Large Reasoning Models (LRMs) have demonstrated impressive performance in reasoningintensive tasks, but they remain vulnerable to harmful content generation, particularly in the mid-to-late steps of their reasoning processes.Current defense methods, however, depend on costly fine-tuning and additional expert knowledge, which limits their scalability.In this work, we propose ReasoningGuard, an inference-time safeguard for LRMs.It injects timely safety aha moments during the reasoning process to guide the model towards harmless yet helpful reasoning.Our approach leverages the internal attention mechanisms of the LRM to accurately identify key points in the reasoning path, triggering safety-oriented reflections.To safeguard both the subsequent reasoning steps and the final answers, we implement a scaling sampling strategy during decoding to select the optimal reasoning path.With minimal additional inference cost, Rea-soningGuard effectively mitigates four types of jailbreak attacks, including recent ones targeting the reasoning process of LRMs.Our approach outperforms nine existing safeguards, providing state-of-the-art defenses while avoiding common exaggerated safety issues.

Causal inference is an essential approach for understanding biological processes. Traditional causal inference methods assume a linear relationship between different biological traits, whereas their true causal relationship may be nonlinear, such as U-shaped. Moreover, when the instrument set includes weak and pleiotropic genetic instruments, accurately capturing the shape of these relationships becomes challenging. To address these issues, we propose model-averaged control function-based instrumental variable regression, a two-stage framework based on a model-averaged control function approach to estimate the marginal effect function, which represents the derivative of the causal relationship. In the first stage, a model averaging technique is employed to estimate the control function, thereby reducing weak genetic instrument bias. In the second stage, B-spline approximation is applied to estimate the marginal effect function, while SCAD penalization is used to minimize pleiotropic instrument bias. We establish the asymptotic properties of the proposed estimator and demonstrate its robust performance through simulations. Application to the Atherosclerosis Risk in Communities dataset highlights a nonlinear causal relationship between body mass index and hypertension, with the proposed method effectively estimating the specific shape and trend of the relationship.

Radiation therapy is a cornerstone of tumor treatment, and accurate prediction of the gamma passing rate (GPR) for intensity-modulated radiation therapy (IMRT) plans is clinically critical. Existing AI-based predictors often lack locational information of dose accuracy. We propose GammaDiff, a diffusion-model framework for gamma distribution prediction, with three main contributions: (1) an advanced noise-prediction network that fuses CNNs for local features with transformers for global context, achieving multi-scale modeling with efficient computation; (2) a multi-scale fusion U-Net (MFUnet) that em-beds fluence maps structure via hierarchical feature integration into the noise-prediction process; and (3) a two-stage diffusion procedure in which the forward process progressively adds noise to form training samples, and the reverse process uses the opti-mized predictor to reconstruct high-fidelity gamma distributions. Extensive experiments show that GammaDiff outperforms prior methods on PSNR and SSIM, with notably higher sensitivity to failure cases, providing a more robust, reliable AI solution for plan-quality verification in radiation therapy.

This paper presents a novel framework for fine-grained martial arts gesture recognition that integrates attention mechanisms to enhance model accuracy.The proposed approach introduces several key innovations to improve the recognition of subtle variations in martial arts movements. First, a multiscale attention mechanism is employed, allowing the model to dynamically focus on both fine-grained body parts and global features, capturing the intricate relationships between different scales of motion. Second, structured attention maps are introduced to help the model better understand the spatial relationships between body parts, further improving recognition accuracy. Additionally, the framework enhances local feature generation, where attention mechanisms refine the model’s focus on critical areas, while attention regularization prevents overfitting by reducing excessive attention to certain body parts, boosting the model’s generalization capabilities. This novel combination of attention mechanisms results in a highly effective and accurate system for martial arts gesture recognition, achieving superior performance compared to traditional methods.

Integrated order batching and picker routing (IOBPR) is a complex combinatorial optimization problem in real-world intelligent manufacturing systems. Heuristics are often used for solving such complex scheduling problems. Manually designing scheduling heuristics suffer from two limitations: 1) few problem features can be taken into account and 2) the design process is time consuming. Genetic programming hyper heuristic (GPHH) approaches have been proposed on many scheduling problems to automatically evolve effective heuristics. However, existing GPHH approaches are often problem specific and requires careful design of problem specific terminal sets and evolution operators. The aim of this work is to develop a GPHH approach to evolve heuristics for the IOBPR problem. In particular, we propose a novel terminal set (NT) with three types of terminals, and a GPHH with elitist mutation (GPHH-EM) algorithm. Extensive experiments demonstrate that the heuristics evolved by GPHH-EM can significantly outperform other state-of-the-art competing algorithms designed by human experts. Further analysis indicates that the three types of terminals effectively complement to improve evolved heuristics for decision making. Furthermore, the newly developed elitist mutation operator expedites the evolutionary process for GPHH to find high-quality heuristics.

Capturing stationary aerial objects on unmanned surface vehicles (USVs) is challenging due to quasiperiodic and fast floating-base motions caused by wave-induced disturbances. It is hard to maintain high motion prediction accuracy due to the stochastic nature of these disturbances, and perform object capture through real-time tracking due to the limited active torque. We introduce confidence analysis in predictive capture. To address the inaccuracy predictions, we calculate a real-time confidence tube to evaluate the prediction quality. To overcome tracking difficulties, we plan a trajectory to capture the object at a future moment while maximizing the confidence of the capture position on the predicted trajectory. All calculations are completed within 0.2 s to ensure a timely response. We validate our approach through experiments, where we simulate disturbances by executing real USV motions using a servo platform. The results demonstrate that our method achieves an 80% success rate.

With the rapid growth of Massive Open Online Courses (MOOCs), it is expensive and time-consuming to extract high-quality knowledgeable concepts taught in the course by human effort to help learners grasp the essence of the course.In this paper, we propose to automatically extract course concepts using distant supervision to eliminate the heavy work of human annotations, which generates labels by matching them with an easily accessed dictionary.However, this matching process suffers from severe noisy and incomplete annotations because of the limited dictionary and diverse MOOCs.To tackle these challenges, we present a novel three-stage framework DS-MOCE, which leverages the power of pretrained language models explicitly and implicitly and employs discipline-embedding models with a self-train strategy based on label generation refinement across different domains.We also provide an expert-labeled dataset spanning 20 academic disciplines.Experimental results demonstrate the superiority of DS-MOCE over the state-of-the-art distantly supervised methods (with 7% absolute F1 score improvement).Code and data are now available at https: //github.com/THU-KEG/MOOC-NER.

The prosperity of massive open online courses provides fodder for plentiful research efforts on adaptive learning. However, current open-access educational datasets are still far from sufficient to meet the need for various topics of adaptive learning. Existing released datasets often cover only small-scale data, lack fine-grained knowledge concepts. They are even difficult to curate and supplement due to platform limitations. In this work, we construct MOOCCubeX, a large, knowledge-centered repository consisting of 4,216 courses, 230,263 videos, 358,265 exercises, 637,572 fine-grained concepts and over 296 million behavioral data of 3,330,294 students, for supporting the research topics on adaptive learning in MOOCs. Licensed by XuetangX, one of the largest MOOC websites in China, we obtain abundant and diverse course resources and student behavioral data and are permitted to make subsequent periodic updates. We propose a framework to accomplish data processing, weakly supervised fine-grained concept graph mining, and data curation to improve usability and richness. Based on the fine-grained concepts, we re-organize the data from the knowledge perspective and acquire more external learning resources from the web. Our repository is now available at https://github.com/THU-KEG/MOOCCubeX.

Jifan Yu, Gan Luo, Tong Xiao, Qingyang Zhong, Yuquan Wang, Wenzheng Feng, Junyi Luo, Chenyu Wang, Lei Hou, Juanzi Li, Zhiyuan Liu, Jie Tang. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. 2020.

Accurate kinematic models and measurements are needed in many robotic applications. However uncertainties related to joint angle measurements and manipulator geometry are unavoidable, especially when grasping and using different tools or when we do not have access to an accurate robot model, e.g. when we construct a robotic system by hand. The generalized inverse kinematics methods are not applicable when a manipulator stay inside its singular region. We derive the upper bounds on the joint measurement errors and geometric uncertainties, in order to guarantee that the open-chain serial manipulators stay outside the singular region. These bounds in other words enable en effective execution of generalized inverse kinematics methods for a robotic system which is prone to geometric uncertainties. In addition to the analytic derivation, We validate the proposed bounds through a trajectory tracing task performed by a PR2 robot simulator.

Teleoperation of a mobile robot manipulating and exploring an object shares many similarities with the manipulation of virtual objects in a 3D design software such as AutoCAD. The user interfaces are however quite different, mainly for historical reasons. In this paper we aim to change that, and draw inspiration from the 3D design community to propose a teleoperation interface control mode that is identical to the ones being used to locally navigate the virtual viewpoint of most Computer Aided Design (CAD) softwares. The proposed mobile manipulator control framework thus allows the user to focus on the 3D objects being manipulated, using control modes such as orbit object and pan object, supported by data from the wrist mounted RGB-D sensor. The gripper of the robot performs the desired motions relative to the object, while the manipulator arm and base moves in a way that realizes the desired gripper motions. The system redundancies are exploited in order to take additional constraints, such as obstacle avoidance, into account, using a constraint based programming framework.

Constraint based programming provides a versatile framework for combining several different constraints into a single robot control scheme. We take advantage of the redundancy of a robot manipulator to improve the execution of a reactive tracking task, in terms of a task-dependent measure which is a weighted sum of velocity transmissions along the current directions of motion. With inspiration from recent work, we provide analytical gradients and computable weights of the task-dependent measure, which enable us to include it in a reactive constraint based programming framework, without relying on inexact numerical approximations and manually tuning weights. The proposed approach is illustrated in a set of simulations, comparing the performance with a standard constraint based programming method.

In the future mobile dual-arm robots are expected to perform many tasks. Kinematically, the configuration of two manipulators that branch from the same common mobile base results in a serial-to-parallel kinematic structure, which makes inverse kinematic computations non-trivial. The motion of the base has to be decided in a trade-off, taking the needs of both arms into account. We propose to use a Virtual Kinematic Chain (VKC) to specify the common motion of the parallel manipulators, instead of using the two manipulators kinematics directly. With this VKC, we formulate a constraint based programming solution for the robot to respond to external disturbances during task execution. The proposed approach is experimentally verified both in a noise-free illustrative simulation and a real human robot co-manipulation task.

The use of redundant sensors brings a rich diversity of information, nevertheless fusing different sensors that run at vastly different frequencies into a proper estimate is still a challenging sensor fusion problem. Instead of using the size-varying measurements and thereby the size-varying filters during each sampling period, we propose to find a substitute of the unavailable low frequency measurements such that we can avoid using different sampling frequencies in one filter. In the gap between the sampling of two low frequency measurements, the use of these substitutes produces smoother estimates. In both the proof of concept simulation and the localization experiment performed on an indoor soccer robot, our proposed approach exhibits an improved performance compared to the size-varying Kalman filter methods.