As large language models (LLMs) become more integrated into societal systems, the risk of them perpetuating and amplifying harmful biases becomes a critical safety concern. Traditional methods for mitigating bias often rely on data filtering or post-hoc output moderation, which treat the model as an opaque black box. In this work, we introduce a complete, end-to-end system that uses techniques from mechanistic interpretability to both identify and actively mitigate bias directly within a model's internal workings. Our method involves two primary stages. First, we train linear "probes" on the internal activations of a model to detect the latent representations of various biases (e.g., gender, race, age). Our experiments on \texttt{gpt2-large} demonstrate that these probes can identify biased content with near-perfect accuracy, revealing that bias representations become most salient in the model's later layers. Second, we leverage these findings to compute "steering vectors" by contrasting the model's activation patterns for biased and neutral statements. By adding these vectors during inference, we can actively steer the model's generative process away from producing harmful, stereotypical, or biased content in real-time. We demonstrate the efficacy of this activation steering technique, showing that it successfully alters biased completions toward more neutral alternatives. We present our work as a robust and reproducible system that offers a more direct and interpretable approach to building safer and more accountable LLMs.

Designing mechanical mechanisms to trace specific paths is a classic yet notoriously difficult engineering problem, characterized by a vast and complex search space of discrete topologies and continuous parameters. We introduce MechaFormer, a Transformer-based model that tackles this challenge by treating mechanism design as a conditional sequence generation task. Our model learns to translate a target curve into a domain-specific language (DSL) string, simultaneously determining the mechanism's topology and geometric parameters in a single, unified process. MechaFormer significantly outperforms existing baselines, achieving state-of-the-art path-matching accuracy and generating a wide diversity of novel and valid designs. We demonstrate a suite of sampling strategies that can dramatically improve solution quality and offer designers valuable flexibility. Furthermore, we show that the high-quality outputs from MechaFormer serve as excellent starting points for traditional optimizers, creating a hybrid approach that finds superior solutions with remarkable efficiency.

Natural disasters such as hurricanes and wildfires increasingly introduce unusual disturbance on economic activities, which are especially likely to reshape commercial land use pattern given their sensitive to customer visitation. However, current modeling approaches are limited in capturing such complex interplay between human activities and commercial land use change under and following disturbances. Such interactions have been more effectively captured in current resilient urban planning theories. This study designs and calibrates a Urban Theory-Informed Spatio-Temporal Attention Model for Predicting Post-Disaster Commercial Land Use Change (Urban-STA4CLC) to predict both the yearly decline and expansion of commercial land use at census block level under cumulative impact of disasters on human activities over two years. Guided by urban theories, Urban-STA4CLC integrates both spatial and temporal attention mechanisms with three theory-informed modules. Resilience theory guides a disaster-aware temporal attention module that captures visitation dynamics. Spatial economic theory informs a multi-relational spatial attention module for inter-block representation. Diffusion theory contributes a regularization term that constrains land use transitions. The model performs significantly better than non-theoretical baselines in predicting commercial land use change under the scenario of recurrent hurricanes, with around 19% improvement in F1 score (0.8763). The effectiveness of the theory-guided modules was further validated through ablation studies. The research demonstrates that embedding urban theory into commercial land use modeling models may substantially enhance the capacity to capture its gains and losses. These advances in commercial land use modeling contribute to land use research that accounts for cumulative impacts of recurrent disasters and shifts in economic activity patterns.

Transformer - based models have made significant progress in time series forecasting. However, a key limitation of deep learning models is their susceptibility to adversarial attacks, which has not been studied enough in the context of time series prediction . In contrast to areas such as computer vision, where adversarial robustness has been extensively studied, frequency domain features of time series data play an important role in the prediction task but have not been sufficiently explored in terms of adver sarial attacks. This paper proposes a time series prediction attack algorithm based on frequency domain loss. Specifically, we adapt an attack method originally designed for classification tasks to the prediction field and optimize the adversarial samples using both time - domain and frequency - domain losses. To the best of our knowledge, there is no relevant research on using frequency information for time - series adversarial attacks. Our experimental results show that these current time series prediction mode ls are vulnerable to adversarial attacks, and our approach achieves excellent performance on major time series forecasting datasets.

Despite their impressive performances, Large Language Models (LLMs) remain prone to hallucination, which critically undermines their trustworthiness. While most of the previous work focused on tackling answer and attribution correctness, a recent line of work investigated faithfulness, with a focus on leveraging internal model signals to reflect a model's actual decision-making process while generating the answer. Nevertheless, these methods induce additional latency and have shown limitations in directly aligning token generation with attribution generation. In this paper, we introduce LoDIT, a method that jointly generates and faithfully attributes answers in RAG by leveraging specific token logits during generation. It consists of two steps: (1) marking the documents with specific token identifiers and then leveraging the logits of these tokens to estimate the contribution of each document to the answer during generation, and (2) aggregating these contributions into document attributions. Experiments on a trustworthiness-focused attributed text-generation benchmark, Trust-Align, show that LoDIT significantly outperforms state-of-the-art models on several metrics. Finally, an in-depth analysis of LoDIT shows both its efficiency in terms of latency and its robustness in different settings.

Large Vision-Language Models face growing safety challenges with multimodal inputs. This paper introduces the concept of Implicit Reasoning Safety, a vulnerability in LVLMs. Benign combined inputs trigger unsafe LVLM outputs due to flawed or hidden reasoning. To showcase this, we developed Safe Semantics, Unsafe Interpretations, the first dataset for this critical issue. Our demonstrations show that even simple In-Context Learning with SSUI significantly mitigates these implicit multimodal threats, underscoring the urgent need to improve cross-modal implicit reasoning.

--This letter introduces EGGCodec, a robust neural Encodec framework engineered for electroglottography (EGG) signal reconstruction and F0 extraction. We propose a multi-scale frequency-domain loss function to capture the nuanced relationship between original and reconstructed EGG signals, complemented by a time-domain correlation loss to improve generalization and accuracy. Unlike conventional Encodec models that extract F0 directly from features, EGGCodec leverages reconstructed EGG signals, which more closely correspond to F0. By removing the conventional GAN discriminator, we streamline EGGCodec's training process without compromising efficiency, incurring only negligible performance degradation. Trained on a widely used EGG-inclusive dataset, extensive evaluations demonstrate that EGGCodec outperforms state-of-the-art F0 extraction schemes, reducing mean absolute error (MAE) from 14.14 Hz to 13.69 Hz, and improving voicing decision error (VDE) by 38.2%.

Ultrasound (US) imaging is increasingly used in spinal procedures due to its real-time, radiation-free capabilities; however, its effectiveness is hindered by shadowing artifacts that obscure deeper tissue structures. Traditional approaches, such as CT-to-US registration, incorporate anatomical information from preoperative CT scans to guide interventions, but they are limited by complex registration requirements, differences in spine curvature, and the need for recent CT imaging. Recent shape completion methods can offer an alternative by reconstructing spinal structures in US data, while being pretrained on large set of publicly available CT scans. However, these approaches are typically offline and have limited reproducibility. In this work, we introduce a novel integrated system that combines robotic ultrasound with real-time shape completion to enhance spinal visualization. Our robotic platform autonomously acquires US sweeps of the lumbar spine, extracts vertebral surfaces from ultrasound, and reconstructs the complete anatomy using a deep learning-based shape completion network. This framework provides interactive, real-time visualization with the capability to autonomously repeat scans and can enable navigation to target locations. This can contribute to better consistency, reproducibility, and understanding of the underlying anatomy. We validate our approach through quantitative experiments assessing shape completion accuracy and evaluations of multiple spine acquisition protocols on a phantom setup. Additionally, we present qualitative results of the visualization on a volunteer scan.

Deep learning-based time series forecasting has found widespread applications. Recently, converting time series data into the frequency domain for forecasting has become popular for accurately exploring periodic patterns. However, existing methods often cannot effectively explore stationary information from complex intertwined frequency components. In this paper, we propose a simple yet effective Amplitude-Phase Reconstruct Network (APRNet) that models the inter-relationships of amplitude and phase, which prevents the amplitude and phase from being constrained by different physical quantities, thereby decoupling the distinct characteristics of signals for capturing stationary information. Specifically, we represent the multivariate time series input across sequence and channel dimensions, highlighting the correlation between amplitude and phase at multiple interaction frequencies. We propose a novel Kolmogorov-Arnold-Network-based Local Correlation (KLC) module to adaptively fit local functions using univariate functions, enabling more flexible characterization of stationary features across different amplitudes and phases. This significantly enhances the model's capability to capture time-varying patterns. Extensive experiments demonstrate the superiority of our APRNet against the state-of-the-arts (SOTAs).

Magnetic resonance (MR) imaging is essential for evaluating central nervous system (CNS) tumors, guiding surgical planning, treatment decisions, and assessing postoperative outcomes and complication risks. While recent work has advanced automated tumor segmentation and report generation, most efforts have focused on preoperative data, with limited attention to postoperative imaging analysis. This study introduces a comprehensive pipeline for standardized postsurtical reporting in CNS tumors. Using the Attention U-Net architecture, segmentation models were trained for the preoperative (non-enhancing) tumor core, postoperative contrast-enhancing residual tumor, and resection cavity. Additionally, MR sequence classification and tumor type identification for contrast-enhancing lesions were explored using the DenseNet architecture. The models were integrated into a reporting pipeline, following the RANO 2.0 guidelines. Training was conducted on multicentric datasets comprising 2000 to 7000 patients, using a 5-fold cross-validation. Evaluation included patient-, voxel-, and object-wise metrics, with benchmarking against the latest BraTS challenge results. The segmentation models achieved average voxel-wise Dice scores of 87%, 66%, 70%, and 77% for the tumor core, non-enhancing tumor core, contrast-enhancing residual tumor, and resection cavity, respectively. Classification models reached 99.5% balanced accuracy in MR sequence classification and 80% in tumor type classification. The pipeline presented in this study enables robust, automated segmentation, MR sequence classification, and standardized report generation aligned with RANO 2.0 guidelines, enhancing postoperative evaluation and clinical decision-making. The proposed models and methods were integrated into Raidionics, open-source software platform for CNS tumor analysis, now including a dedicated module for postsurgical analysis.