Rethinking Reinforcement fine-tuning of LLMs: A Multi-armed Bandit Learning Perspective

A large number of heuristics have been proposed to optimize the reinforcement fine-tuning of LLMs. However, inconsistent claims are made from time to time, making this area elusive. Reflecting on this situation, two fundamental questions still lack a clear understanding: 1) what is the role of each optimizing choice? 2) which ones are the bottlenecks? This paper aims to shed light on them, and it faces the challenge of several entangled confounding factors in the fine-tuning process. To tackle this challenge, we propose a bottom-up experiment pipeline. The bottom layer is composed of a minimalist configuration: one training data, one rollout per round and the reward directly serve as the learning signal without advantage function design. This minimalist configuration connects to multi-armed bandit learning with extremely large discrete action space, which offers theories to corroborate the experiment findings. The up procedure of the experiment pipeline expanding the minimalist configuration layer by layer, examining the role of each design choice. Experimental results on three LLMs and two reasoning datasets not only reveal new understanding of the design choice but also yield essential insights to shape the area.

Diagnostic Performance of Universal-Learning Ultrasound AI Across Multiple Organs and Tasks: the UUSIC25 Challenge

IMPORTANCE: Current ultrasound AI remains fragmented into single-task tools, limiting clinical utility compared to versatile modern ultrasound systems. OBJECTIVE: To evaluate the diagnostic accuracy and efficiency of single general-purpose deep learning models for multi-organ classification and segmentation. DESIGN: The Universal UltraSound Image Challenge 2025 (UUSIC25) involved developing algorithms on 11,644 images (public/private). Evaluation used an independent, multi-center test set of 2,479 images, including data from a center completely unseen during training to assess generalization. OUTCOMES: Diagnostic performance (Dice Similarity Coefficient [DSC]; Area Under the Receiver Operating Characteristic Curve [AUC]) and computational efficiency (inference time, GPU memory). RESULTS: Of 15 valid algorithms, the top model (SMART) achieved a macro-averaged DSC of 0.854 across 5 segmentation tasks and AUC of 0.766 for binary classification. Models showed high capability in segmentation (e.g., fetal head DSC: 0.942) but variability in complex tasks subject to domain shift. Notably, in breast cancer molecular subtyping, the top model's performance dropped from AUC 0.571 (internal) to 0.508 (unseen external center), highlighting generalization challenges. CONCLUSIONS: General-purpose AI models achieve high accuracy and efficiency across multiple tasks using a single architecture. However, performance degradation on unseen data suggests domain generalization is critical for future clinical deployment.

QSMnet-INR: Single-Orientation Quantitative Susceptibility Mapping via Implicit Neural Representation in k-Space

Quantitative Susceptibility Mapping (QSM) quantifies tissue magnetic susceptibility from magnetic-resonance phase data and plays a crucial role in brain microstructure imaging, iron-deposition assessment, and neurological-disease research. However, single-orientation QSM inversion remains highly ill-posed because the dipole kernel exhibits a cone-null region in the Fourier domain, leading to streaking artifacts and structural loss. To overcome this limitation, we propose QSMnet-INR, a deep, physics-informed framework that integrates an Implicit Neural Representation (INR) into the k-space domain. The INR module continuously models multi-directional dipole responses and explicitly completes the cone-null region, while a frequency-domain residual-weighted Dipole Loss enforces physical consistency. The overall network combines a 3D U-Net-based QSMnet backbone with the INR module through alternating optimization for end-to-end joint training. Experiments on the 2016 QSM Reconstruction Challenge, a multi-orientation GRE dataset, and both in-house and public single-orientation clinical data demonstrate that QSMnet-INR consistently outperforms conventional and recent deep-learning approaches across multiple quantitative metrics. The proposed framework shows notable advantages in structural recovery within cone-null regions and in artifact suppression. Ablation studies further confirm the complementary contributions of the INR module and Dipole Loss to detail preservation and physical stability. Overall, QSMnet-INR effectively alleviates the ill-posedness of single-orientation QSM without requiring multi-orientation acquisition, achieving high accuracy, robustness, and strong cross-scenario generalization-highlighting its potential for clinical translation.

SUGAR: Learning Skeleton Representation with Visual-Motion Knowledge for Action Recognition

Large Language Models (LLMs) hold rich implicit knowledge and powerful transferability. In this paper, we explore the combination of LLMs with the human skeleton to perform action classification and description. However, when treating LLM as a recognizer, two questions arise: 1) How can LLMs understand skeleton? 2) How can LLMs distinguish among actions? To address these problems, we introduce a novel paradigm named learning Skeleton representation with visUal-motion knowledGe for Action Recognition (SUGAR). In our pipeline, we first utilize off-the-shelf large-scale video models as a knowledge base to generate visual, motion information related to actions. Then, we propose to supervise skeleton learning through this prior knowledge to yield discrete representations. Finally, we use the LLM with untouched pre-training weights to understand these representations and generate the desired action targets and descriptions. Notably, we present a Temporal Query Projection (TQP) module to continuously model the skeleton signals with long sequences. Experiments on several skeleton-based action classification benchmarks demonstrate the efficacy of our SUGAR. Moreover, experiments on zero-shot scenarios show that SUGAR is more versatile than linear-based methods.

A Diffusion-Driven Temporal Super-Resolution and Spatial Consistency Enhancement Framework for 4D MRI imaging

In medical imaging, 4D MRI enables dynamic 3D visualization, yet the trade-off between spatial and temporal resolution requires prolonged scan time that can compromise temporal fidelity--especially during rapid, large-amplitude motion. Traditional approaches typically rely on registration-based interpolation to generate intermediate frames. However, these methods struggle with large deformations, resulting in misregistration, artifacts, and diminished spatial consistency. To address these challenges, we propose TSSC-Net, a novel framework that generates intermediate frames while preserving spatial consistency. To improve temporal fidelity under fast motion, our diffusion-based temporal super-resolution network generates intermediate frames using the start and end frames as key references, achieving 6x temporal super-resolution in a single inference step. Additionally, we introduce a novel tri-directional Mamba-based module that leverages long-range contextual information to effectively resolve spatial inconsistencies arising from cross-slice misalignment, thereby enhancing volumetric coherence and correcting cross-slice errors. Extensive experiments were performed on the public ACDC cardiac MRI dataset and a real-world dynamic 4D knee joint dataset. The results demonstrate that TSSC-Net can generate high-resolution dynamic MRI from fast-motion data while preserving structural fidelity and spatial consistency.

Implicit Neural Representation-Based MRI Reconstruction Method with Sensitivity Map Constraints

Magnetic Resonance Imaging (MRI) is a widely utilized diagnostic tool in clinical settings, but its application is limited by the relatively long acquisition time. As a result, fast MRI reconstruction has become a significant area of research. In recent years, Implicit Neural Representation (INR), as a scan-specific method, has demonstrated outstanding performance in fast MRI reconstruction without fully-sampled images for training. High acceleration reconstruction poses a challenging problem, and a key component in achieving high-quality reconstruction with much few data is the accurate estimation of coil sensitivity maps. However, most INR-based methods apply regularization constraints solely to the generated images, while overlooking the characteristics of the coil sensitivity maps. To handle this, this work proposes a joint coil sensitivity map and image estimation network, termed INR-CRISTAL. The proposed INR-CRISTAL introduces an extra sensitivity map regularization in the INR networks to make use of the smooth characteristics of the sensitivity maps. Experimental results show that INR-CRISTAL provides more accurate coil sensitivity estimates with fewer artifacts, and delivers superior reconstruction performance in terms of artifact removal and structure preservation. Moreover, INR-CRISTAL demonstrates stronger robustness to automatic calibration signals and the acceleration rate compared to existing methods.

MambaControl: Anatomy Graph-Enhanced Mamba ControlNet with Fourier Refinement for Diffusion-Based Disease Trajectory Prediction

Modelling disease progression in precision medicine requires capturing complex spatio-temporal dynamics while preserving anatomical integrity. Existing methods often struggle with longitudinal dependencies and structural consistency in progressive disorders. To address these limitations, we introduce MambaControl, a novel framework that integrates selective state-space modelling with diffusion processes for high-fidelity prediction of medical image trajectories. To better capture subtle structural changes over time while maintaining anatomical consistency, MambaControl combines Mamba-based long-range modelling with graph-guided anatomical control to more effectively represent anatomical correlations. Furthermore, we introduce Fourier-enhanced spectral graph representations to capture spatial coherence and multiscale detail, enabling MambaControl to achieve state-of-the-art performance in Alzheimer's disease prediction. Quantitative and regional evaluations demonstrate improved progression prediction quality and anatomical fidelity, highlighting its potential for personalised prognosis and clinical decision support.

BioVFM-21M: Benchmarking and Scaling Self-Supervised Vision Foundation Models for Biomedical Image Analysis

Scaling up model and data size have demonstrated impressive performance improvement over a wide range of tasks. Despite extensive studies on scaling behaviors for general-purpose tasks, medical images exhibit substantial differences from natural data. It remains unclear the key factors in developing medical vision foundation models at scale due to the absence of an extensive understanding of scaling behavior in the medical domain. In this paper, we explored the scaling behavior across model sizes, training algorithms, data sizes, and imaging modalities in developing scalable medical vision foundation models by self-supervised learning. To support scalable pretraining, we introduce BioVFM-21M, a large-scale biomedical image dataset encompassing a wide range of biomedical image modalities and anatomies. We observed that scaling up does provide benefits but varies across tasks. Additional analysis reveals several factors correlated with scaling benefits. Finally, we propose BioVFM, a large-scale medical vision foundation model pretrained on 21 million biomedical images, which outperforms the previous state-of-the-art foundation models across 12 medical benchmarks. Our results highlight that while scaling up is beneficial for pursuing better performance, task characteristics, data diversity, pretraining methods, and computational efficiency remain critical considerations for developing scalable medical foundation models.

PhysLLM: Harnessing Large Language Models for Cross-Modal Remote Physiological Sensing

Remote photoplethysmography (rPPG) enables non-contact physiological measurement but remains highly susceptible to illumination changes, motion artifacts, and limited temporal modeling. Large Language Models (LLMs) excel at capturing long-range dependencies, offering a potential solution but struggle with the continuous, noise-sensitive nature of rPPG signals due to their text-centric design. To bridge this gap, we introduce PhysLLM, a collaborative optimization framework that synergizes LLMs with domain-specific rPPG components. Specifically, the Text Prototype Guidance (TPG) strategy is proposed to establish cross-modal alignment by projecting hemodynamic features into LLM-interpretable semantic space, effectively bridging the representational gap between physiological signals and linguistic tokens. Besides, a novel Dual-Domain Stationary (DDS) Algorithm is proposed for resolving signal instability through adaptive time-frequency domain feature re-weighting. Finally, rPPG task-specific cues systematically inject physiological priors through physiological statistics, environmental contextual answering, and task description, leveraging cross-modal learning to integrate both visual and textual information, enabling dynamic adaptation to challenging scenarios like variable illumination and subject movements. Evaluation on four benchmark datasets, PhysLLM achieves state-of-the-art accuracy and robustness, demonstrating superior generalization across lighting variations and motion scenarios.

From Gaze to Insight: Bridging Human Visual Attention and Vision Language Model Explanation for Weakly-Supervised Medical Image Segmentation

Medical image segmentation remains challenging due to the high cost of pixel-level annotations for training. In the context of weak supervision, clinician gaze data captures regions of diagnostic interest; however, its sparsity limits its use for segmentation. In contrast, vision-language models (VLMs) provide semantic context through textual descriptions but lack the explanation precision required. Recognizing that neither source alone suffices, we propose a teacher-student framework that integrates both gaze and language supervision, leveraging their complementary strengths. Our key insight is that gaze data indicates where clinicians focus during diagnosis, while VLMs explain why those regions are significant. To implement this, the teacher model first learns from gaze points enhanced by VLM-generated descriptions of lesion morphology, establishing a foundation for guiding the student model. The teacher then directs the student through three strategies: (1) Multi-scale feature alignment to fuse visual cues with textual semantics; (2) Confidence-weighted consistency constraints to focus on reliable predictions; (3) Adaptive masking to limit error propagation in uncertain areas. Experiments on the Kvasir-SEG, NCI-ISBI, and ISIC datasets show that our method achieves Dice scores of 80.78%, 80.53%, and 84.22%, respectively-improving 3-5% over gaze baselines without increasing the annotation burden. By preserving correlations among predictions, gaze data, and lesion descriptions, our framework also maintains clinical interpretability. This work illustrates how integrating human visual attention with AI-generated semantic context can effectively overcome the limitations of individual weak supervision signals, thereby advancing the development of deployable, annotation-efficient medical AI systems. Code is available at: https://github.com/jingkunchen/FGI.git.