Wei Wang

XinYang He, Wei Wang, Bing Zhao et al.

Sparse autoencoders (SAEs) have emerged as a powerful tool for decomposing superposed language model representations into sparse and interpretable features. However, training SAEs is computationally expensive, and available open-source SAE models remain limited. In this work, we introduce \textbf{Qwen3-Instruct SAE}, a comprehensive suite of SAEs trained on the Qwen3 instruction-tuned model family, covering Qwen3-1.7B, Qwen3-4B, and Qwen3-8B. For Qwen3-1.7B and Qwen3-4B, we train layer-wise SAEs at three key activation sites: residual streams, MLP outputs, and attention outputs. For Qwen3-8B, we train SAEs on a subset of residual stream layers. We systematically evaluate these SAEs using both activation-level reconstruction metrics and model-level recovery metrics, revealing distinct sparsity--fidelity trade-offs across layers and components. Finally, we demonstrate the utility of Qwen3-Instruct SAE through a refusal-steering case study, showing that selected SAE features can causally steer instruction-tuned Qwen3 models toward refusal behavior. Our release provides a practical resource for studying sparse representations, feature-level mechanisms, and behavioral interventions in instruction-tuned language models

Feng Pan, Chunran Zheng, Bing Xue et al.

Large-scale point cloud maps are essential for robotics and spatial intelligence tasks. UAVs provide an efficient means for large-scale map acquisition; however, due to limited flight endurance and onboard storage, mapping a large-scale scene within a single flight remains difficult. Existing multi-session map merging methods can extend the mapping range, yet in UAV scenarios they still struggle to simultaneously suppress long-range drift and preserve local geometric accuracy. To address this issue, an uncertainty-aware multi-session point cloud map merging and coarse-to-fine optimization system is proposed. The proposed method first performs initial multi-session map merging based on a scene graph, and then incorporates RTK observations through an RTK spatiotemporal alignment module, where temporal offsets are estimated using Dynamic Time Warping (DTW), and continuous RTK constraints are recovered using Multi-Output Gaussian Processes (MOGP) under incomplete sampling and frame dropouts. On this basis, a unified uncertainty-aware factor graph is constructed, and local geometric accuracy is further improved through iterative plane-factor refinement. Experiments on real-world datasets validate the effectiveness and robustness of the proposed method. To facilitate further research and development in the community, our code and dataset will be publicly released.

Rui Zhang, Fuwang Dong, Wei Wang et al.

In sea-air communication networks composed of an uncrewed aerial vehicle (UAV) and an uncrewed surface vehicle (USV), the extended target characteristics and three degree of freedom motion of the USV under sea induced disturbances cause beam misalignment in the UAV's tracking of the USV. To address these issues, this paper proposes a predictive beam tracking scheme based on integrated sensing and communication (ISAC) for sea-air networks. We develop a wide and narrow beam switching scheme based on sub-array selection, where a time allocation factor is optimized to balance robust state sensing in the wide beam mode and high-rate communication in the narrow beam mode. Specifically, a wide beam mode provides full USV coverage and state sensing, while a narrow beam mode exploits the estimated state for high-gain communication with the communication receiver (CR) mounted on the USV. To characterize the CR motion, a sea-air state evolution model is derived by jointly considering the surge, sway, yaw, and sea induced disturbances of the USV. For the extended target USV, the measurement equation is constructed from multiple scatterer observations, with the measurement noise caused by sea clutter modeled, and an extended Kalman filter (EKF) based CR state prediction and estimation method is developed. In addition, the effect of sea clutter on sensing accuracy is incorporated into the time allocation optimization problem to adjust the time of the wide beam mode. Simulation results demonstrate that the proposed scheme achieves higher tracking accuracy than the state-of-the-art benchmark schemes.