Siyu Zhang

Shangqing Tu, Zheyuan Zhang, Jifan Yu et al. · tsinghua

Teaching assistants have played essential roles in the long history of education. However, few MOOC platforms are providing human or virtual teaching assistants to support learning for massive online students due to the complexity of real-world online education scenarios and the lack of training data. In this paper, we present a virtual MOOC teaching assistant, LittleMu with minimum labeled training data, to provide question answering and chit-chat services. Consisting of two interactive modules of heterogeneous retrieval and language model prompting, LittleMu first integrates structural, semi- and unstructured knowledge sources to support accurate answers for a wide range of questions. Then, we design delicate demonstrations named "Chain of Teach" prompts to exploit the large-scale pre-trained model to handle complex uncollected questions. Except for question answering, we develop other educational services such as knowledge-grounded chit-chat. We test the system's performance via both offline evaluation and online deployment. Since May 2020, our LittleMu system has served over 80,000 users with over 300,000 queries from over 500 courses on XuetangX MOOC platform, which continuously contributes to a more convenient and fair education. Our code, services, and dataset will be available at https://github.com/THU-KEG/VTA.

Shuheng Cao, Ruiqi Chen, Renjie Cao et al.

Biomedical NER is deceptively simple for modern LLMs: plausible biomedical mentions are easy to surface, but corpus-convention correctness depends on annotation conventions, span boundaries, entity granularity, and type schemas. Multi-LLM agreement is a salience signal, not corpus-convention correctness. We introduce a candidate-level panel-output benchmark for panel-surfaced candidate verification, where the unit is an aligned candidate surfaced by an explicitly defined multi-model panel rather than a standalone extractor output. The benchmark aligns eight LLMs' predictions over five public biomedical NER datasets into a candidate master table. BioConCal is an in-domain supervised scorer that instantiates this layer with inference-time gold-free agreement, mention, surface-availability, and document features for a fixed candidate stream. In domain, BioConCal improves AUROC from 0.753 for raw agreement to 0.910. At a validation-selected 0.95 precision target it selects 1,340 candidates at empirical test precision 0.939, compared with 293 for raw agreement. This corresponds to candidate-level recall 0.592 and corpus-level recall 0.523 against a within-panel row-label ceiling of 0.883. The main benefit is not recovering entities missed by every panel member, but reshaping a noisy panel stream into a higher-yield review queue. Under entity-type shift, thresholds require target-domain validation, and exact character localization remains a separate deterministic post-processing step.

Jiaming Sun, Zihao Wang, Siyu Zhang et al.

We propose a new method named OnePose for object pose estimation. Unlike existing instance-level or category-level methods, OnePose does not rely on CAD models and can handle objects in arbitrary categories without instance- or category-specific network training. OnePose draws the idea from visual localization and only requires a simple RGB video scan of the object to build a sparse SfM model of the object. Then, this model is registered to new query images with a generic feature matching network. To mitigate the slow runtime of existing visual localization methods, we propose a new graph attention network that directly matches 2D interest points in the query image with the 3D points in the SfM model, resulting in efficient and robust pose estimation. Combined with a feature-based pose tracker, OnePose is able to stably detect and track 6D poses of everyday household objects in real-time. We also collected a large-scale dataset that consists of 450 sequences of 150 objects.

Haoqin Tu, Zhongliang Yang, Jinshuai Yang et al.

Controllable text generation has taken a gigantic step forward these days. Yet existing methods are either constrained in a one-off pattern or not efficient enough for receiving multiple conditions at every generation stage. We propose a model-agnostic framework Plug-in Conditional Auto-Encoder for Controllable Text Generation (PCAE) towards flexible and semi-supervised text generation. Our framework is "plug-and-play" with partial parameters to be fine-tuned in the pre-trained model (less than a half). Crucial to the success of PCAE is the proposed broadcasting label fusion network for navigating the global latent code to a specified local and confined space. Visualization of the local latent prior well confirms the primary devotion in hidden space of the proposed model. Moreover, extensive experiments across five related generation tasks (from 2 conditions up to 10 conditions) on both RNN- based and pre-trained BART [26] based auto-encoders reveal the high capability of PCAE, which enables generation that is highly manipulable, syntactically diverse and time-saving with minimum labeled samples. We will release our code at https://github.com/ImKeTT/pcae.

Xiyang Wang, Chunyun Fu, Jiawei He et al.

In the classical tracking-by-detection (TBD) paradigm, detection and tracking are separately and sequentially conducted, and data association must be properly performed to achieve satisfactory tracking performance. In this paper, a new end-to-end multi-object tracking framework is proposed, which integrates object detection and multi-object tracking into a single model. The proposed tracking framework eliminates the complex data association process in the classical TBD paradigm, and requires no additional training. Secondly, the regression confidence of historical trajectories is investigated, and the possible states of a trajectory (weak object or strong object) in the current frame are predicted. Then, a confidence fusion module is designed to guide non-maximum suppression for trajectories and detections to achieve ordered and robust tracking. Thirdly, by integrating historical trajectory features, the regression performance of the detector is enhanced, which better reflects the occlusion and disappearance patterns of objects in real world. Lastly, extensive experiments are conducted on the commonly used KITTI and Waymo datasets. The results show that the proposed framework can achieve robust tracking by using only a 2D detector and a 3D detector, and it is proven more accurate than many of the state-of-the-art TBD-based multi-modal tracking methods. The source codes of the proposed method are available at https://github.com/wangxiyang2022/YONTD-MOT.

Xiyang Wang, Shouzheng Qi, Jieyou Zhao et al.

This paper introduces MCTrack, a new 3D multi-object tracking method that achieves state-of-the-art (SOTA) performance across KITTI, nuScenes, and Waymo datasets. Addressing the gap in existing tracking paradigms, which often perform well on specific datasets but lack generalizability, MCTrack offers a unified solution. Additionally, we have standardized the format of perceptual results across various datasets, termed BaseVersion, facilitating researchers in the field of multi-object tracking (MOT) to concentrate on the core algorithmic development without the undue burden of data preprocessing. Finally, recognizing the limitations of current evaluation metrics, we propose a novel set that assesses motion information output, such as velocity and acceleration, crucial for downstream tasks. The source codes of the proposed method are available at this link: https://github.com/megvii-research/MCTrack}{https://github.com/megvii-research/MCTrack

Siyu Zhang, Yeming Chen, Yaoru Sun et al.

Visual question answering (VQA) has been intensively studied as a multimodal task that requires effort in bridging vision and language to infer answers correctly. Recent attempts have developed various attention-based modules for solving VQA tasks. However, the performance of model inference is largely bottlenecked by visual processing for semantics understanding. Most existing detection methods rely on bounding boxes, remaining a serious challenge for VQA models to understand the causal nexus of object semantics in images and correctly infer contextual information. To this end, we propose a finer model framework without bounding boxes in this work, termed Looking Out of Instance Semantics (LOIS) to tackle this important issue. LOIS enables more fine-grained feature descriptions to produce visual facts. Furthermore, to overcome the label ambiguity caused by instance masks, two types of relation attention modules: 1) intra-modality and 2) inter-modality, are devised to infer the correct answers from the different multi-view features. Specifically, we implement a mutual relation attention module to model sophisticated and deeper visual semantic relations between instance objects and background information. In addition, our proposed attention model can further analyze salient image regions by focusing on important word-related questions. Experimental results on four benchmark VQA datasets prove that our proposed method has favorable performance in improving visual reasoning capability.

Haoran Wang, Zeshen Tang, Leya Yang et al.

Goal-conditioned hierarchical reinforcement learning (HRL) presents a promising approach for enabling effective exploration in complex, long-horizon reinforcement learning (RL) tasks through temporal abstraction. Empirically, heightened inter-level communication and coordination can induce more stable and robust policy improvement in hierarchical systems. Yet, most existing goal-conditioned HRL algorithms have primarily focused on the subgoal discovery, neglecting inter-level cooperation. Here, we propose a goal-conditioned HRL framework named Guided Cooperation via Model-based Rollout (GCMR), aiming to bridge inter-layer information synchronization and cooperation by exploiting forward dynamics. Firstly, the GCMR mitigates the state-transition error within off-policy correction via model-based rollout, thereby enhancing sample efficiency. Secondly, to prevent disruption by the unseen subgoals and states, lower-level Q-function gradients are constrained using a gradient penalty with a model-inferred upper bound, leading to a more stable behavioral policy conducive to effective exploration. Thirdly, we propose a one-step rollout-based planning, using higher-level critics to guide the lower-level policy. Specifically, we estimate the value of future states of the lower-level policy using the higher-level critic function, thereby transmitting global task information downwards to avoid local pitfalls. These three critical components in GCMR are expected to facilitate inter-level cooperation significantly. Experimental results demonstrate that incorporating the proposed GCMR framework with a disentangled variant of HIGL, namely ACLG, yields more stable and robust policy improvement compared to various baselines and significantly outperforms previous state-of-the-art algorithms.

Yeming Chen, Siyu Zhang, Yaoru Sun et al.

With the success of self-supervised learning, multimodal foundation models have rapidly adapted a wide range of downstream tasks driven by vision and language (VL) pretraining. State-of-the-art methods achieve impressive performance by pre-training on large-scale datasets. However, bridging the semantic gap between the two modalities remains a nonnegligible challenge for VL tasks. In this work, we propose an efficient computation framework for multimodal alignment by introducing a novel visual semantic module to further improve the performance of the VL tasks. Specifically, we propose a flexible model, namely Artificial-Spiking Hierarchical Networks (ASH-Nets), which combines the complementary advantages of Artificial neural networks (ANNs) and Spiking neural networks (SNNs) to enrich visual semantic representations. In particular, a visual concrete encoder and a semantic abstract encoder are constructed to learn continuous and discrete latent variables to enhance the flexibility of semantic encoding. Considering the spatio-temporal properties of SNNs modeling, we introduce a contrastive learning method to optimize the inputs of similar samples. This can improve the computational efficiency of the hierarchical network, while the augmentation of hard samples is beneficial to the learning of visual representations. Furthermore, the Spiking to Text Uni-Alignment Learning (STUA) pre-training method is proposed, which only relies on text features to enhance the encoding ability of abstract semantics. We validate the performance on multiple well-established downstream VL tasks. Experiments show that the proposed ASH-Nets achieve competitive results.

Siyu Zhang, Yeming Chen, Yaoru Sun et al.

The key to integrating visual language tasks is to establish a good alignment strategy. Recently, visual semantic representation has achieved fine-grained visual understanding by dividing grids or image patches. However, the coarse-grained semantic interactions in image space should not be ignored, which hinders the extraction of complex contextual semantic relations at the scene boundaries. This paper proposes superpixels as comprehensive and robust visual primitives, which mine coarse-grained semantic interactions by clustering perceptually similar pixels, speeding up the subsequent processing of primitives. To capture superpixel-level semantic features, we propose a Multiscale Difference Graph Convolutional Network (MDGCN). It allows parsing the entire image as a fine-to-coarse visual hierarchy. To reason actual semantic relations, we reduce potential noise interference by aggregating difference information between adjacent graph nodes. Finally, we propose a multi-level fusion rule in a bottom-up manner to avoid understanding deviation by mining complementary spatial information at different levels. Experiments show that the proposed method can effectively promote the learning of multiple downstream tasks. Encouragingly, our method outperforms previous methods on all metrics. Our code will be released upon publication.

Siyu Zhang, Lianlei Shan, Runhe Qiu

Multimodal fusion of remote sensing images serves as a core technology for overcoming the limitations of single-source data and improving the accuracy of surface information extraction, which exhibits significant application value in fields such as environmental monitoring and urban planning. To address the deficiencies of existing methods, including the failure of fixed resolutions to balance efficiency and detail, as well as the lack of semantic hierarchy in single-scale alignment, this study proposes a Vision-language Model (VLM) framework integrated with two key innovations: the Dynamic Resolution Input Strategy (DRIS) and the Multi-scale Vision-language Alignment Mechanism (MS-VLAM).Specifically, the DRIS adopts a coarse-to-fine approach to adaptively allocate computational resources according to the complexity of image content, thereby preserving key fine-grained features while reducing redundant computational overhead. The MS-VLAM constructs a three-tier alignment mechanism covering object, local-region and global levels, which systematically captures cross-modal semantic consistency and alleviates issues of semantic misalignment and granularity imbalance.Experimental results on the RS-GPT4V dataset demonstrate that the proposed framework significantly improves the accuracy of semantic understanding and computational efficiency in tasks including image captioning and cross-modal retrieval. Compared with conventional methods, it achieves superior performance in evaluation metrics such as BLEU-4 and CIDEr for image captioning, as well as R@10 for cross-modal retrieval. This technical framework provides a novel approach for constructing efficient and robust multimodal remote sensing systems, laying a theoretical foundation and offering technical guidance for the engineering application of intelligent remote sensing interpretation.

Beining Wu, Fuyou Mao, Jiong Lin et al.

Generative engines (GEs) are reshaping information access by replacing ranked links with citation-grounded answers, yet current Generative Engine Optimization (GEO) methods optimize each instance in isolation, unable to accumulate or transfer effective strategies across tasks and engines. We reframe GEO as a strategy learning problem and propose MAGEO, a multi-agent framework in which coordinated planning, editing, and fidelity-aware evaluation serve as the execution layer, while validated editing patterns are progressively distilled into reusable, engine-specific optimization skills. To enable controlled assessment, we introduce a Twin Branch Evaluation Protocol for causal attribution of content edits and DSV-CF, a dual-axis metric that unifies semantic visibility with attribution accuracy. We further release MSME-GEO-Bench, a multi-scenario, multi-engine benchmark grounded in real-world queries. Experiments on three mainstream engines show that MAGEO substantially outperforms heuristic baselines in both visibility and citation fidelity, with ablations confirming that engine-specific preference modeling and strategy reuse are central to these gains, suggesting a scalable learning-driven paradigm for trustworthy GEO. Code is available at https://github.com/Wu-beining/MAGEO

Yu Huo, Kun Zeng, Siyu Zhang et al.

Repository-level code completion benefits from retrieval-augmented generation (RAG). However, controlling cross-file evidence is difficult because chunk utility is often interaction-dependent: some snippets help only when paired with complementary context, while others harm decoding when they conflict. We propose RepoShapley, a coalition-aware context filtering framework supervised by Shapley-style marginal contributions. Our offline labeling module, ChunkShapley, estimates signed per-chunk effects via teacher-forced probing, feeds them into a lightweight surrogate game that captures saturation and interference, computes exact Shapley values for small retrieval sets, and selects a decoding-optimal coalition through bounded post-verification with the frozen generator. The verified keep/drop decisions and retrieval triggers are then distilled into a single model via discrete control tokens. Experiments across benchmarks and backbones show that RepoShapley improves completion quality while reducing harmful context and unnecessary retrieval.

Shen Cai, Zhanhao Wu, Lingxi Guo et al.

In this paper, we present two fast and interpretable decomposition methods for 2D homography, which are named Similarity-Kernel-Similarity (SKS) and Affine-Core-Affine (ACA) transformations respectively. Under the minimal $4$-point configuration, the first and the last similarity transformations in SKS are computed by two anchor points on target and source planes, respectively. Then, the other two point correspondences can be exploited to compute the middle kernel transformation with only four parameters. Furthermore, ACA uses three anchor points to compute the first and the last affine transformations, followed by computation of the middle core transformation utilizing the other one point correspondence. ACA can compute a homography up to a scale with only $85$ floating-point operations (FLOPs), without even any division operations. Therefore, as a plug-in module, ACA facilitates the traditional feature-based Random Sample Consensus (RANSAC) pipeline, as well as deep homography pipelines estimating $4$-point offsets. In addition to the advantages of geometric parameterization and computational efficiency, SKS and ACA can express each element of homography by a polynomial of input coordinates ($7$th degree to $9$th degree), extend the existing essential Similarity-Affine-Projective (SAP) decomposition and calculate 2D affine transformations in a unified way. Source codes are released in https://github.com/cscvlab/SKS-Homography.

Yifu Guo, Yuquan Lu, Wentao Zhang et al.

Continual Semantic Segmentation (CSS) requires learning new classes without forgetting previously acquired knowledge, addressing the fundamental challenge of catastrophic forgetting in dense prediction tasks. However, existing CSS methods typically employ single-stage encoder-decoder architectures where segmentation masks and class labels are tightly coupled, leading to interference between old and new class learning and suboptimal retention-plasticity balance. We introduce DecoupleCSS, a novel two-stage framework for CSS. By decoupling class-aware detection from class-agnostic segmentation, DecoupleCSS enables more effective continual learning, preserving past knowledge while learning new classes. The first stage leverages pre-trained text and image encoders, adapted using LoRA, to encode class-specific information and generate location-aware prompts. In the second stage, the Segment Anything Model (SAM) is employed to produce precise segmentation masks, ensuring that segmentation knowledge is shared across both new and previous classes. This approach improves the balance between retention and adaptability in CSS, achieving state-of-the-art performance across a variety of challenging tasks. Our code is publicly available at: https://github.com/euyis1019/Decoupling-Continual-Semantic-Segmentation.

Yu Huo, Siyu Zhang, Kun Zeng et al.

Multimodal models for text-to-image generation have achieved strong visual fidelity, yet they remain brittle under compositional structural constraints-notably generative numeracy, attribute binding, and part-level relations. To address these challenges, we propose Shape-of-Thought (SoT), a visual CoT framework that enables progressive shape assembly via coherent 2D projections without external engines at inference time. SoT trains a unified multimodal autoregressive model to generate interleaved textual plans and rendered intermediate states, helping the model capture shape-assembly logic without producing explicit geometric representations. To support this paradigm, we introduce SoT-26K, a large-scale dataset of grounded assembly traces derived from part-based CAD hierarchies, and T2S-CompBench, a benchmark for evaluating structural integrity and trace faithfulness. Fine-tuning on SoT-26K achieves 88.4% on component numeracy and 84.8% on structural topology, outperforming text-only baselines by around 20%. SoT establishes a new paradigm for transparent, process-supervised compositional generation. The code is available at https://anonymous.4open.science/r/16FE/. The SoT-26K dataset will be released upon acceptance.

Yuquan Lu, Yifu Guo, Zishan Xu et al.

Continual semantic segmentation (CSS) is a cornerstone task in computer vision that enables a large number of downstream applications, but faces the catastrophic forgetting challenge. In conventional class-incremental semantic segmentation (CISS) frameworks using Softmax-based classification heads, catastrophic forgetting originates from Catastrophic forgetting and task affiliation probability. We formulate these problems and provide a theoretical analysis to more deeply understand the limitations in existing CISS methods, particularly Strict Parameter Isolation (SPI). To address these challenges, we follow a dual-phase intuition from human annotators, and introduce Cognitive Cascade Segmentation (CogCaS), a novel dual-phase cascade formulation for CSS tasks in the CISS setting. By decoupling the task into class-existence detection and class-specific segmentation, CogCaS enables more effective continual learning, preserving previously learned knowledge while incorporating new classes. Using two benchmark datasets PASCAL VOC 2012 and ADE20K, we have shown significant improvements in a variety of challenging scenarios, particularly those with long sequence of incremental tasks, when compared to exsiting state-of-the-art methods. Our code will be made publicly available upon paper acceptance.

Kun Zeng, Yu Huo, Siyu Zhang et al.

Skill-augmented agents increasingly rely on large reusable skill libraries, but retrieving relevant skills is not the same as presenting usable context. Existing methods typically return atomic skills or dependency-aware bundles whose internal roles remain implicit, leaving the agent to infer the execution entry point, support skills, visible requirements, and failure-avoidance guidance. We introduce Group of Skills (GoSkills), an inference-time group-structured retrieval method that changes the agent-facing retrieval object from a flat skill list to a compact, role-labeled execution context. GoSkills builds anchor-centered skill groups from a typed skill graph, expands support groups through a group graph, bottlenecks the selected group plan into a bounded set of atomic skill payloads, and renders a fixed execution contract with Start, Support, Check, and Avoid fields, without changing the downstream agent, skill payloads, or execution environment. Experiments on SkillsBench and ALFWorld show that GoSkills preserves visible-requirement coverage under a small skill budget, improves over flat skill-access baselines, and often improves reward and agent-only runtime relative to structural retrieval references.

Zhiqing Cui, Haotong Xie, Jiahao Yuan et al.

Large language model (LLM) multi-agent systems typically rely on rigid orchestration, committing either to flat per-query routing or to hand-engineered task decomposition, so decomposition depth, worker choice, and inference budget are not jointly optimized under one objective. We introduce Uno-Orchestra, a unified orchestration policy that selectively decomposes a task and dispatches each subtask to an admissible (model, primitive) pair, with both decisions learned together from curated RL trajectories grounded in real worker interactions. Against 22 baselines on a 13-benchmark suite spanning math, code, knowledge, long-context, and agentic tool-use, Uno-Orchestra reaches 77.0% macro pass@1, roughly 16% above the strongest workflow baseline, at roughly an order of magnitude lower per-query cost, advancing the accuracy-efficiency frontier of selective delegation.

Team Seawead, Ceyuan Yang, Zhijie Lin et al.

This technical report presents a cost-efficient strategy for training a video generation foundation model. We present a mid-sized research model with approximately 7 billion parameters (7B) called Seaweed-7B trained from scratch using 665,000 H100 GPU hours. Despite being trained with moderate computational resources, Seaweed-7B demonstrates highly competitive performance compared to contemporary video generation models of much larger size. Design choices are especially crucial in a resource-constrained setting. This technical report highlights the key design decisions that enhance the performance of the medium-sized diffusion model. Empirically, we make two observations: (1) Seaweed-7B achieves performance comparable to, or even surpasses, larger models trained on substantially greater GPU resources, and (2) our model, which exhibits strong generalization ability, can be effectively adapted across a wide range of downstream applications either by lightweight fine-tuning or continue training. See the project page at https://seaweed.video/

Diankun Zhang, Guoan Wang, Runwen Zhu et al.

End-to-End paradigms use a unified framework to implement multi-tasks in an autonomous driving system. Despite simplicity and clarity, the performance of end-to-end autonomous driving methods on sub-tasks is still far behind the single-task methods. Meanwhile, the widely used dense BEV features in previous end-to-end methods make it costly to extend to more modalities or tasks. In this paper, we propose a Sparse query-centric paradigm for end-to-end Autonomous Driving (SparseAD), where the sparse queries completely represent the whole driving scenario across space, time and tasks without any dense BEV representation. Concretely, we design a unified sparse architecture for perception tasks including detection, tracking, and online mapping. Moreover, we revisit motion prediction and planning, and devise a more justifiable motion planner framework. On the challenging nuScenes dataset, SparseAD achieves SOTA full-task performance among end-to-end methods and significantly narrows the performance gap between end-to-end paradigms and single-task methods. Codes will be released soon.

Sirui Cheng, Siyu Zhang, Jiayi Wu et al.

Within the multimodal field, large vision-language models (LVLMs) have made significant progress due to their strong perception and reasoning capabilities in the visual and language systems. However, LVLMs are still plagued by the two critical issues of object hallucination and factual accuracy, which limit the practicality of LVLMs in different scenarios. Furthermore, previous evaluation methods focus more on the comprehension and reasoning of language content but lack a comprehensive evaluation of multimodal interactions, thereby resulting in potential limitations. To this end, we propose a novel KNVQA-Eval, which is devoted to knowledge-based VQA task evaluation to reflect the factuality of multimodal LVLMs. To ensure the robustness and scalability of the evaluation, we develop a new KNVQA dataset by incorporating human judgment and perception, aiming to evaluate the accuracy of standard answers relative to AI-generated answers in knowledge-based VQA. This work not only comprehensively evaluates the contextual information of LVLMs using reliable human annotations, but also further analyzes the fine-grained capabilities of current methods to reveal potential avenues for subsequent optimization of LVLMs-based estimators. Our proposed VQA-Eval and corresponding dataset KNVQA will facilitate the development of automatic evaluation tools with the advantages of low cost, privacy protection, and reproducibility. Our code will be released upon publication.

Siyu Zhang

Sparse neural networks are often hypothesized to be more interpretable than dense models, motivated by findings that weight sparsity can produce compact circuits in language models. However, it remains unclear whether structural sparsity itself leads to improved semantic interpretability. In this work, we systematically evaluate the relationship between weight sparsity and interpretability in Vision Transformers using DeiT-III B/16 models pruned with Wanda. To assess interpretability comprehensively, we introduce \textbf{IMPACT}, a multi-level framework that evaluates interpretability across four complementary levels: neurons, layer representations, task circuits, and model-level attribution. Layer representations are analyzed using BatchTopK sparse autoencoders, circuits are extracted via learnable node masking, and explanations are evaluated with transformer attribution using insertion and deletion metrics. Our results reveal a clear structural effect but limited interpretability gains. Sparse models produce circuits with approximately $2.5\times$ fewer edges than dense models, yet the fraction of active nodes remains similar or higher, indicating that pruning redistributes computation rather than isolating simpler functional modules. Consistent with this observation, sparse models show no systematic improvements in neuron-level selectivity, SAE feature interpretability, or attribution faithfulness. These findings suggest that structural sparsity alone does not reliably yield more interpretable vision models, highlighting the importance of evaluation frameworks that assess interpretability beyond circuit compactness.

Lizhi Bai, Chunqi Tian, Jun Yang et al.

3D Gaussian Splatting has emerged as a promising technique for high-quality 3D rendering, leading to increasing interest in integrating 3DGS into realism SLAM systems. However, existing methods face challenges such as Gaussian primitives redundancy, forgetting problem during continuous optimization, and difficulty in initializing primitives in monocular case due to lack of depth information. In order to achieve efficient and photorealistic mapping, we propose RP-SLAM, a 3D Gaussian splatting-based vision SLAM method for monocular and RGB-D cameras. RP-SLAM decouples camera poses estimation from Gaussian primitives optimization and consists of three key components. Firstly, we propose an efficient incremental mapping approach to achieve a compact and accurate representation of the scene through adaptive sampling and Gaussian primitives filtering. Secondly, a dynamic window optimization method is proposed to mitigate the forgetting problem and improve map consistency. Finally, for the monocular case, a monocular keyframe initialization method based on sparse point cloud is proposed to improve the initialization accuracy of Gaussian primitives, which provides a geometric basis for subsequent optimization. The results of numerous experiments demonstrate that RP-SLAM achieves state-of-the-art map rendering accuracy while ensuring real-time performance and model compactness.

Siyu Zhang, Kenneth Mcmillan

Interpretable and faithful explanations for specific neural inferences are crucial for understanding and evaluating model behavior. Our work introduces \textbf{F}aithfulness-guided \textbf{E}nsemble \textbf{I}nterpretation (\textbf{FEI}), an innovative framework that enhances the breadth and effectiveness of faithfulness, advancing interpretability by providing superior visualization. Through an analysis of existing evaluation benchmarks, \textbf{FEI} employs a smooth approximation to elevate quantitative faithfulness scores. Diverse variations of \textbf{FEI} target enhanced faithfulness in hidden layer encodings, expanding interpretability. Additionally, we propose a novel qualitative metric that assesses hidden layer faithfulness. In extensive experiments, \textbf{FEI} surpasses existing methods, demonstrating substantial advances in qualitative visualization and quantitative faithfulness scores. Our research establishes a comprehensive framework for elevating faithfulness in neural network explanations, emphasizing both breadth and precision

Siyu Zhang, Wenzhe Liu, Yeming Chen et al.

To bridge the semantic gap between vision and language (VL), it is necessary to develop a good alignment strategy, which includes handling semantic diversity, abstract representation of visual information, and generalization ability of models. Recent works use detector-based bounding boxes or patches with regular partitions to represent visual semantics. While current paradigms have made strides, they are still insufficient for fully capturing the nuanced contextual relations among various objects. This paper proposes a comprehensive visual semantic representation module, necessitating the utilization of panoptic segmentation to generate coherent fine-grained semantic features. Furthermore, we propose a novel Graph Spiking Hybrid Network (GSHN) that integrates the complementary advantages of Spiking Neural Networks (SNNs) and Graph Attention Networks (GATs) to encode visual semantic information. Intriguingly, the model not only encodes the discrete and continuous latent variables of instances but also adeptly captures both local and global contextual features, thereby significantly enhancing the richness and diversity of semantic representations. Leveraging the spatiotemporal properties inherent in SNNs, we employ contrastive learning (CL) to enhance the similarity-based representation of embeddings. This strategy alleviates the computational overhead of the model and enriches meaningful visual representations by constructing positive and negative sample pairs. We design an innovative pre-training method, Spiked Text Learning (STL), which uses text features to improve the encoding ability of discrete semantics. Experiments show that the proposed GSHN exhibits promising results on multiple VL downstream tasks.

Jiaheng Zhuang, Guoan Wang, Siyu Zhang et al.

3D multi-object tracking and trajectory prediction are two crucial modules in autonomous driving systems. Generally, the two tasks are handled separately in traditional paradigms and a few methods have started to explore modeling these two tasks in a joint manner recently. However, these approaches suffer from the limitations of single-frame training and inconsistent coordinate representations between tracking and prediction tasks. In this paper, we propose a streaming and unified framework for joint 3D Multi-Object Tracking and trajectory Prediction (StreamMOTP) to address the above challenges. Firstly, we construct the model in a streaming manner and exploit a memory bank to preserve and leverage the long-term latent features for tracked objects more effectively. Secondly, a relative spatio-temporal positional encoding strategy is introduced to bridge the gap of coordinate representations between the two tasks and maintain the pose-invariance for trajectory prediction. Thirdly, we further improve the quality and consistency of predicted trajectories with a dual-stream predictor. We conduct extensive experiments on popular nuSences dataset and the experimental results demonstrate the effectiveness and superiority of StreamMOTP, which outperforms previous methods significantly on both tasks. Furthermore, we also prove that the proposed framework has great potential and advantages in actual applications of autonomous driving.

Yongdong Huang, Yuanzhan Li, Xulong Cao et al.

Using neural networks to represent 3D objects has become popular. However, many previous works employ neural networks with fixed architecture and size to represent different 3D objects, which lead to excessive network parameters for simple objects and limited reconstruction accuracy for complex objects. For each 3D model, it is desirable to have an end-to-end neural network with as few parameters as possible to achieve high-fidelity reconstruction. In this paper, we propose an efficient voxel reconstruction method utilizing neural architecture search (NAS) and binary classification. Taking the number of layers, the number of nodes in each layer, and the activation function of each layer as the search space, a specific network architecture can be obtained based on reinforcement learning technology. Furthermore, to get rid of the traditional surface reconstruction algorithms (e.g., marching cube) used after network inference, we complete the end-to-end network by classifying binary voxels. Compared to other signed distance field (SDF) prediction or binary classification networks, our method achieves significantly higher reconstruction accuracy using fewer network parameters.

Yuanzhan Li, Yuqi Liu, Yujie Lu et al.

In recent years, neural signed distance function (SDF) has become one of the most effective representation methods for 3D models. By learning continuous SDFs in 3D space, neural networks can predict the distance from a given query space point to its closest object surface,whose positive and negative signs denote inside and outside of the object, respectively. Training a specific network for each 3D model, which individually embeds its shape, can realize compressed representation of objects by storing fewer network (and possibly latent) parameters. Consequently, reconstruction through network inference and surface recovery can be achieved. In this paper, we propose an SDF prediction network using explicit key spheres as input. Key spheres are extracted from the internal space of objects, whose centers either have relatively larger SDF values (sphere radii), or are located at essential positions. By inputting the spatial information of multiple spheres which imply different local shapes, the proposed method can significantly improve the reconstruction accuracy with a negligible storage cost. Compared to previous works, our method achieves the high-fidelity and high-compression 3D object coding and reconstruction. Experiments conducted on three datasets verify the superior performance of our method.

Siyu Zhang, Zhongliang Yang, Haoqin Tu et al.

In this letter, we explored generative image steganography based on autoregressive models. We proposed Pixel-Stega, which implements pixel-level information hiding with autoregressive models and arithmetic coding algorithm. Firstly, one of the autoregressive models, PixelCNN++, is utilized to produce explicit conditional probability distribution of each pixel. Secondly, secret messages are encoded to the selection of pixels through steganographic sampling (stegosampling) based on arithmetic coding. We carried out qualitative and quantitative assessment on gray-scale and colour image datasets. Experimental results show that Pixel-Stega is able to embed secret messages adaptively according to the entropy of the pixels to achieve both high embedding capacity (up to 4.3 bpp) and nearly perfect imperceptibility (about 50% detection accuracy).

Siyu Zhang, Zhongliang Yang, Jinshuai Yang et al.

Generative linguistic steganography mainly utilized language models and applied steganographic sampling (stegosampling) to generate high-security steganographic text (stegotext). However, previous methods generally lead to statistical differences between the conditional probability distributions of stegotext and natural text, which brings about security risks. In this paper, to further ensure security, we present a novel provably secure generative linguistic steganographic method ADG, which recursively embeds secret information by Adaptive Dynamic Grouping of tokens according to their probability given by an off-the-shelf language model. We not only prove the security of ADG mathematically, but also conduct extensive experiments on three public corpora to further verify its imperceptibility. The experimental results reveal that the proposed method is able to generate stegotext with nearly perfect security.

Siyu Zhang, Hui Cao, Yuqi Liu et al.

Using deep learning techniques to process 3D objects has achieved many successes. However, few methods focus on the representation of 3D objects, which could be more effective for specific tasks than traditional representations, such as point clouds, voxels, and multi-view images. In this paper, we propose a Sphere Node Graph (SN-Graph) to represent 3D objects. Specifically, we extract a certain number of internal spheres (as nodes) from the signed distance field (SDF), and then establish connections (as edges) among the sphere nodes to construct a graph, which is seamlessly suitable for 3D analysis using graph neural network (GNN). Experiments conducted on the ModelNet40 dataset show that when there are fewer nodes in the graph or the tested objects are rotated arbitrarily, the classification accuracy of SN-Graph is significantly higher than the state-of-the-art methods.

Ji Luo, Hui Cao, Jie Wang et al.

Voxel-based 3D object classification has been thoroughly studied in recent years. Most previous methods convert the classic 2D convolution into a 3D form that will be further applied to objects with binary voxel representation for classification. However, the binary voxel representation is not very effective for 3D convolution in many cases. In this paper, we propose a hybrid cascade architecture for voxel-based 3D object classification. It consists of three stages composed of fully connected and convolutional layers, dealing with easy, moderate, and hard 3D models respectively. Both accuracy and speed can be balanced in our proposed method. By giving each voxel a signed distance value, an obvious gain regarding the accuracy can be observed. Besides, the mean inference time can be speeded up hugely compared with the state-of-the-art point cloud and voxel based methods.

Lamei Zhang, Siyu Zhang, Bin Zou et al.

Deep learning and convolutional neural networks (CNNs) have made progress in polarimetric synthetic aperture radar (PolSAR) image classification over the past few years. However, a crucial issue has not been addressed, i.e., the requirement of CNNs for abundant labeled samples versus the insufficient human annotations of PolSAR images. It is well-known that following the supervised learning paradigm may lead to the overfitting of training data, and the lack of supervision information of PolSAR images undoubtedly aggravates this problem, which greatly affects the generalization performance of CNN-based classifiers in large-scale applications. To handle this problem, in this paper, learning transferrable representations from unlabeled PolSAR data through convolutional architectures is explored for the first time. Specifically, a PolSAR-tailored contrastive learning network (PCLNet) is proposed for unsupervised deep PolSAR representation learning and few-shot classification. Different from the utilization of optical processing methods, a diversity stimulation mechanism is constructed to narrow the application gap between optics and PolSAR. Beyond the conventional supervised methods, PCLNet develops an unsupervised pre-training phase based on the proxy objective of instance discrimination to learn useful representations from unlabeled PolSAR data. The acquired representations are transferred to the downstream task, i.e., few-shot PolSAR classification. Experiments on two widely-used PolSAR benchmark datasets confirm the validity of PCLNet. Besides, this work may enlighten how to efficiently utilize the massive unlabeled PolSAR data to alleviate the greedy demands of CNN-based methods for human annotations.

Jiaming Sun, Linghao Chen, Yiming Xie et al.

In this paper, we propose a novel system named Disp R-CNN for 3D object detection from stereo images. Many recent works solve this problem by first recovering a point cloud with disparity estimation and then apply a 3D detector. The disparity map is computed for the entire image, which is costly and fails to leverage category-specific prior. In contrast, we design an instance disparity estimation network (iDispNet) that predicts disparity only for pixels on objects of interest and learns a category-specific shape prior for more accurate disparity estimation. To address the challenge from scarcity of disparity annotation in training, we propose to use a statistical shape model to generate dense disparity pseudo-ground-truth without the need of LiDAR point clouds, which makes our system more widely applicable. Experiments on the KITTI dataset show that, even when LiDAR ground-truth is not available at training time, Disp R-CNN achieves competitive performance and outperforms previous state-of-the-art methods by 20% in terms of average precision.

Hui Cao, Haikuan Du, Siyu Zhang et al.

In this paper, we present an InSphereNet method for the problem of 3D object classification. Unlike previous methods that use points, voxels, or multi-view images as inputs of deep neural network (DNN), the proposed method constructs a class of more representative features named infilling spheres from signed distance field (SDF). Because of the admirable spatial representation of infilling spheres, we can not only utilize very fewer number of spheres to accomplish classification task, but also design a lightweight InSphereNet with less layers and parameters than previous methods. Experiments on ModelNet40 show that the proposed method leads to superior performance than PointNet and PointNet++ in accuracy. In particular, if there are only a few dozen sphere inputs or about 100000 DNN parameters, the accuracy of our method remains at a very high level (over 88%). This further validates the conciseness and effectiveness of the proposed InSphere 3D representation. Keywords: 3D object classification , signed distance field , deep learning , infilling sphere

Hongwei Dong, Siyu Zhang, Bin Zou et al.

Convolutional neural networks (CNNs) have shown good performance in polarimetric synthetic aperture radar (PolSAR) image classification due to the automation of feature engineering. Excellent hand-crafted architectures of CNNs incorporated the wisdom of human experts, which is an important reason for CNN's success. However, the design of the architectures is a difficult problem, which needs a lot of professional knowledge as well as computational resources. Moreover, the architecture designed by hand might be suboptimal, because it is only one of thousands of unobserved but objective existed paths. Considering that the success of deep learning is largely due to its automation of the feature engineering process, how to design automatic architecture searching methods to replace the hand-crafted ones is an interesting topic. In this paper, we explore the application of neural architecture search (NAS) in PolSAR area for the first time. Different from the utilization of existing NAS methods, we propose a differentiable architecture search (DAS) method which is customized for PolSAR classification. The proposed DAS is equipped with a PolSAR tailored search space and an improved one-shot search strategy. By DAS, the weights parameters and architecture parameters (corresponds to the hyperparameters but not the topologies) can be optimized by stochastic gradient descent method during the training. The optimized architecture parameters should be transformed into corresponding CNN architecture and re-train to achieve high-precision PolSAR classification. In addition, complex-valued DAS is developed to take into account the characteristics of PolSAR images so as to further improve the performance. Experiments on three PolSAR benchmark datasets show that the CNNs obtained by searching have better classification performance than the hand-crafted ones.