With the advance of Artificial Intelligence (AI), Large Language Models (LLMs) have gained prominence and been applied in diverse contexts. As they evolve into more sophisticated versions, it is essential to assess whether they reproduce biases, such as discrimination and racialization, while maintaining hegemonic discourses. Current bias detection approaches rely mostly on quantitative, automated methods, which often overlook the nuanced ways in which biases emerge in natural language. This study proposes a qualitative, discursive framework to complement such methods. Through manual analysis of LLM-generated short stories featuring Black and white women, we investigate gender and racial biases. We contend that qualitative methods such as the one proposed here are fundamental to help both developers and users identify the precise ways in which biases manifest in LLM outputs, thus enabling better conditions to mitigate them. Results show that Black women are portrayed as tied to ancestry and resistance, while white women appear in self-discovery processes. These patterns reflect how language models replicate crystalized discursive representations, reinforcing essentialization and a sense of social immobility. When prompted to correct biases, models offered superficial revisions that maintained problematic meanings, revealing limitations in fostering inclusive narratives. Our results demonstrate the ideological functioning of algorithms and have significant implications for the ethical use and development of AI. The study reinforces the need for critical, interdisciplinary approaches to AI design and deployment, addressing how LLM-generated discourses reflect and perpetuate inequalities.

-- Accurate tissue motion tracking is critical to ensure treatment outcome and safety in 2D-Cine MRIguided radiotherapy. This is typically achieved by registration of sequential images, but existing methods often face challenges with large misalignments and lack of in-terpretability. In this paper, we introduce DINOMotion, a novel deep learning framework based on DINOv2 with Low-Rank Adaptation (LoRA) layers for robust, efficient, and interpretable motion tracking. DINOMotion automatically detects corresponding landmarks to derive optimal image registration, enhancing interpretability by providing explicit visual correspondences between sequential images. The integration of LoRA layers reduces trainable parameters, improving training efficiency, while DINOv2's powerful feature representations offer robustness against large misalignments. Unlike iterative optimization-based methods, DINOMotion directly computes image registration at test time. Our experiments on volunteer and patient datasets demonstrate its effectiveness in estimating both linear and nonlinear transformations, achieving Dice scores of 92.07% DINOMotion processes each scan in approximately 30ms and consistently outperforms state-of-the-art methods, particularly in handling large misalignments. Soorena Salari and Yiming Xiao are with the Department of Computer Science and Software Engineering, Concordia University, Montreal, Canada (e-mail: soorena.salari@concordia.ca and yim-ing.xiao@concordia.ca). Hassan Rivaz is with the Department of Electrical and Computer Engineering, Concordia University, Montreal, Canada (e-mail: has-san.rivaz@concordia.ca). Catherine Spino, Laurie-Anne Pharand, and Fabienne Lath-uiliere were with the Elekta Ltd., Montreal, Canada (e-mail: catherine.spino@polymtl.ca,

Transformer models have demonstrated exceptional performance and have become indispensable in computer vision (CV) and natural language processing (NLP) tasks. However, recent studies reveal that transformers are susceptible to backdoor attacks. Prior backdoor attack methods typically rely on retraining with clean data or altering the model architecture, both of which can be resource-intensive and intrusive. In this paper, we propose Head-wise Pruning and Malicious Injection (HPMI), a novel retraining-free backdoor attack on transformers that does not alter the model's architecture. Our approach requires only a small subset of the original data and basic knowledge of the model architecture, eliminating the need for retraining the target transformer. Technically, HPMI works by pruning the least important head and injecting a pre-trained malicious head to establish the backdoor. We provide a rigorous theoretical justification demonstrating that the implanted backdoor resists detection and removal by state-of-the-art defense techniques, under reasonable assumptions. Experimental evaluations across multiple datasets further validate the effectiveness of HPMI, showing that it 1) incurs negligible clean accuracy loss, 2) achieves at least 99.55% attack success rate, and 3) bypasses four advanced defense mechanisms. Additionally, relative to state-of-the-art retraining-dependent attacks, HPMI achieves greater concealment and robustness against diverse defense strategies, while maintaining minimal impact on clean accuracy.

In-context learning (ICL) has emerged as a powerful capability of transformer-based language models, enabling them to perform tasks by conditioning on a small number of examples presented at inference time, without any parameter updates. Prior work has shown that transformers can generalize over simple function classes like linear functions, decision trees, even neural networks, purely from context, focusing on numerical or symbolic reasoning over underlying well-structured functions. Instead, we propose a novel application of ICL into the domain of cryptographic function learning, specifically focusing on ciphers such as mono-alphabetic substitution and Vigenère ciphers, two classes of private-key encryption schemes. These ciphers involve a fixed but hidden bijective mapping between plain text and cipher text characters. Given a small set of (cipher text, plain text) pairs, the goal is for the model to infer the underlying substitution and decode a new cipher text word. This setting poses a structured inference challenge, which is well-suited for evaluating the inductive biases and generalization capabilities of transformers under the ICL paradigm. Code is available at https://github.com/adistomar/CS182-project.

Bioacoustics, the study of animal sounds, offers a non-invasive method to monitor ecosystems. Extracting embeddings from audio-pretrained deep learning (DL) models without fine-tuning has become popular for obtaining bioacoustic features for tasks. However, a recent benchmark study reveals that while fine-tuned audio-pretrained VGG and transformer models achieve state-of-the-art performance in some tasks, they fail in others. This study benchmarks 11 DL models on the same tasks by reducing their learned embeddings' dimensionality and evaluating them through clustering. We found that audio-pretrained DL models 1) without fine-tuning even underperform fine-tuned AlexNet, 2) both with and without fine-tuning fail to separate the background from labeled sounds, but ResNet does, and 3) outperform other models when fewer background sounds are included during fine-tuning. This study underscores the necessity of fine-tuning audio-pretrained models and checking the embeddings after fine-tuning. Our codes are available: https://github.com/NeuroscienceAI/Audio\_Embeddings

This project explores emoji prediction from short text sequences using four deep learning architectures: a feed-forward network, CNN, transformer, and BERT. Using the TweetEval dataset, we address class imbalance through focal loss and regularization techniques. Results show BERT achieves the highest overall performance due to it's pre-training advantage, while CNN demonstrates superior efficacy on rare emoji classes. This research shows the importance of architecture selection and hyperparameter tuning for sentiment-aware emoji prediction, contributing to improved human-computer interaction.

Traditional models for pricing catastrophe (CAT) bonds struggle to capture the complex, relational data inherent in these instruments. This paper introduces CATNet, a novel framework that applies a geometric deep learning architecture, the Relational Graph Convolutional Network (R-GCN), to model the CAT bond primary market as a graph, leveraging its underlying network structure for spread prediction. Our analysis reveals that the CAT bond market exhibits the characteristics of a scale-free network, a structure dominated by a few highly connected and influential hubs. CATNet demonstrates high predictive performance, significantly outperforming a strong Random Forest benchmark. The inclusion of topological centrality measures as features provides a further, significant boost in accuracy. Interpretability analysis confirms that these network features are not mere statistical artifacts; they are quantitative proxies for long-held industry intuition regarding issuer reputation, underwriter influence, and peril concentration. This research provides evidence that network connectivity is a key determinant of price, offering a new paradigm for risk assessment and proving that graph-based models can deliver both state-of-the-art accuracy and deeper, quantifiable market insights.

Evaluation of reasoning language models gained importance after it was observed that they can combine their existing capabilities into novel traces of intermediate steps before task completion and that the traces can sometimes help them to generalize better than past models. As reasoning becomes the next scaling dimension of large language models, careful study of their capabilities in critical tasks is needed. We show that better performance is not always caused by test-time algorithmic improvements or model sizes but also by using impactful benchmarks as curricula for learning. We call this benchmark-driven selection of AI and show its effects on DeepSeek-R1 using our sequential decision-making problem from Humanity's Last Exam. Steering development of AI by impactful benchmarks trades evaluation for learning and makes novelty of test tasks key for measuring generalization capabilities of reasoning models. Consequently, some benchmarks could be seen as curricula for training rather than unseen test sets.

Recent advances in Large Reasoning Models (LRMs) have demonstrated strong performance on complex tasks through long Chain-of-Thought (CoT) reasoning. However, their lengthy outputs increase computational costs and may lead to overthinking, raising challenges in balancing reasoning effectiveness and efficiency. Current methods for efficient reasoning often compromise reasoning quality or require extensive resources. This paper investigates efficient methods to reduce the generation length of LRMs. We analyze generation path distributions and filter generated trajectories through difficulty estimation. Subsequently, we analyze the convergence behaviors of the objectives of various preference optimization methods under a Bradley-Terry loss based framework. Based on the analysis, we propose Length Controlled Preference Optimization (LCPO) that directly balances the implicit reward related to NLL loss. LCPO can effectively learn length preference with limited data and training. Extensive experiments demonstrate that our approach significantly reduces the average output length by over 50\% across multiple benchmarks while maintaining the reasoning performance. Our work highlights the potential for computationally efficient approaches in guiding LRMs toward efficient reasoning.

The current advancements in generative artificial intelligence (GenAI) models have paved the way for new possibilities for generating high-resolution synthetic images, thereby offering a promising alternative to traditional image acquisition for training computer vision models in agriculture. In the context of crop disease diagnosis, GenAI models are being used to create synthetic images of various diseases, potentially facilitating model creation and reducing the dependency on resource-intensive in-field data collection. However, limited research has been conducted on evaluating the effectiveness of integrating real with synthetic images to improve disease classification performance. Therefore, this study aims to investigate whether combining a limited number of real images with synthetic images can enhance the prediction accuracy of an EfficientNetV2-L model for classifying watermelon \textit{(Citrullus lanatus)} diseases. The training dataset was divided into five treatments: H0 (only real images), H1 (only synthetic images), H2 (1:1 real-to-synthetic), H3 (1:10 real-to-synthetic), and H4 (H3 + random images to improve variability and model generalization). All treatments were trained using a custom EfficientNetV2-L architecture with enhanced fine-tuning and transfer learning techniques. Models trained on H2, H3, and H4 treatments demonstrated high precision, recall, and F1-score metrics. Additionally, the weighted F1-score increased from 0.65 (on H0) to 1.00 (on H3-H4) signifying that the addition of a small number of real images with a considerable volume of synthetic images improved model performance and generalizability. Overall, this validates the findings that synthetic images alone cannot adequately substitute for real images; instead, both must be used in a hybrid manner to maximize model performance for crop disease classification.