Visual question answering (VQA) in medical imaging aims to support clinical diagnosis by automatically interpreting complex imaging data in response to natural language queries. Existing studies typically rely on distinct visual and textual encoders to independently extract features from medical images and clinical questions, which are subsequently combined to generate answers. Specifically, in computed tomography (CT), such approaches are similar to the conventional practices in medical image analysis. However, these approaches pay less attention to the spatial continuity and inter-slice correlations in the volumetric CT data, leading to fragmented and imprecise responses. In this paper, we propose a novel large language model (LLM)-based framework enhanced by a graph representation of salient features. Different from conventional multimodal encoding strategies, our approach constructs a cross-modal graph integrating both visual and textual features, treating individual CT slices and question tokens as nodes within the graph. We further leverage an attentive graph convolutional network to dynamically fuse information within this structure. The resulting aggregated graph features then serve as a soft prompt to guide a large language model in generating accurate answers. Extensive experiments on the M3D-VQA benchmark demonstrate that our approach consistently outperforms baselines across multiple evaluation metrics, offering more robust reasoning capabilities.

Mobile GUI agents are designed to autonomously execute diverse device-control tasks by interpreting and interacting with mobile screens. Despite notable advancements, their resilience in real-world scenarios where screen content may be partially manipulated by untrustworthy third parties remains largely unexplored. Owing to their black-box and autonomous nature, these agents are vulnerable to manipulations that could compromise user devices. In this work, we present the first systematic investigation into the vulnerabilities of mobile GUI agents. We introduce a scalable attack simulation framework AgentHazard, which enables flexible and targeted modifications of screen content within existing applications. Leveraging this framework, we develop a comprehensive benchmark suite comprising both a dynamic task execution environment and a static dataset of vision-language-action tuples, totaling over 3,000 attack scenarios. The dynamic environment encompasses 58 reproducible tasks in an emulator with various types of hazardous UI content, while the static dataset is constructed from 210 screenshots collected from 14 popular commercial apps. Importantly, our content modifications are designed to be feasible for unprivileged third parties. We evaluate 7 widely-used mobile GUI agents and 5 common backbone models using our benchmark. Our findings reveal that all examined agents are significantly influenced by misleading third-party content (with an average misleading rate of 28.8% in human-crafted attack scenarios) and that their vulnerabilities are closely linked to the employed perception modalities and backbone LLMs. Furthermore, we assess training-based mitigation strategies, highlighting both the challenges and opportunities for enhancing the robustness of mobile GUI agents. Our code and data will be released at https://agenthazard.github.io.

Learning rate (LR) schedules in large language model (LLM) training often follow empirical templates: warm-up, constant plateau/stable phase, and decay (WSD). However, the mechanistic explanation for this strategy remains underexplored, and the choice of plateau height and decay schedule is largely heuristic. In this paper, we connect training dynamics to a thermodynamic analogy via the Mpemba effect - a phenomenon in which a hotter system cools faster than a colder one when quenched into the same bath. We analyze a class of "valley-river" loss landscapes, where sharp (valley) directions equilibrate quickly, while flatter (river) directions govern global descent. The Mpemba effect provides an explanation for the necessity of the warm-up phase and motivates a high plateau - rather than a low one - for accelerating loss decrease during decay. We show that for certain loss landscapes, there exists an optimal plateau learning rate - the "strong Mpemba point" - at which the slowest mode vanishes, resulting in faster convergence during the decay phase. We derive analytical conditions for its existence and estimate decay dynamics required to preserve the Mpemba advantage. Our minimal model and analysis offer a principled justification for plateau-based schedulers and provide guidance for tuning LR in LLMs with minimal hyperparameter sweep.

Although the amount of available spoken content is steadily increasing, extracting information and knowledge from speech recordings remains challenging. Beyond enhancing traditional information retrieval methods such as speech search and keyword spotting, novel approaches for navigating and searching spoken content need to be explored and developed. In this paper, we propose a novel navigational method for speech archives that leverages recent advances in language and multimodal generative models. We demonstrate our approach with a Web application that organizes data into a structured format using interactive mind maps and image generation tools. The system is implemented using the TED-LIUM~3 dataset, which comprises over 2,000 speech transcripts and audio files of TED Talks. Initial user tests using a System Usability Scale (SUS) questionnaire indicate the application's potential to simplify the exploration of large speech collections.

Recent works on large language models (LLMs) have demonstrated the impact of prompting strategies and fine-tuning techniques on their reasoning capabilities. Yet, their effectiveness on clinical natural language inference (NLI) remains underexplored. This study presents the first controlled evaluation of how prompt structure and efficient fine-tuning jointly shape model performance in clinical NLI. We inspect four classes of prompting strategies to elicit reasoning in LLMs at different levels of abstraction, and evaluate their impact on a range of clinically motivated reasoning types. For each prompting strategy, we construct high-quality demonstrations using a frontier model to distil multi-step reasoning capabilities into smaller models (4B parameters) via Low-Rank Adaptation (LoRA). Across different language models fine-tuned on the NLI4CT benchmark, we found that prompt type alone accounts for up to 44% of the variance in macro-F1. Moreover, LoRA fine-tuning yields consistent gains of +8 to 12 F1, raises output alignment above 97%, and narrows the performance gap to GPT-4o-mini to within 7.1%. Additional experiments on reasoning generalisation reveal that LoRA improves performance in 75% of the models on MedNLI and TREC Clinical Trials Track. Overall, these findings demonstrate that (i) prompt structure is a primary driver of clinical reasoning performance, (ii) compact models equipped with strong prompts and LoRA can rival frontier-scale systems, and (iii) reasoning-type-aware evaluation is essential to uncover prompt-induced trade-offs. Our results highlight the promise of combining prompt design and lightweight adaptation for more efficient and trustworthy clinical NLP systems, providing insights on the strengths and limitations of widely adopted prompting and parameter-efficient techniques in highly specialised domains.

Retrieval-augmented generation (RAG) enhances large language models (LLMs) by incorporating external documents at inference time, enabling up-to-date knowledge access without costly retraining. However, conventional RAG methods retrieve passages independently, often leading to redundant, noisy, or insufficiently diverse context-particularly problematic - particularly problematic in noisy corpora and for multi-hop questions. To address this, we propose Adaptive Passage Combination Retrieval (AdaPCR), a novel framework for open-domain question answering with black-box LMs. AdaPCR explicitly models dependencies between passages by considering passage combinations as units for retrieval and reranking. It consists of a context-aware query reformulation using concatenated passages, and a reranking step trained with a predictive objective aligned with downstream answer likelihood. Crucially, AdaPCR adaptively selects the number of retrieved passages without additional stopping modules. Experiments across several QA benchmarks show that AdaPCR outperforms baselines, particularly in multi-hop reasoning, demonstrating the effectiveness of modeling inter-passage dependencies for improved retrieval.

Topological materials occupy a frontier in condensed-matter physics thanks to their remarkable electronic and quantum properties, yet their cross-scale design remains bottlenecked by inefficient discovery workflows. Here, we introduce TopoMAS (Topological materials Multi-Agent System), an interactive human-AI framework that seamlessly orchestrates the entire materials-discovery pipeline: from user-defined queries and multi-source data retrieval, through theoretical inference and crystal-structure generation, to first-principles validation. Crucially, TopoMAS closes the loop by autonomously integrating computational outcomes into a dynamic knowledge graph, enabling continuous knowledge refinement. In collaboration with human experts, it has already guided the identification of novel topological phases SrSbO3, confirmed by first-principles calculations. Comprehensive benchmarks demonstrate robust adaptability across base Large Language Model, with the lightweight Qwen2.5-72B model achieving 94.55% accuracy while consuming only 74.3-78.4% of tokens required by Qwen3-235B and 83.0% of DeepSeek-V3's usage--delivering responses twice as fast as Qwen3-235B. This efficiency establishes TopoMAS as an accelerator for computation-driven discovery pipelines. By harmonizing rational agent orchestration with a self-evolving knowledge graph, our framework not only delivers immediate advances in topological materials but also establishes a transferable, extensible paradigm for materials-science domain.

--As large language models (LLMs) become more common in educational tools and programming environments, questions arise about how these systems should interact with users. This study investigates how different interaction styles with ChatGPT -4o (passive, proactive, and collaborative) affect user performance on simple programming tasks. I conducted a within-subjects experiment where fifteen high school students participated, completing three problems under three distinct versions of the model. Each version was designed to represent a specific style of AI support: responding only when asked, offering suggestions automatically, or engaging the user in back-and-forth dialogue.Quantitative analysis revealed that the collaborative interaction style significantly improved task completion time compared to the passive and proactive conditions. Participants also reported higher satisfaction and perceived helpfulness when working with the collaborative version. These findings suggest that the way an LLM communicates, how it guides, prompts, and responds, can meaningfully impact learning and performance.This research highlights the importance of designing LLMs that go beyond functional correctness to support more interactive, adaptive, and user-centered experiences, especially for novice programmers. Large language models (LLMs) are rapidly changing how people learn to code.

While Large Language Models (LLMs) have demonstrated impressive abilities across various domains, they still struggle with complex problems characterized by multi-objective optimization, precise constraint satisfaction, immense solution spaces, etc. To address the limitation, drawing on the superior semantic understanding ability of LLMs and also the outstanding global search and optimization capability of genetic algorithms, we propose to capitalize on their respective strengths and introduce Lyria, a general LLM-driven genetic algorithm framework, comprising 7 essential components. Through conducting extensive experiments with 4 LLMs across 3 types of problems, we demonstrated the efficacy of Lyria. Additionally, with 7 additional ablation experiments, we further systematically analyzed and elucidated the factors that affect its performance.

Effective communication is essential in cancer care, yet patients often face challenges in preparing for complex medical visits. We present an interactive, Retrieval-augmented Generation-assisted system that helps patients progress from uninformed to visit-ready. Our system adapts the Ottawa Personal Decision Guide into a dynamic retrieval-augmented generation workflow, helping users bridge knowledge gaps, clarify personal values and generate useful questions for their upcoming visits. Focusing on localized prostate cancer, we conduct a user study with patients and a clinical expert. Results show high system usability (UMUX Mean = 6.0 out of 7), strong relevance of generated content (Mean = 6.7 out of 7), minimal need for edits, and high clinical faithfulness (Mean = 6.82 out of 7). This work demonstrates the potential of combining patient-centered design with language models to enhance clinical preparation in oncology care.