Weikun Zhang

Zichao Dong, Weikun Zhang, Xufeng Huang et al.

Human robot interaction is an exciting task, which aimed to guide robots following instructions from human. Since huge gap lies between human natural language and machine codes, end to end human robot interaction models is fair challenging. Further, visual information receiving from sensors of robot is also a hard language for robot to perceive. In this work, HuBo-VLM is proposed to tackle perception tasks associated with human robot interaction including object detection and visual grounding by a unified transformer based vision language model. Extensive experiments on the Talk2Car benchmark demonstrate the effectiveness of our approach. Code would be publicly available in https://github.com/dzcgaara/HuBo-VLM.

Xiaoxiang Han, Yang Chen, Qiaohong Liu et al.

Cardiac cine magnetic resonance imaging (MRI) is one of the important means to assess cardiac functions and vascular abnormalities. Mitigating artifacts arising during image reconstruction and accelerating cardiac cine MRI acquisition to obtain high-quality images is important. A novel end-to-end deep learning network is developed to improve cardiac cine MRI reconstruction. First, a U-Net is adopted to obtain the initial reconstructed images in k-space. Further to remove the motion artifacts, the motion-guided deformable alignment (MGDA) module with second-order bidirectional propagation is introduced to align the adjacent cine MRI frames by maximizing spatial-temporal information to alleviate motion artifacts. Finally, the multi-resolution fusion (MRF) module is designed to correct the blur and artifacts generated from alignment operation and obtain the last high-quality reconstructed cardiac images. At an 8$\times$ acceleration rate, the numerical measurements on the ACDC dataset are structural similarity index (SSIM) of 78.40%$\pm$.57%, peak signal-to-noise ratio (PSNR) of 30.46$\pm$1.22dB, and normalized mean squared error (NMSE) of 0.0468$\pm$0.0075. On the ACMRI dataset, the results are SSIM of 87.65%$\pm$4.20%, PSNR of 30.04$\pm$1.18dB, and NMSE of 0.0473$\pm$0.0072. The proposed method exhibits high-quality results with richer details and fewer artifacts for cardiac cine MRI reconstruction on different accelerations.

Zichao Dong, Hang Ji, Weikun Zhang et al.

Occupancy prediction tasks focus on the inference of both geometry and semantic labels for each voxel, which is an important perception mission. However, it is still a semantic segmentation task without distinguishing various instances. Further, although some existing works, such as Open-Vocabulary Occupancy (OVO), have already solved the problem of open vocabulary detection, visual grounding in occupancy has not been solved to the best of our knowledge. To tackle the above two limitations, this paper proposes Occupancy Grounding (OG), a novel method that equips vanilla occupancy instance segmentation ability and could operate visual grounding in a voxel manner with the help of grounded-SAM. Keys to our approach are (1) affinity field prediction for instance clustering and (2) association strategy for aligning 2D instance masks and 3D occupancy instances. Extensive experiments have been conducted whose visualization results and analysis are shown below. Our code will be publicly released soon.

Zichao Dong, Hang Ji, Xufeng Huang et al.

Point encoder is of vital importance for point cloud recognition. As the very beginning step of whole model pipeline, adding features from diverse sources and providing stronger feature encoding mechanism would provide better input for downstream modules. In our work, we proposed a novel PeP module to tackle above issue. PeP contains two main parts, a refined point painting method and a LM-based point encoder. Experiments results on the nuScenes and KITTI datasets validate the superior performance of our PeP. The advantages leads to strong performance on both semantic segmentation and object detection, in both lidar and multi-modal settings. Notably, our PeP module is model agnostic and plug-and-play. Our code will be publicly available soon.

Zhiyu Tan, Zichao Dong, Cheng Zhang et al.

Semantic occupancy prediction aims to infer dense geometry and semantics of surroundings for an autonomous agent to operate safely in the 3D environment. Existing occupancy prediction methods are almost entirely trained on human-annotated volumetric data. Although of high quality, the generation of such 3D annotations is laborious and costly, restricting them to a few specific object categories in the training dataset. To address this limitation, this paper proposes Open Vocabulary Occupancy (OVO), a novel approach that allows semantic occupancy prediction of arbitrary classes but without the need for 3D annotations during training. Keys to our approach are (1) knowledge distillation from a pre-trained 2D open-vocabulary segmentation model to the 3D occupancy network, and (2) pixel-voxel filtering for high-quality training data generation. The resulting framework is simple, compact, and compatible with most state-of-the-art semantic occupancy prediction models. On NYUv2 and SemanticKITTI datasets, OVO achieves competitive performance compared to supervised semantic occupancy prediction approaches. Furthermore, we conduct extensive analyses and ablation studies to offer insights into the design of the proposed framework. Our code is publicly available at https://github.com/dzcgaara/OVO.

FaQiang Qian, WeiKun Zhang, Ziliang Wang et al.

Shaping powerful LLMs to be beneficial and safe is central to AI alignment. We argue that post-training alignment is fundamentally a unified Preference Learning problem, involving two modalities: demonstrated preferences (e.g., Supervised Fine-Tuning, SFT) and comparative preferences (e.g., Reinforcement Learning, RL).The standard sequential pipeline-SFT followed by RL-is flawed due to a critical distributional mismatch: SFT uses static expert data, but as the policy evolves, its generation distribution drifts, making SFT knowledge brittle. Subsequent RL then explores without direct access to the rich, ground-truth knowledge in expert demonstrations, leading to inefficient, ungrounded updates. This separation prevents mutual regularization between data sources. To address this, we reframe alignment as a constrained optimization problem and propose Unified Adversarial Preference Learning (UniAPL),a novel framework that dynamically aligns the policy's distribution with the expert's. UniAPL implements a single-stage unified training objective, jointly learning from mixed batches of SFT and preference data. In every gradient step, dense expert demonstrations directly ground and regularize online exploration, inherently resolving distributional mismatch and maximizing data synergy.We evaluate UniAPL on instruction-following tasks using Qwen3-235B-Instruct-2507 as the teacher. Our models match or exceed strong GRPO baselines: +5.77% on Qwen3-0.6B (matching a 32B model) and +3.75% on Qwen3-4B,even outperforming the teacher. Analyses of response length and log-probability distributions confirm that UniAPL outputs closely mimic expert demonstrations, achieving both stronger performance and better behavioral alignment.

Ziliang Wang, Kang An, Xuhui Zheng et al.

While search-augmented large language models (LLMs) exhibit impressive capabilities, their reliability in complex multi-hop reasoning remains limited. This limitation arises from three fundamental challenges: decomposition errors, where tasks are incorrectly broken down; retrieval missing, where key evidence fails to be retrieved; and reasoning errors, where flawed logic propagates through the reasoning chain. A single failure in any of these stages can derail the final answer. We propose Erasable Reinforcement Learning (ERL), a novel framework that transforms fragile reasoning into a robust process. ERL explicitly identifies faulty steps, erases them, and regenerates reasoning in place, preventing defective logic from propagating through the reasoning chain. This targeted correction mechanism turns brittle reasoning into a more resilient process. Models trained with ERL, termed ESearch, achieve substantial improvements on HotpotQA, MuSiQue, 2Wiki, and Bamboogle, with the 3B model achieving +8.48% EM and +11.56% F1, and the 7B model achieving +5.38% EM and +7.22% F1 over previous state-of-the-art(SOTA) results. These findings suggest that erasable reinforcement learning provides a powerful paradigm shift for robust multi-step reasoning in LLMs.

Weikun Zhang, Jianke Zhu

With the growing popularity of neural rendering, there has been an increasing number of neural implicit multi-view reconstruction methods. While many models have been enhanced in terms of positional encoding, sampling, rendering, and other aspects to improve the reconstruction quality, current methods do not fully leverage the information among neighboring pixels during the reconstruction process. To address this issue, we propose an enhanced model called BundleRecon. In the existing approaches, sampling is performed by a single ray that corresponds to a single pixel. In contrast, our model samples a patch of pixels using a bundle of rays, which incorporates information from neighboring pixels. Furthermore, we design bundle-based constraints to further improve the reconstruction quality. Experimental results demonstrate that BundleRecon is compatible with the existing neural implicit multi-view reconstruction methods and can improve their reconstruction quality.