Large Language Models (LLMs) can perform new tasks from in-context demonstrations, a phenomenon known as in-context learning (ICL). Recent work suggests that these demonstrations are compressed into task vectors (TVs), compact task representations that LLMs exploit for predictions. However, prior studies typically extract TVs from model outputs or hidden states using cumbersome and opaque methods, and they rarely elucidate the mechanisms by which TVs influence computation. In this work, we address both limitations. First, we propose directly training Learned Task Vectors (LTVs), which surpass extracted TVs in accuracy and exhibit superior flexibility-acting effectively at arbitrary layers, positions, and even with ICL prompts. Second, through systematic analysis, we investigate the mechanistic role of TVs, showing that at the low level they steer predictions primarily through attention-head OV circuits, with a small subset of "key heads" most decisive. At a higher level, we find that despite Transformer nonlinearities, TV propagation is largely linear: early TVs are rotated toward task-relevant subspaces to improve logits of relevant labels, while later TVs are predominantly scaled in magnitude. Taken together, LTVs not only provide a practical approach for obtaining effective TVs but also offer a principled lens into the mechanistic foundations of ICL.

Autoencoders have emerged as powerful models for visualization and dimensionality reduction based on the fundamental assumption that high-dimensional data is generated from a low-dimensional manifold. A critical challenge in autoencoder design is to preserve the geometric structure of data in the latent space, with existing approaches typically focusing on either global or local geometric properties separately. Global approaches often encounter errors in distance approximation that accumulate, while local methods frequently converge to suboptimal solutions that distort large-scale relationships. We propose Multi-Scale Geometric Autoencoder (MAE), which introduces an asymmetric architecture that simultaneously preserves both scales of the geometric structure by applying global distance constraints to the encoder and local geometric constraints to the decoder. Through theoretical analysis, we establish that this asymmetric design aligns naturally with the distinct roles of the encoder and decoder components. Our comprehensive experiments on both synthetic manifolds and real-world datasets demonstrate that MAE consistently outperforms existing methods across various evaluation metrics.

Machine unlearning in large language models (LLMs) is essential for privacy and safety; however, existing approaches remain unstable and unreliable. A widely used strategy, the gradient difference method, applies gradient descent on retained data while performing gradient ascent on forget data, the data whose influence should be removed. However, when combined with cross-entropy loss, this procedure causes unbounded growth of weights and gradients, leading to training instability and degrading both forgetting and retention. We provide a theoretical framework that explains this failure, explicitly showing how ascent on the forget set destabilizes optimization in the feedforward MLP layers of LLMs. Guided by this insight, we propose Bounded Parameter-Efficient Unlearning, a parameter-efficient approach that stabilizes LoRA-based fine-tuning by applying bounded functions to MLP adapters. This simple modification controls the weight dynamics during ascent, enabling the gradient difference method to converge reliably. Across the TOFU, TDEC, and MUSE benchmarks, and across architectures and scales from 125M to 8B parameters, our method achieves substantial improvements in forgetting while preserving retention, establishing a novel theoretically grounded and practically scalable framework for unlearning in LLMs.

Adolescent Idiopathic Scoliosis (AIS) is a complex three-dimensional spinal deformity, and accurate morphological assessment requires evaluating both coronal and sagittal alignment. While previous research has made significant progress in developing radiation-free methods for coronal plane assessment, reliable and accurate evaluation of sagittal alignment without ionizing radiation remains largely underexplored. To address this gap, we propose LatXGen, a novel generative framework that synthesizes realistic lateral spinal radiographs from posterior Red-Green-Blue and Depth (RGBD) images of unclothed backs. This enables accurate, radiation-free estimation of sagittal spinal alignment. LatXGen tackles two core challenges: (1) inferring sagittal spinal morphology changes from a lateral perspective based on posteroanterior surface geometry, and (2) performing cross-modality translation from RGBD input to the radiographic domain. The framework adopts a dual-stage architecture that progressively estimates lateral spinal structure and synthesizes corresponding radiographs. To enhance anatomical consistency, we introduce an attention-based Fast Fourier Convolution (FFC) module for integrating anatomical features from RGBD images and 3D landmarks, and a Spatial Deformation Network (SDN) to model morphological variations in the lateral view. Additionally, we construct the first large-scale paired dataset for this task, comprising 3,264 RGBD and lateral radiograph pairs. Experimental results demonstrate that LatXGen produces anatomically accurate radiographs and outperforms existing GAN-based methods in both visual fidelity and quantitative metrics. This study offers a promising, radiation-free solution for sagittal spine assessment and advances comprehensive AIS evaluation.

We investigate the mechanistic underpinnings of in-context learning (ICL) in large language models by reconciling two dominant perspectives: the component-level analysis of attention heads and the holistic decomposition of ICL into Task Recognition (TR) and Task Learning (TL). We propose a novel framework based on Task Subspace Logit Attribution (TSLA) to identify attention heads specialized in TR and TL, and demonstrate their distinct yet complementary roles. Through correlation analysis, ablation studies, and input perturbations, we show that the identified TR and TL heads independently and effectively capture the TR and TL components of ICL. Using steering experiments with geometric analysis of hidden states, we reveal that TR heads promote task recognition by aligning hidden states with the task subspace, while TL heads rotate hidden states within the subspace toward the correct label to facilitate prediction. We further show how previous findings on ICL mechanisms, including induction heads and task vectors, can be reconciled with our attention-head-level analysis of the TR-TL decomposition. Our framework thus provides a unified and interpretable account of how large language models execute ICL across diverse tasks and settings.

Large reasoning models (LRMs) exhibit unprecedented capabilities in solving complex problems through Chain-of-Thought (CoT) reasoning. However, recent studies reveal that their final answers often contradict their own reasoning traces. We hypothesize that this inconsistency stems from two competing mechanisms for generating answers: CoT reasoning and memory retrieval. To test this hypothesis, we conduct controlled experiments that challenge LRMs with misleading cues during reasoning and/or corrupted answers during retrieval. Our results across models and datasets confirm that both mechanisms operate simultaneously, with their relative dominance influenced by multiple factors: problem domains, model scales, and fine-tuning approaches (e.g., reinforcement learning vs. distillation). The findings reveal a critical limitation in current reasoning fine-tuning paradigms: models can exploit the retrieval mechanism as a shortcut, effectively "hacking" the reward signal and undermining genuine reasoning development. To address this challenge, we introduce FARL, a novel fine-tuning framework that integrates memory unlearning with reinforcement learning. By carefully suppressing retrieval shortcuts during the fine-tuning process, FARL promotes reasoning-dominant behavior and enhances generalizable reasoning capabilities.

The early and accurate classification of brain tumors is crucial for guiding effective treatment strategies and improving patient outcomes. This study presents BrainFusion, a significant advancement in brain tumor analysis using magnetic resonance imaging (MRI) by combining fine-tuned convolutional neural networks (CNNs) for tumor classification--including VGG16, ResNet50, and Xception--with YOLOv8 for precise tumor localization with bounding boxes. Leveraging the Brain Tumor MRI Dataset, our experiments reveal that the fine-tuned VGG16 model achieves test accuracy of 99.86%, substantially exceeding previous benchmarks. Beyond setting a new accuracy standard, the integration of bounding-box localization and explainable AI techniques further enhances both the clinical interpretability and trustworthiness of the system's outputs. Overall, this approach underscores the transformative potential of deep learning in delivering faster, more reliable diagnoses, ultimately contributing to improved patient care and survival rates.

Current comparisons of large reasoning models (LRMs) focus on macro-level statistics such as task accuracy or reasoning length. Whether different LRMs reason differently remains an open question. To address this gap, we introduce the LLM-proposed Open Taxonomy (LOT), a classification method that uses a generative language model to compare reasoning traces from two LRMs and articulate their distinctive features in words. LOT then models how these features predict the source LRM of a reasoning trace based on their empirical distributions across LRM outputs. Iterating this process over a dataset of reasoning traces yields a human-readable taxonomy that characterizes how models think. We apply LOT to compare the reasoning of 12 open-source LRMs on tasks in math, science, and coding. LOT identifies systematic differences in their thoughts, achieving 80-100% accuracy in distinguishing reasoning traces from LRMs that differ in scale, base model family, or objective domain. Beyond classification, LOT's natural-language taxonomy provides qualitative explanations of how LRMs think differently. Finally, in a case study, we link the reasoning differences to performance: aligning the reasoning style of smaller Qwen3 models with that of the largest Qwen3 during test time improves their accuracy on GPQA by 3.3-5.7%. Following the success of GPT -o1 and DeepSeek-R1, a wave of large reasoning models (LRMs) has recently become available. These models differ in training recipes and report varying benchmark performance, but far less is known about whether they also reason differently from one another. In this paper, we pose a fundamental question: can LRMs be distinguished by their reasoning patterns, and if so, what are the key distinguishing traits? A growing body of work has begun probing the reasoning style of individual LRMs, yielding insights into how artificial thinkers "think". Marjanovi c et al. (2025), for example, finds that DeepSeek-R1's reasoning depth correlates with the human cognitive load when processing complex sentences. Bogdan et al. (2025) annotates functions of DeepSeek-R1-Distill-Qwen's reasoning steps, showing that plan generation and re-evaluation are critical for solving math problems. However, only a few studies attempt broader comparisons across LRMs and on multiple reasoning behaviors. Gandhi et al. (2025) compares the reasoning patterns of base large language models (LLMs) and their RL fine-tuned variants, finding that the habits of base models correlate with fine-tuning gains.

Abstract--Accurate cyclone forecasting is essential for minimizing loss of life, infrastructure damage, and economic disruption. Traditional numerical weather prediction models, though effective, are computationally intensive and prone to error due to the chaotic nature of atmospheric systems. This study proposes a machine learning (ML) approach to forecasting tropical cyclone trajectory and status using time series data from the National Hurricane Center, including recently added best track wind radii. A two-stage ML pipeline is developed: a regression model first predicts cyclone features--maximum wind speed, minimum pressure, trajectory length, and directional change--using a sliding window of historical data. These outputs are then input into classification models to predict the cyclone's categorical status. Gradient boosting regression and three classifiers--random forest (RF), support vector machine (SVM), and multi-layer perceptron (MLP)--are evaluated. After hyperparameter tuning and synthetic minority oversampling (SMOTE), the RF classifier achieves the highest performance with 93% accuracy, outperforming SVM and MLP across precision, recall, and F1 score. The RF model is particularly robust in identifying minority cyclone statuses and minimizing false negatives. Regression results yield low mean absolute errors, with pressure and wind predictions within 2.2 mb and 2.4 kt, respectively. These findings demonstrate that ML models, especially ensemble-based classifiers, offer an effective, scalable alternative to traditional forecasting methods, with potential for real-time cyclone prediction and integration into decision-support systems.

Retinal disease diagnosis is critical in preventing vision loss and reducing socioeconomic burdens. Globally, over 2.2 billion people are affected by some form of vision impairment, resulting in annual productivity losses estimated at $411 billion. Traditional manual grading of retinal fundus images by ophthalmologists is time-consuming and subjective. In contrast, deep learning has revolutionized medical diagnostics by automating retinal image analysis and achieving expert-level performance. In this study, we present EYE-DEX, an automated framework for classifying 10 retinal conditions using the large-scale Retinal Disease Dataset comprising 21,577 eye fundus images. We benchmark three pre-trained Convolutional Neural Network (CNN) models--VGG16, VGG19, and ResNet50--with our finetuned VGG16 achieving a state-of-the-art global benchmark test accuracy of 92.36%. To enhance transparency and explainability, we integrate the Gradient-weighted Class Activation Mapping (Grad-CAM) technique to generate visual explanations highlighting disease-specific regions, thereby fostering clinician trust and reliability in AI-assisted diagnostics.