compressibility. The improvements in efficiency and cost reduction provide substantial value for intelligent tutoring systems and significantly advance the field of intelligent education.