In this paper, we present our fact-checking pipeline which has scored first in FEVER 8 shared task. Our fact-checking system is a simple two-step RAG pipeline based on our last year's submission. We show how the pipeline can be redeployed on-premise, achieving state-of-the-art fact-checking performance (in sense of Ev2R test-score), even under the constraint of a single NVidia A10 GPU, 23GB of graphical memory and 60s running time per claim.

Graphical user interface visual grounding (GUI-VG), a core capability for GUI agents, has primarily relied on supervised fine-tuning (SFT) of multimodal large language models (MLLMs), which demands extensive data curation and significant training costs. However, as MLLMs continue to advance and even cover GUI domains during pretraining, the necessity of exhaustive SFT post-training becomes increasingly questionable. Meanwhile, recent successes of rule-based reinforcement fine-tuning (RFT) suggest a more efficient alternative. Despite this promise, the optimal manner of applying RFT for GUI-VG remains unexplored. To bridge this gap, we introduce GuirlVG, a reinforcement learning-based GUI-VG method built on a systematic empirical study and a novel stabilization technique. We find that naive application of RFT underperforms the SFT baseline, motivating a deeper exploration. First, we decompose RFT into its core components and analyze the optimal formulation of each. Second, we propose a novel Adversarial KL Factor that dynamically stabilizes training to mitigate reward over-optimization. Third, we further explore the training configurations of RFT to enhance effectiveness. Extensive experiments show that GuirlVG, with only 5.2K training samples, outperforms SFT methods trained on over 10M samples, achieving a 7.7% improvement on ScreenSpot, a 17.2% improvement on ScreenSpotPro, and 91.9% accuracy on ScreenSpotV2.

This paper introduces the Learned User Significance Tracker (LUST), a framework designed to analyze video content and quantify the thematic relevance of its segments in relation to a user-provided textual description of significance. LUST leverages a multi-modal analytical pipeline, integrating visual cues from video frames with textual information extracted via Automatic Speech Recognition (ASR) from the audio track. The core innovation lies in a hierarchical, two-stage relevance scoring mechanism employing Large Language Models (LLMs). An initial "direct relevance" score, $S_{d,i}$, assesses individual segments based on immediate visual and auditory content against the theme. This is followed by a "contextual relevance" score, $S_{c,i}$, that refines the assessment by incorporating the temporal progression of preceding thematic scores, allowing the model to understand evolving narratives. The LUST framework aims to provide a nuanced, temporally-aware measure of user-defined significance, outputting an annotated video with visualized relevance scores and comprehensive analytical logs.

Previous work has demonstrated that AI methods for analysing scientific literature benefit significantly from annotating sentences in papers according to their rhetorical roles, such as research gaps, results, limitations, extensions of existing methodologies, and others. Such representations also have the potential to support the development of a new generation of systems capable of producing high-quality literature reviews. However, achieving this goal requires the definition of a relevant annotation schema and effective strategies for large-scale annotation of the literature. This paper addresses these challenges by 1) introducing a novel annotation schema specifically designed to support literature review generation and 2) conducting a comprehensive evaluation of a wide range of state-of-the-art large language models (LLMs) in classifying rhetorical roles according to this schema. To this end, we also present Sci-Sentence, a novel multidisciplinary benchmark comprising 700 sentences manually annotated by domain experts and 2,240 sentences automatically labelled using LLMs. We evaluate 37 LLMs on this benchmark, spanning diverse model families and sizes, using both zero-shot learning and fine-tuning approaches. The experiments yield several novel insights that advance the state of the art in this challenging domain. First, the current generation of LLMs performs remarkably well on this task when fine-tuned on high-quality data, achieving performance levels above 96\% F1. Second, while large proprietary models like GPT-4o achieve the best results, some lightweight open-source alternatives also demonstrate excellent performance. Finally, enriching the training data with semi-synthetic examples generated by LLMs proves beneficial, enabling small encoders to achieve robust results and significantly enhancing the performance of several open decoder models.

Large language models (LLMs) show significant potential in healthcare, prompting numerous benchmarks to evaluate their capabilities. However, concerns persist regarding the reliability of these benchmarks, which often lack clinical fidelity, robust data management, and safety-oriented evaluation metrics. To address these shortcomings, we introduce MedCheck, the first lifecycle-oriented assessment framework specifically designed for medical benchmarks. Our framework deconstructs a benchmark's development into five continuous stages, from design to governance, and provides a comprehensive checklist of 46 medically-tailored criteria. Using MedCheck, we conducted an in-depth empirical evaluation of 53 medical LLM benchmarks. Our analysis uncovers widespread, systemic issues, including a profound disconnect from clinical practice, a crisis of data integrity due to unmitigated contamination risks, and a systematic neglect of safety-critical evaluation dimensions like model robustness and uncertainty awareness. Based on these findings, MedCheck serves as both a diagnostic tool for existing benchmarks and an actionable guideline to foster a more standardized, reliable, and transparent approach to evaluating AI in healthcare.

Large language models (LLMs) are now being used with increasing frequency as chat bots, tasked with the summarizing information or generating text and code in accordance with user instructions. The rapid increase in reasoning capabilities and inference speed of LLMs has revealed their remarkable potential for applications extending beyond the domain of chat bots to general machine learning tasks. This work is conducted out of the curiosity about such potential. In this work, we propose a framework Mockingbird to adapt LLMs to general machine learning tasks and evaluate its performance and scalability on several general machine learning tasks. The core concept of this framework is instructing LLMs to role-play functions and reflect on its mistakes to improve itself. Our evaluation and analysis result shows that LLM-driven machine learning methods, such as Mockingbird, can achieve acceptable results on common machine learning tasks; however, solely reflecting on its own currently cannot outperform the effect of domain-specific documents and feedback from human experts.

BalancedBio, a theoretically-grounded framework for parameter-efficient biomedical reasoning that addresses the fundamental challenge of multi-capability integration in domain-specific AI alignment. We establish the Biomedical Multi-Capability Convergence Theorem, proving that balanced development of domain expertise, reasoning, and instruction-following requires orthogonal gradient spaces to prevent capability interference--a critical requirement for safe biomedical AI deployment. Our approach introduces two key innovations: (1) Medical Knowledge-Grounded Synthetic Generation (MKGSG), which extends Source2Synth by incorporating clinical workflow constraints and medical ontology validation to ensure both factual accuracy and clinical safety; and (2) Capability-A ware Group Relative Policy Optimization, where we theoretically derive optimal hybrid reward weighting strategies that maintain capability orthogonality during reinforcement learning, incorporating a reward model that scores business data adapted to biomedical downstream tasks, achieving true multi-dimensional hybrid RL with both rule-based and model-based scores . Through rigorous mathematical analysis, we prove that our training objective achieves Pareto-optimal convergence where improvements in one capability domain preserve performance in others--addressing a fundamental alignment challenge in medical AI. BalancedBio demonstrates state-of-the-art performance within its parameter class: domain expertise (80.95% BIOMED-MMLU, +15.32% over best baseline), reasoning capabilities (61.94%, +7.75%), instruction-following (67.95%, +6.44%), and integration score (86.7%, +18.5%). Critically, we provide theoretical safety guarantees with formal bounds on capability preservation and clinical accuracy maintenance. Real-world deployment across healthcare institutions validates practical impact: 78% cost reduction, 23% improved diagnostic accuracy, and 89% clinician acceptance rate. Our work establishes a principled methodology for biomedical AI alignment, demonstrating that sophisticated reasoning capabilities can be achieved efficiently while maintaining safety and reliability constraints essential for medical applications. We will release the 0.5B V ersion of our model.

Key-Value (KV) cache quantization has become a widely adopted optimization technique for efficient large language models (LLMs) inference by reducing KV cache memory usage and mitigating memory-bound constraints. Recent studies have emphasized the importance of preserving the original precision of KVs for the first few tokens to ensure the protection of attention sinks. While this approach has proven effective in mitigating performance degradation, its underlying principles remain insufficiently understood. Moreover, it fails to address the recent discovery that attention sinks can emerge beyond the initial token positions. In this work, we elucidate the underlying mechanisms of attention sinks during inference by examining their role in the cross-layer evolution of extreme activation outliers. Additionally, we provide a comprehensive analysis of the interplay between attention sinks and KV cache quantization. Based on our enhanced understanding, we introduce \textit{\textbf{KVSink}}, a plug-and-play method that effectively predicts sink tokens with negligible overhead, enabling more thorough preservation. Extensive experiments demonstrate that KVSink outperforms the existing Preserve-First-N (PFN) strategy, offering more effective preservation of attention sinks during KV cache quantization. Moreover, when applied to the well-established KVQuant method, KVSink further improves perplexity (PPL) and reduces reliance on 16-bit numerical outliers.

The increasing demand for mental health services has outpaced the availability of real training data to develop clinical professionals, leading to limited support for the diagnosis of depression. This shortage has motivated the development of simulated or virtual patients to assist in training and evaluation, but existing approaches often fail to generate clinically valid, natural, and diverse symptom presentations. In this work, we embrace the recent advanced language models as the backbone and propose a novel clinician-in-the-loop patient simulation pipeline, TalkDep, with access to diversified patient profiles to develop simulated patients. By conditioning the model on psychiatric diagnostic criteria, symptom severity scales, and contextual factors, our goal is to create authentic patient responses that can better support diagnostic model training and evaluation. We verify the reliability of these simulated patients with thorough assessments conducted by clinical professionals. The availability of validated simulated patients offers a scalable and adaptable resource for improving the robustness and generalisability of automatic depression diagnosis systems.

Hierarchical Text Classification (HTC) aims to assign texts to structured label hierarchies; however, it faces challenges due to data scarcity and model complexity. This study explores the feasibility of using black box Large Language Models (LLMs) accessed via APIs for HTC, as an alternative to traditional machine learning methods that require extensive labeled data and computational resources. We evaluate three prompting strategies -- Direct Leaf Label Prediction (DL), Direct Hierarchical Label Prediction (DH), and Top-down Multi-step Hierarchical Label Prediction (TMH) -- in both zero-shot and few-shot settings, comparing the accuracy and cost-effectiveness of these strategies. Experiments on two datasets show that a few-shot setting consistently improves classification accuracy compared to a zero-shot setting. While a traditional machine learning model achieves high accuracy on a dataset with a shallow hierarchy, LLMs, especially DH strategy, tend to outperform the machine learning model on a dataset with a deeper hierarchy. API costs increase significantly due to the higher input tokens required for deeper label hierarchies on DH strategy. These results emphasize the trade-off between accuracy improvement and the computational cost of prompt strategy. These findings highlight the potential of black box LLMs for HTC while underscoring the need to carefully select a prompt strategy to balance performance and cost.