Recent reinforcement learning (RL) techniques have yielded impressive reasoning improvements in language models, yet it remains unclear whether post-training truly extends a model's reasoning ability beyond what it acquires during pre-training. A central challenge is the lack of control in modern training pipelines: large-scale pre-training corpora are opaque, mid-training is often underexamined, and RL objectives interact with unknown prior knowledge in complex ways. To resolve this ambiguity, we develop a fully controlled experimental framework that isolates the causal contributions of pre-training, mid-training, and RL-based post-training. Our approach employs synthetic reasoning tasks with explicit atomic operations, parseable step-by-step reasoning traces, and systematic manipulation of training distributions. We evaluate models along two axes: extrapolative generalization to more complex compositions and contextual generalization across surface contexts. Using this framework, we reconcile competing views on RL's effectiveness. We show that: 1) RL produces true capability gains (pass@128) only when pre-training leaves sufficient headroom and when RL data target the model's edge of competence, tasks at the boundary that are difficult but not yet out of reach. 2) Contextual generalization requires minimal yet sufficient pre-training exposure, after which RL can reliably transfer. 3) Mid-training significantly enhances performance under fixed compute compared with RL only, demonstrating its central but underexplored role in training pipelines. 4) Process-level rewards reduce reward hacking and improve reasoning fidelity. Together, these results clarify the interplay between pre-training, mid-training, and RL, offering a foundation for understanding and improving reasoning LM training strategies.

Leveraging a dataset of paired narratives, we investigate the extent to which large language models (LLMs) can reliably separate incoherent and coherent stories. A probing study finds that LLMs' internal representations can reliably identify incoherent narratives. However, LLMs generate responses to rating questions that fail to satisfactorily separate the coherent and incoherent narratives across several prompt variations, hinting at a gap in LLM's understanding of storytelling. The reasoning LLMs tested do not eliminate these deficits, indicating that thought strings may not be able to fully address the discrepancy between model internal state and behavior. Additionally, we find that LLMs appear to be more sensitive to incoherence resulting from an event that violates the setting (e.g., a rainy day in the desert) than to incoherence arising from a character violating an established trait (e.g., Mary, a vegetarian, later orders a cheeseburger), suggesting that LLMs may rely more on prototypical world knowledge than building meaning-based narrative coherence. The consistent asymmetry found in our results suggests that LLMs do not have a complete grasp on narrative coherence.

Clinical ultrasound acquisition is highly operator-dependent, where rapid probe motion and brightness fluctuations often lead to reconstruction errors that reduce trust and clinical utility. We present UltrasODM, a dual-stream framework that assists sonographers during acquisition through calibrated per-frame uncertainty, saliency-based diagnostics, and actionable prompts. UltrasODM integrates (i) a contrastive ranking module that groups frames by motion similarity, (ii) an optical-flow stream fused with Dual-Mamba temporal modules for robust 6-DoF pose estimation, and (iii) a Human-in-the-Loop (HITL) layer combining Bayesian uncertainty, clinician-calibrated thresholds, and saliency maps highlighting regions of low confidence. When uncertainty exceeds the threshold, the system issues unobtrusive alerts suggesting corrective actions such as re-scanning highlighted regions or slowing the sweep. Evaluated on a clinical freehand ultrasound dataset, UltrasODM reduces drift by 15.2%, distance error by 12.1%, and Hausdorff distance by 10.1% relative to UltrasOM, while producing per-frame uncertainty and saliency outputs. By emphasizing transparency and clinician feedback, UltrasODM improves reconstruction reliability and supports safer, more trustworthy clinical workflows. Our code is publicly available at https://github.com/AnandMayank/UltrasODM.

Action recognition is also key for applications ranging from robotics to healthcare monitoring. Action information can be extracted from the body pose and movements, as well as from the background scene. However, the extent to which deep neural networks (DNNs) make use of information about the body and information about the background remains unclear. Since these two sources of information may be correlated within a training dataset, DNNs might learn to rely predominantly on one of them, without taking full advantage of the other. Unlike DNNs, humans have domain-specific brain regions selective for perceiving bodies, and regions selective for perceiving scenes. The present work tests whether humans are thus more effective at extracting information from both body and background, and whether building brain-inspired deep network architectures with separate domain-specific streams for body and scene perception endows them with more human-like performance. We first demonstrate that DNNs trained using the HAA500 dataset perform almost as accurately on versions of the stimuli that show both body and background and on versions of the stimuli from which the body was removed, but are at chance-level for versions of the stimuli from which the background was removed. Conversely, human participants (N=28) can recognize the same set of actions accurately with all three versions of the stimuli, and perform significantly better on stimuli that show only the body than on stimuli that show only the background. Finally, we implement and test a novel architecture patterned after domain specificity in the brain with separate streams to process body and background information. We show that 1) this architecture improves action recognition performance, and 2) its accuracy across different versions of the stimuli follows a pattern that matches more closely the pattern of accuracy observed in human participants.

Existing data-driven approaches in modeling and predicting time series data include ARIMA (Autoregressive Integrated Moving Average), Transformer-based models, LSTM (Long Short-Term Memory) and TCN (Temporal Convolutional Network). These approaches, and in particular deep learning-based models such as LSTM and TCN, have shown great results in predicting time series data. With the advancement of leveraging pre-trained foundation models such as Large Language Models (LLMs) and more notably Google's recent foundation model for time series data, {\it TimesFM} (Time Series Foundation Model), it is of interest to investigate whether these foundation models have the capability of outperforming existing modeling approaches in analyzing and predicting time series data. This paper investigates the performance of using LLM models for time series data prediction. We investigate the in-context learning methodology in the training of LLM models that are specific to the underlying application domain. More specifically, the paper explores training LLMs through in-context, zero-shot and few-shot learning and forecasting time series data with OpenAI {\tt o4-mini} and Gemini 2.5 Flash Lite, as well as the recent Google's Transformer-based TimesFM, a time series-specific foundation model, along with two deep learning models, namely TCN and LSTM networks. The findings indicate that TimesFM has the best overall performance with the lowest RMSE value (0.3023) and the competitive inference time (266 seconds). Furthermore, OpenAI's o4-mini also exhibits a good performance based on Zero Shot learning. These findings highlight pre-trained time series foundation models as a promising direction for real-time forecasting, enabling accurate and scalable deployment with minimal model adaptation.

Large foundation models have emerged in the last years and are pushing performance boundaries for a variety of tasks. Training or even finetuning such models demands vast datasets and computational resources, which are often scarce and costly. Adaptation methods provide a computationally efficient solution to address these limitations by allowing such models to be finetuned on small amounts of data and computing power . This is achieved by appending new trainable modules to frozen backbones with only a fraction of the trainable parameters and fitting only these modules on novel tasks. Recently, the V eRA adapter was shown to excel in parameter-efficient adaptations by utilizing a pair of frozen random low-rank matrices shared across all layers. In this paper, we propose PV eRA, a probabilistic version of the V eRA adapter, which modifies the low-rank matrices of V eRA in a probabilistic manner . This modification naturally allows handling inherent ambiguities in the input and allows for different sampling configurations during training and testing. A comprehensive evaluation was performed on the VTAB-1k benchmark and seven adapters, with PV eRA outperforming V eRA and other adapters. Our code for training models with PV eRA and benchmarking all adapters is available here.

In pre-market drug safety review, grouping related adverse event terms into standardised MedDRA queries or the FDA Office of New Drugs Custom Medical Queries (OCMQs) is critical for signal detection. We present a novel quantitative artificial intelligence system that understands and processes medical terminology and automatically retrieves relevant MedDRA Preferred Terms (PTs) for a given input query, ranking them by a relevance score using multi-criteria statistical methods. The system (SafeTerm) embeds medical query terms and MedDRA PTs in a multidimensional vector space, then applies cosine similarity and extreme-value clustering to generate a ranked list of PTs. Validation was conducted against the FDA OCMQ v3.0 (104 queries), restricted to valid MedDRA PTs. Precision, recall and F1 were computed across similarity-thresholds. High recall (>95%) is achieved at moderate thresholds. Higher thresholds improve precision (up to 86%). The optimal threshold (~0.70 - 0.75) yielded recall ~50% and precision ~33%. Narrow-term PT subsets performed similarly but required slightly higher similarity thresholds. The SafeTerm AI-driven system provides a viable supplementary method for automated MedDRA query generation. A similarity threshold of ~0.60 is recommended initially, with increased thresholds for refined term selection.

Deep learning holds immense promise for transforming medical image analysis, yet its clinical generalization remains profoundly limited. A major barrier is data heterogeneity. This is particularly true in Magnetic Resonance Imaging, where scanner hardware differences, diverse acquisition protocols, and varying sequence parameters introduce substantial domain shifts that obscure underlying biological signals. Data harmonization methods aim to reduce these instrumental and acquisition variability, but existing approaches remain insufficient. When applied to imaging data, image-based harmonization approaches are often restricted by the need for target images, while existing text-guided methods rely on simplistic labels that fail to capture complex acquisition details or are typically restricted to datasets with limited variability, failing to capture the heterogeneity of real-world clinical environments. To address these limitations, we propose DIST-CLIP (Disentangled Style Transfer with CLIP Guidance), a unified framework for MRI harmonization that flexibly uses either target images or DICOM metadata for guidance. Our framework explicitly disentangles anatomical content from image contrast, with the contrast representations being extracted using pre-trained CLIP encoders. These contrast embeddings are then integrated into the anatomical content via a novel Adaptive Style Transfer module. We trained and evaluated DIST-CLIP on diverse real-world clinical datasets, and showed significant improvements in performance when compared against state-of-the-art methods in both style translation fidelity and anatomical preservation, offering a flexible solution for style transfer and standardizing MRI data. Our code and weights will be made publicly available upon publication.

Effective motion representation is crucial for enabling robots to imitate expressive behaviors in real time, yet existing motion controllers often ignore inherent patterns in motion. Previous efforts in representation learning do not attempt to jointly capture structured periodic patterns and irregular variations in human and animal movement. To address this, we present Multi-Domain Motion Embedding (MDME), a motion representation that unifies the embedding of structured and unstructured features using a wavelet-based encoder and a probabilistic embedding in parallel. This produces a rich representation of reference motions from a minimal input set, enabling improved generalization across diverse motion styles and morphologies. We evaluate MDME on retargeting-free real-time motion imitation by conditioning robot control policies on the learned embeddings, demonstrating accurate reproduction of complex trajectories on both humanoid and quadruped platforms. Our comparative studies confirm that MDME outperforms prior approaches in reconstruction fidelity and generalizability to unseen motions. Furthermore, we demonstrate that MDME can reproduce novel motion styles in real-time through zero-shot deployment, eliminating the need for task-specific tuning or online retargeting. These results position MDME as a generalizable and structure-aware foundation for scalable real-time robot imitation.

Large Language Models (LLMs) have demonstrated remarkable performance in code intelligence tasks such as code generation, summarization, and translation. However, their reliance on linearized token sequences limits their ability to understand the structural semantics of programs. While prior studies have explored graphaugmented prompting and structure-aware pretraining, they either suffer from prompt length constraints or require task-specific architectural changes that are incompatible with large-scale instructionfollowing LLMs. To address these limitations, this paper proposes CGBridge, a novel plug-and-play method that enhances LLMs with Code Graph information through an external, trainable Bridge module. CGBridge first pre-trains a code graph encoder via selfsupervised learning on a large-scale dataset of 270K code graphs to learn structural code semantics. It then trains an external module to bridge the modality gap among code, graph, and text by aligning their semantics through cross-modal attention mechanisms. Finally, the bridge module generates structure-informed prompts, which are injected into a frozen LLM, and is fine-tuned for downstream code intelligence tasks. Experiments show that CGBridge achieves notable improvements over both the original model and the graphaugmented prompting method. Specifically, it yields a 16.19% and 9.12% relative gain in LLM-as-a-Judge on code summarization, and a 9.84% and 38.87% relative gain in Execution Accuracy on code translation. Moreover, CGBridge achieves over 4x faster inference than LoRA-tuned models, demonstrating both effectiveness and efficiency in structure-aware code understanding.