Wenyuan Zhang

Yan Wang, Wenjie Luo, Junjie Bai et al. · nvidia

End-to-end architectures trained via imitation learning have advanced autonomous driving by scaling model size and data, yet performance remains brittle in safety-critical long-tail scenarios where supervision is sparse and causal understanding is limited. To address this, we introduce Alpamayo-R1 (AR1), a vision-language-action model (VLA) that integrates Chain of Causation reasoning with trajectory planning to enhance decision-making in complex driving scenarios. Our approach features three key innovations: (1) the Chain of Causation (CoC) dataset, built through a hybrid auto-labeling and human-in-the-loop pipeline producing decision-grounded, causally linked reasoning traces aligned with driving behaviors; (2) a modular VLA architecture combining Cosmos-Reason, a Vision-Language Model pre-trained for Physical AI applications, with a diffusion-based trajectory decoder that generates dynamically feasible plans in real time; (3) a multi-stage training strategy using supervised fine-tuning to elicit reasoning and reinforcement learning (RL) to optimize reasoning quality via large reasoning model feedback and enforce reasoning-action consistency. Evaluation shows AR1 achieves up to a 12% improvement in planning accuracy on challenging cases compared to a trajectory-only baseline, with a 35% reduction in off-road rate and 25% reduction in close encounter rate in closed-loop simulation. RL post-training improves reasoning quality by 45% as measured by a large reasoning model critic and reasoning-action consistency by 37%. Model scaling from 0.5B to 7B parameters shows consistent improvements. On-vehicle road tests confirm real-time performance (99 ms latency) and successful urban deployment. By bridging interpretable reasoning with precise control, AR1 demonstrates a practical path towards Level 4 autonomous driving. We plan to release AR1 models and a subset of the CoC in a future update.

Wenyuan Zhang, Ruofan Xing, Yunfan Zeng et al. · tsinghua

Learning radiance fields has shown remarkable results for novel view synthesis. The learning procedure usually costs lots of time, which motivates the latest methods to speed up the learning procedure by learning without neural networks or using more efficient data structures. However, these specially designed approaches do not work for most of radiance fields based methods. To resolve this issue, we introduce a general strategy to speed up the learning procedure for almost all radiance fields based methods. Our key idea is to reduce the redundancy by shooting much fewer rays in the multi-view volume rendering procedure which is the base for almost all radiance fields based methods. We find that shooting rays at pixels with dramatic color change not only significantly reduces the training burden but also barely affects the accuracy of the learned radiance fields. In addition, we also adaptively subdivide each view into a quadtree according to the average rendering error in each node in the tree, which makes us dynamically shoot more rays in more complex regions with larger rendering error. We evaluate our method with different radiance fields based methods under the widely used benchmarks. Experimental results show that our method achieves comparable accuracy to the state-of-the-art with much faster training.

Chuyang Zhao, Haobo Chen, Wenyuan Zhang et al.

Natural language (NL) based vehicle retrieval aims to search specific vehicle given text description. Different from the image-based vehicle retrieval, NL-based vehicle retrieval requires considering not only vehicle appearance, but also surrounding environment and temporal relations. In this paper, we propose a Symmetric Network with Spatial Relationship Modeling (SSM) method for NL-based vehicle retrieval. Specifically, we design a symmetric network to learn the unified cross-modal representations between text descriptions and vehicle images, where vehicle appearance details and vehicle trajectory global information are preserved. Besides, to make better use of location information, we propose a spatial relationship modeling methods to take surrounding environment and mutual relationship between vehicles into consideration. The qualitative and quantitative experiments verify the effectiveness of the proposed method. We achieve 43.92% MRR accuracy on the test set of the 6th AI City Challenge on natural language-based vehicle retrieval track, yielding the 1st place among all valid submissions on the public leaderboard. The code is available at https://github.com/hbchen121/AICITY2022_Track2_SSM.

Wenyuan Zhang, Chunsheng Wang, Kanle Shi et al.

Unsigned distance functions (UDFs) have been a vital representation for open surfaces. With different differentiable renderers, current methods are able to train neural networks to infer a UDF by minimizing the rendering errors with the UDF to the multi-view ground truth. However, these differentiable renderers are mainly handcrafted, which makes them either biased on ray-surface intersections, or sensitive to unsigned distance outliers, or not scalable to large scenes. To resolve these issues, we present a novel differentiable renderer to infer UDFs more accurately. Instead of using handcrafted equations, our differentiable renderer is a neural network which is pre-trained in a data-driven manner. It learns how to render unsigned distances into depth images, leading to a prior knowledge, dubbed volume rendering priors. To infer a UDF for an unseen scene from multiple RGB images, we generalize the learned volume rendering priors to map inferred unsigned distances in alpha blending for RGB image rendering. To reduce the bias of sampling in UDF inference, we utilize an auxiliary point sampling prior as an indicator of ray-surface intersection, and propose novel schemes towards more accurate and uniform sampling near the zero-level sets. We also propose a new strategy that leverages our pretrained volume rendering prior to serve as a general surface refiner, which can be integrated with various Gaussian reconstruction methods to optimize the Gaussian distributions and refine geometric details. Our results show that the learned volume rendering prior is unbiased, robust, scalable, 3D aware, and more importantly, easy to learn. Further experiments show that the volume rendering prior is also a general strategy to enhance other neural implicit representations such as signed distance function and occupancy.

Yao Chen, Jiawei Sheng, Wenyuan Zhang et al.

The significant computational demands of large language models have increased interest in distilling reasoning abilities into smaller models via Chain-of-Thought (CoT) distillation. Current CoT distillation methods mainly focus on transferring teacher-generated rationales for complex reasoning to student models. However, they do not adequately explore teachers' dynamic attention toward critical information during reasoning. We find that language models exhibit progressive attention shifts towards key information during reasoning, which implies essential clues for drawing conclusions. Building on this observation and analysis, we introduce a novel CoT distillation framework that transfers the teacher's stepwise attention on key information to the student model. This establishes structured guidance for the student's progressive concentration on key information during reasoning. More importantly, we develop a Mixture of Layers module enabling dynamic alignment that adapts to different layers between the teacher and student. Our method achieves consistent performance improvements across multiple mathematical and commonsense reasoning datasets. To our knowledge, it is the first method to leverage stepwise attention within CoT distillation to improve small model reasoning.

Yingkai Tang, Taoyu Su, Wenyuan Zhang et al.

Multi-table entity matching (MEM) addresses the limitations of dual-table approaches by enabling simultaneous identification of equivalent entities across multiple data sources without unique identifiers. However, existing methods relying on pre-trained language models struggle to handle semantic inconsistencies caused by numerical attribute variations. Inspired by the powerful language understanding capabilities of large language models (LLMs), we propose a novel LLM-based framework for multi-table entity matching, termed LLM4MEM. Specifically, we first propose a multi-style prompt-enhanced LLM attribute coordination module to address semantic inconsistencies. Then, to alleviate the matching efficiency problem caused by the surge in the number of entities brought by multiple data sources, we develop a transitive consensus embedding matching module to tackle entity embedding and pre-matching issues. Finally, to address the issue of noisy entities during the matching process, we introduce a density-aware pruning module to optimize the quality of multi-table entity matching. We conducted extensive experiments on 6 MEM datasets, and the results show that our model improves by an average of 5.1% in F1 compared with the baseline model. Our code is available at https://github.com/Ymeki/LLM4MEM.

Shiyao Cui, Zhenyu Zhang, Yilong Chen et al.

The widespread of generative artificial intelligence has heightened concerns about the potential harms posed by AI-generated texts, primarily stemming from factoid, unfair, and toxic content. Previous researchers have invested much effort in assessing the harmlessness of generative language models. However, existing benchmarks are struggling in the era of large language models (LLMs), due to the stronger language generation and instruction following capabilities, as well as wider applications. In this paper, we propose FFT, a new benchmark with 2116 elaborated-designed instances, for LLM harmlessness evaluation with factuality, fairness, and toxicity. To investigate the potential harms of LLMs, we evaluate 9 representative LLMs covering various parameter scales, training stages, and creators. Experiments show that the harmlessness of LLMs is still under-satisfactory, and extensive analysis derives some insightful findings that could inspire future research for harmless LLM research.

Wenyuan Zhang, Shuaiyi Nie, Jiawei Sheng et al.

Large language model (LLM) role-playing has gained widespread attention. Authentic character knowledge is crucial for constructing realistic LLM role-playing agents. However, existing works usually overlook the exploration of LLMs' ability to detect characters' known knowledge errors (KKE) and unknown knowledge errors (UKE) while playing roles, which would lead to low-quality automatic construction of character trainable corpus. In this paper, we propose RoleKE-Bench to evaluate LLMs' ability to detect errors in KKE and UKE. The results indicate that even the latest LLMs struggle to detect these two types of errors effectively, especially when it comes to familiar knowledge. We experimented with various reasoning strategies and propose an agent-based reasoning method, Self-Recollection and Self-Doubt (S$^2$RD), to explore further the potential for improving error detection capabilities. Experiments show that our method effectively improves the LLMs' ability to detect error character knowledge, but it remains an issue that requires ongoing attention.

Junsheng Zhou, Baorui Ma, Wenyuan Zhang et al.

Cross-modality registration between 2D images from cameras and 3D point clouds from LiDARs is a crucial task in computer vision and robotic. Previous methods estimate 2D-3D correspondences by matching point and pixel patterns learned by neural networks, and use Perspective-n-Points (PnP) to estimate rigid transformation during post-processing. However, these methods struggle to map points and pixels to a shared latent space robustly since points and pixels have very different characteristics with patterns learned in different manners (MLP and CNN), and they also fail to construct supervision directly on the transformation since the PnP is non-differentiable, which leads to unstable registration results. To address these problems, we propose to learn a structured cross-modality latent space to represent pixel features and 3D features via a differentiable probabilistic PnP solver. Specifically, we design a triplet network to learn VoxelPoint-to-Pixel matching, where we represent 3D elements using both voxels and points to learn the cross-modality latent space with pixels. We design both the voxel and pixel branch based on CNNs to operate convolutions on voxels/pixels represented in grids, and integrate an additional point branch to regain the information lost during voxelization. We train our framework end-to-end by imposing supervisions directly on the predicted pose distribution with a probabilistic PnP solver. To explore distinctive patterns of cross-modality features, we design a novel loss with adaptive-weighted optimization for cross-modality feature description. The experimental results on KITTI and nuScenes datasets show significant improvements over the state-of-the-art methods. The code and models are available at https://github.com/junshengzhou/VP2P-Match.

Yuqing Li, Wenyuan Zhang, Binbin Li et al.

Conversational aspect-based sentiment quadruple analysis (DiaASQ) aims to extract the quadruple of target-aspect-opinion-sentiment within a dialogue. In DiaASQ, a quadruple's elements often cross multiple utterances. This situation complicates the extraction process, emphasizing the need for an adequate understanding of conversational context and interactions. However, existing work independently encodes each utterance, thereby struggling to capture long-range conversational context and overlooking the deep inter-utterance dependencies. In this work, we propose a novel Dynamic Multi-scale Context Aggregation network (DMCA) to address the challenges. Specifically, we first utilize dialogue structure to generate multi-scale utterance windows for capturing rich contextual information. After that, we design a Dynamic Hierarchical Aggregation module (DHA) to integrate progressive cues between them. In addition, we form a multi-stage loss strategy to improve model performance and generalization ability. Extensive experimental results show that the DMCA model outperforms baselines significantly and achieves state-of-the-art performance.

Minghao Tang, Yongquan He, Yongxiu Xu et al.

Named entity recognition (NER) is a fundamental task in natural language processing that aims to identify and classify named entities in text. However, span-based methods for NER typically assign entity types to text spans, resulting in an imbalanced sample space and neglecting the connections between non-entity and entity spans. To address these issues, we propose a novel approach for NER, named the Boundary Offset Prediction Network (BOPN), which predicts the boundary offsets between candidate spans and their nearest entity spans. By leveraging the guiding semantics of boundary offsets, BOPN establishes connections between non-entity and entity spans, enabling non-entity spans to function as additional positive samples for entity detection. Furthermore, our method integrates entity type and span representations to generate type-aware boundary offsets instead of using entity types as detection targets. We conduct experiments on eight widely-used NER datasets, and the results demonstrate that our proposed BOPN outperforms previous state-of-the-art methods.

Wenyuan Zhang, Xinghua Zhang, Haiyang Yu et al.

Experience intervention in web agents emerges as a promising technical paradigm, enhancing agent interaction capabilities by providing valuable insights from accumulated experiences. However, existing methods predominantly inject experience passively as global context before task execution, struggling to adapt to dynamically changing contextual observations during agent-environment interaction. We propose ExpSeek, which shifts experience toward step-level proactive seeking: (1) estimating step-level entropy thresholds to determine intervention timing using the model's intrinsic signals; (2) designing step-level tailor-designed experience content. Experiments on Qwen3-8B and 32B models across four challenging web agent benchmarks demonstrate that ExpSeek achieves absolute improvements of 9.3% and 7.5%, respectively. Our experiments validate the feasibility and advantages of entropy as a self-triggering signal, reveal that even a 4B small-scale experience model can significantly boost the performance of larger agent models.

Shuaiyi Nie, Siyu Ding, Wenyuan Zhang et al.

Large reasoning models trained with reinforcement learning and verifiable rewards (RLVR) achieve strong performance on complex reasoning tasks, yet often overthink, generating redundant reasoning without performance gains. Existing trajectory-level length penalties often fail to effectively shorten reasoning length and degrade accuracy, as they uniformly treat all reasoning steps and lack fine-grained signals to distinguish redundancy from necessity. Meanwhile, process-supervised methods are typically resource-intensive and suffer from inaccurate credit assignment. To address these issues, we propose ATTNPO, a low-overhead process-supervised RL framework that leverages the model's intrinsic attention signals for step-level credit assignment. We first identify a set of special attention heads that naturally focus on essential steps while suppressing redundant ones. By leveraging the attention scores of these heads, We then employ two sub-strategies to mitigate overthinking by discouraging redundant steps while preserving accuracy by reducing penalties on essential steps. Experimental results show that ATTNPO substantially reduces reasoning length while significantly improving performance across 9 benchmarks.

Minghao Tang, Yongquan He, Yongxiu Xu et al.

Fine-grained entity typing (FET) is an essential task in natural language processing that aims to assign semantic types to entities in text. However, FET poses a major challenge known as the noise labeling problem, whereby current methods rely on estimating noise distribution to identify noisy labels but are confused by diverse noise distribution deviation. To address this limitation, we introduce Co-Prediction Prompt Tuning for noise correction in FET, which leverages multiple prediction results to identify and correct noisy labels. Specifically, we integrate prediction results to recall labeled labels and utilize a differentiated margin to identify inaccurate labels. Moreover, we design an optimization objective concerning divergent co-predictions during fine-tuning, ensuring that the model captures sufficient information and maintains robustness in noise identification. Experimental results on three widely-used FET datasets demonstrate that our noise correction approach significantly enhances the quality of various types of training samples, including those annotated using distant supervision, ChatGPT, and crowdsourcing.

Jimin Tang, Wenyuan Zhang, Junsheng Zhou et al.

Gaussian Splatting has achieved remarkable progress in multi-view surface reconstruction, yet it exhibits notable degradation when only few views are available. Although recent efforts alleviate this issue by enhancing multi-view consistency to produce plausible surfaces, they struggle to infer unseen, occluded, or weakly constrained regions beyond the input coverage. To address this limitation, we present VidSplat, a training-free generative reconstruction framework that leverages powerful video diffusion priors to iteratively synthesize novel views that compensate for missing input coverage, and thereby recover complete 3D scenes from sparse inputs. Specifically, we tackle two key challenges that enable the effective integration of generation and reconstruction. First, for 3D consistent generation, we elaborate a training-free, stage-wise denoising strategy that adaptively guides the denoising direction toward the underlying geometry using the rendered RGB and mask images. Second, to enhance the reconstruction, we develop an iterative mechanism that samples camera trajectories, explores unobserved regions, synthesizes novel views, and supplements training through confidence weighted refinement. VidSplat performs robustly to sparse input and even a single image. Extensive experiments on widely used benchmarks demonstrate our superior performance in sparse-view scene reconstruction.

Liyuan Cui, Wentao Hu, Wenyuan Zhang et al.

Real-time talking avatar generation requires low latency and minute-level temporal stability. Autoregressive (AR) forcing enables streaming inference but suffers from exposure bias, which causes errors to accumulate and become irreversible over long rollouts. In contrast, full-sequence diffusion transformers mitigate drift but remain computationally prohibitive for real-time long-form synthesis. We present AvatarForcing, a one-step streaming diffusion framework that denoises a fixed local-future window with heterogeneous noise levels and emits one clean block per step under constant per-step cost. To stabilize unbounded streams, the method introduces dual-anchor temporal forcing: a style anchor that re-indexes RoPE to maintain a fixed relative position with respect to the active window and applies anchor-audio zero-padding, and a temporal anchor that reuses recently emitted clean blocks to ensure smooth transitions. Real-time one-step inference is enabled by two-stage streaming distillation with offline ODE backfill and distribution matching. Experiments on standard benchmarks and a new 400-video long-form benchmark show strong visual quality and lip synchronization at 34 ms/frame using a 1.3B-parameter student model for realtime streaming. Our page is available at: https://cuiliyuan121.github.io/AvatarForcing/

Yongquan He, Peng Zhang, Luchen Liu et al.

In recent years, temporal knowledge graph (TKG) reasoning has received significant attention. Most existing methods assume that all timestamps and corresponding graphs are available during training, which makes it difficult to predict future events. To address this issue, recent works learn to infer future events based on historical information. However, these methods do not comprehensively consider the latent patterns behind temporal changes, to pass historical information selectively, update representations appropriately and predict events accurately. In this paper, we propose the Historical Information Passing (HIP) network to predict future events. HIP network passes information from temporal, structural and repetitive perspectives, which are used to model the temporal evolution of events, the interactions of events at the same time step, and the known events respectively. In particular, our method considers the updating of relation representations and adopts three scoring functions corresponding to the above dimensions. Experimental results on five benchmark datasets show the superiority of HIP network, and the significant improvements on Hits@1 prove that our method can more accurately predict what is going to happen.

Wenyuan Zhang, Yu-Shen Liu, Zhizhong Han

It is vital to infer a signed distance function (SDF) in multi-view based surface reconstruction. 3D Gaussian splatting (3DGS) provides a novel perspective for volume rendering, and shows advantages in rendering efficiency and quality. Although 3DGS provides a promising neural rendering option, it is still hard to infer SDFs for surface reconstruction with 3DGS due to the discreteness, the sparseness, and the off-surface drift of 3D Gaussians. To resolve these issues, we propose a method that seamlessly merge 3DGS with the learning of neural SDFs. Our key idea is to more effectively constrain the SDF inference with the multi-view consistency. To this end, we dynamically align 3D Gaussians on the zero-level set of the neural SDF using neural pulling, and then render the aligned 3D Gaussians through the differentiable rasterization. Meanwhile, we update the neural SDF by pulling neighboring space to the pulled 3D Gaussians, which progressively refine the signed distance field near the surface. With both differentiable pulling and splatting, we jointly optimize 3D Gaussians and the neural SDF with both RGB and geometry constraints, which recovers more accurate, smooth, and complete surfaces with more geometry details. Our numerical and visual comparisons show our superiority over the state-of-the-art results on the widely used benchmarks.

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.

Wenyuan Zhang, Shuaiyi Nie, Xinghua Zhang et al.

We introduce S1-Bench, a novel benchmark designed to evaluate the performance of Large Reasoning Models (LRMs) on simple tasks that favor intuitive system 1 thinking rather than deliberative system 2 reasoning. While LRMs have achieved significant breakthroughs in complex reasoning tasks through explicit chains of thought, their heavy reliance on system 2 thinking may limit their system 1 thinking capabilities. However, there is a lack of an appropriate benchmark for evaluating LRM's system 1 thinking capabilities. To fill this gap, S1-Bench introduces a suite of simple, diverse, and natural questions across multiple domains and languages, specifically designed to assess LRMs' performance on questions more suitable for system 1 . We conduct extensive evaluations across 28 LRMs, revealing their inefficiency, inadequate accuracy, and limited robustness when handling simple questions. Additionally, we observe a gap between their difficulty perception and generation length. Overall, this work paves the way toward dual-system compatibility in the development of LRMs.

Xiaodong Li, Jiawei Sheng, Jiangxia Cao et al.

Cross-domain recommendation (CDR) has demonstrated to be an effective solution for alleviating the user cold-start issue. By leveraging rich user-item interactions available in a richly informative source domain, CDR could improve the recommendation performance for cold-start users in the target domain. Previous CDR approaches mostly adhere the Embedding and Mapping (EMCDR) paradigm, which learns a user-shared mapping function to transfer users' preference from the source domain to the target domain, neglecting users' personalized preference. Recent CDR approaches further leverage the meta-learning paradigm, considering the CDR task for each user independently and learning user-specific mapping functions for each user. However, they mostly learn representations for each user individually, which ignores the common preference between different users, neglecting valuable information for CDR. In addition, all these approaches usually summarize the user's preference into an overall representation, which can hardly capture the user's multi-interest preference. To this end, we propose a personalized multi-interest modeling framework for CDR to cold-start users, termed as NF-NPCDR. Specifically, we propose a personalized preference encoder that enhances the neural process (NP) with the normalizing flow (NF) to convert the Gaussian (unimodal) distribution to a multimodal distribution, providing a novel way to capture the user's personalized multi-interest preference. Then, we propose a common preference encoder with a preference pool to capture the common preference between different users. Furthermore, we introduce a stochastic adaptive decoder to incorporate both the personalized and common preference for cold-start users, adaptively modulating both preference for better recommendation.

Jian Liu, Sipeng Zhang, Chuixin Kong et al.

This technical report presents our solution, "occTransformer" for the 3D occupancy prediction track in the autonomous driving challenge at CVPR 2023. Our method builds upon the strong baseline BEVFormer and improves its performance through several simple yet effective techniques. Firstly, we employed data augmentation to increase the diversity of the training data and improve the model's generalization ability. Secondly, we used a strong image backbone to extract more informative features from the input data. Thirdly, we incorporated a 3D unet head to better capture the spatial information of the scene. Fourthly, we added more loss functions to better optimize the model. Additionally, we used an ensemble approach with the occ model BevDet and SurroundOcc to further improve the performance. Most importantly, we integrated 3D detection model StreamPETR to enhance the model's ability to detect objects in the scene. Using these methods, our solution achieved 49.23 miou on the 3D occupancy prediction track in the autonomous driving challenge.

Wenyuan Zhang, Xinghua Zhang, Shiyao Cui et al.

Aspect sentiment quad prediction (ASQP) aims to predict the quad sentiment elements for a given sentence, which is a critical task in the field of aspect-based sentiment analysis. However, the data imbalance issue has not received sufficient attention in ASQP task. In this paper, we divide the issue into two-folds, quad-pattern imbalance and aspect-category imbalance, and propose an Adaptive Data Augmentation (ADA) framework to tackle the imbalance issue. Specifically, a data augmentation process with a condition function adaptively enhances the tail quad patterns and aspect categories, alleviating the data imbalance in ASQP. Following previous studies, we also further explore the generative framework for extracting complete quads by introducing the category prior knowledge and syntax-guided decoding target. Experimental results demonstrate that data augmentation for imbalance in ASQP task can improve the performance, and the proposed ADA method is superior to naive data oversampling.

Wenyuan Zhang, Yixiao Yang, Han Huang et al.

Monocular depth priors have been widely adopted by neural rendering in multi-view based tasks such as 3D reconstruction and novel view synthesis. However, due to the inconsistent prediction on each view, how to more effectively leverage monocular cues in a multi-view context remains a challenge. Current methods treat the entire estimated depth map indiscriminately, and use it as ground truth supervision, while ignoring the inherent inaccuracy and cross-view inconsistency in monocular priors. To resolve these issues, we propose MonoInstance, a general approach that explores the uncertainty of monocular depths to provide enhanced geometric priors for neural rendering and reconstruction. Our key insight lies in aligning each segmented instance depths from multiple views within a common 3D space, thereby casting the uncertainty estimation of monocular depths into a density measure within noisy point clouds. For high-uncertainty areas where depth priors are unreliable, we further introduce a constraint term that encourages the projected instances to align with corresponding instance masks on nearby views. MonoInstance is a versatile strategy which can be seamlessly integrated into various multi-view neural rendering frameworks. Our experimental results demonstrate that MonoInstance significantly improves the performance in both reconstruction and novel view synthesis under various benchmarks.

Wenyuan Zhang, Emily Yue-ting Jia, Junsheng Zhou et al.

Recently, it has shown that priors are vital for neural implicit functions to reconstruct high-quality surfaces from multi-view RGB images. However, current priors require large-scale pre-training, and merely provide geometric clues without considering the importance of color. In this paper, we present NeRFPrior, which adopts a neural radiance field as a prior to learn signed distance fields using volume rendering for surface reconstruction. Our NeRF prior can provide both geometric and color clues, and also get trained fast under the same scene without additional data. Based on the NeRF prior, we are enabled to learn a signed distance function (SDF) by explicitly imposing a multi-view consistency constraint on each ray intersection for surface inference. Specifically, at each ray intersection, we use the density in the prior as a coarse geometry estimation, while using the color near the surface as a clue to check its visibility from another view angle. For the textureless areas where the multi-view consistency constraint does not work well, we further introduce a depth consistency loss with confidence weights to infer the SDF. Our experimental results outperform the state-of-the-art methods under the widely used benchmarks.

Wenyuan Zhang, Tianyun Liu, Mengxiao Song et al.

Despite the abundance of prior social strategies possessed by humans, there remains a paucity of research dedicated to their transfer and integration into social agents. Our proposed SOTOPIA-$Ω$ framework aims to address and bridge this gap, with a particular focus on enhancing the social capabilities of language agents. This framework dynamically injects multi-step reasoning strategies inspired by negotiation theory and two simple direct strategies into expert agents, thereby automating the construction of a high-quality social dialogue training corpus. Additionally, we introduce the concept of Social Instruction Following (S-IF) and propose two new S-IF evaluation metrics that complement social capability. We demonstrate that several 7B models trained on high-quality corpus not only significantly surpass the expert agent (GPT-4) in achieving social goals but also enhance S-IF performance. Analysis and variant experiments validate the advantages of dynamic construction, which can especially break the agent's prolonged deadlock.

Yongquan He, Wenyuan Zhang, Xuancheng Huang et al.

Instruction tuning for large language models (LLMs) can drive them to produce results consistent with human goals in specific downstream tasks. However, the process of continual instruction tuning (CIT) for LLMs may bring about the catastrophic forgetting (CF) problem, where previously learned abilities are degraded. Recent methods try to alleviate the CF problem by modifying models or replaying data, which may only remember the surface-level pattern of instructions and get confused on held-out tasks. In this paper, we propose a novel continual instruction tuning method based on Key-part Information Gain (KPIG). Our method computes the information gain on masked parts to dynamically replay data and refine the training objective, which enables LLMs to capture task-aware information relevant to the correct response and alleviate overfitting to general descriptions in instructions. In addition, we propose two metrics, P-score and V-score, to measure the generalization and instruction-following abilities of LLMs. Experiments demonstrate our method achieves superior performance on both seen and held-out tasks.

Juwei Yue, Chuanrui Hu, Jiawei Sheng et al.

Long-term memory is essential for conversational agents to maintain coherence, track persistent tasks, and provide personalized interactions across extended dialogues. However, existing approaches as Retrieval-Augmented Generation (RAG) and graph-based memory mostly rely on pairwise relations, which can hardly capture high-order associations, i.e., joint dependencies among multiple elements, causing fragmented retrieval. To this end, we propose HyperMem, a hypergraph-based hierarchical memory architecture that explicitly models such associations using hyperedges. Particularly, HyperMem structures memory into three levels: topics, episodes, and facts, and groups related episodes and their facts via hyperedges, unifying scattered content into coherent units. Leveraging this structure, we design a hybrid lexical-semantic index and a coarse-to-fine retrieval strategy, supporting accurate and efficient retrieval of high-order associations. Experiments on the LoCoMo benchmark show that HyperMem achieves state-of-the-art performance with 92.73% LLM-as-a-judge accuracy, demonstrating the effectiveness of HyperMem for long-term conversations.

Yulun Wu, Han Huang, Wenyuan Zhang et al.

In recent years, reconstructing indoor scene geometry from multi-view images has achieved encouraging accomplishments. Current methods incorporate monocular priors into neural implicit surface models to achieve high-quality reconstructions. However, these methods require hundreds of images for scene reconstruction. When only a limited number of views are available as input, the performance of monocular priors deteriorates due to scale ambiguity, leading to the collapse of the reconstructed scene geometry. In this paper, we propose a new method, named Sparis, for indoor surface reconstruction from sparse views. Specifically, we investigate the impact of monocular priors on sparse scene reconstruction, introducing a novel prior based on inter-image matching information. Our prior offers more accurate depth information while ensuring cross-view matching consistency. Additionally, we employ an angular filter strategy and an epipolar matching weight function, aiming to reduce errors due to view matching inaccuracies, thereby refining the inter-image prior for improved reconstruction accuracy. The experiments conducted on widely used benchmarks demonstrate superior performance in sparse-view scene reconstruction.

Weiqi Zhang, Junsheng Zhou, Haotian Geng et al.

3D Gaussian Splatting (3DGS) has demonstrated its advantages in achieving fast and high-quality rendering. As point clouds serve as a widely-used and easily accessible form of 3D representation, bridging the gap between point clouds and Gaussians becomes increasingly important. Recent studies have explored how to convert the colored points into Gaussians, but directly generating Gaussians from colorless 3D point clouds remains an unsolved challenge. In this paper, we propose GAP, a novel approach that gaussianizes raw point clouds into high-fidelity 3D Gaussians with text guidance. Our key idea is to design a multi-view optimization framework that leverages a depth-aware image diffusion model to synthesize consistent appearances across different viewpoints. To ensure geometric accuracy, we introduce a surface-anchoring mechanism that effectively constrains Gaussians to lie on the surfaces of 3D shapes during optimization. Furthermore, GAP incorporates a diffuse-based inpainting strategy that specifically targets at completing hard-to-observe regions. We evaluate GAP on the Point-to-Gaussian generation task across varying complexity levels, from synthetic point clouds to challenging real-world scans, and even large-scale scenes. Project Page: https://weiqi-zhang.github.io/GAP.

Junsheng Zhou, Zhifan Yang, Liang Han et al.

This paper tackles the challenge of recovering 4D dynamic scenes from videos captured by as few as four portable cameras. Learning to model scene dynamics for temporally consistent novel-view rendering is a foundational task in computer graphics, where previous works often require dense multi-view captures using camera arrays of dozens or even hundreds of views. We propose \textbf{4C4D}, a novel framework that enables high-fidelity 4D Gaussian Splatting from video captures of extremely sparse cameras. Our key insight lies that the geometric learning under sparse settings is substantially more difficult than modeling appearance. Driven by this observation, we introduce a Neural Decaying Function on Gaussian opacities for enhancing the geometric modeling capability of 4D Gaussians. This design mitigates the inherent imbalance between geometry and appearance modeling in 4DGS by encouraging the 4DGS gradients to focus more on geometric learning. Extensive experiments across sparse-view datasets with varying camera overlaps show that 4C4D achieves superior performance over prior art. Project page at: https://junshengzhou.github.io/4C4D.

Ming Chen, Liyuan Cui, Wenyuan Zhang et al.

Recently, interactive digital human video generation has attracted widespread attention and achieved remarkable progress. However, building such a practical system that can interact with diverse input signals in real time remains challenging to existing methods, which often struggle with heavy computational cost and limited controllability. In this work, we introduce an autoregressive video generation framework that enables interactive multimodal control and low-latency extrapolation in a streaming manner. With minimal modifications to a standard large language model (LLM), our framework accepts multimodal condition encodings including audio, pose, and text, and outputs spatially and semantically coherent representations to guide the denoising process of a diffusion head. To support this, we construct a large-scale dialogue dataset of approximately 20,000 hours from multiple sources, providing rich conversational scenarios for training. We further introduce a deep compression autoencoder with up to 64$\times$ reduction ratio, which effectively alleviates the long-horizon inference burden of the autoregressive model. Extensive experiments on duplex conversation, multilingual human synthesis, and interactive world model highlight the advantages of our approach in low latency, high efficiency, and fine-grained multimodal controllability.

Wenyuan Zhang, Jimin Tang, Weiqi Zhang et al.

Modeling reflections from 2D images is essential for photorealistic rendering and novel view synthesis. Recent approaches enhance Gaussian primitives with reflection-related material attributes to enable physically based rendering (PBR) with Gaussian Splatting. However, the material inference often lacks sufficient constraints, especially under limited environment modeling, resulting in illumination aliasing and reduced generalization. In this work, we revisit the problem from a multi-view perspective and show that multi-view consistent material inference with more physically-based environment modeling is key to learning accurate reflections with Gaussian Splatting. To this end, we enforce 2D Gaussians to produce multi-view consistent material maps during deferred shading. We also track photometric variations across views to identify highly reflective regions, which serve as strong priors for reflection strength terms. To handle indirect illumination caused by inter-object occlusions, we further introduce an environment modeling strategy through ray tracing with 2DGS, enabling photorealistic rendering of indirect radiance. Experiments on widely used benchmarks show that our method faithfully recovers both illumination and geometry, achieving state-of-the-art rendering quality in novel views synthesis.

Yikang Ding, Jiwen Liu, Wenyuan Zhang et al.

Recent advances in audio-driven avatar video generation have significantly enhanced audio-visual realism. However, existing methods treat instruction conditioning merely as low-level tracking driven by acoustic or visual cues, without modeling the communicative purpose conveyed by the instructions. This limitation compromises their narrative coherence and character expressiveness. To bridge this gap, we introduce Kling-Avatar, a novel cascaded framework that unifies multimodal instruction understanding with photorealistic portrait generation. Our approach adopts a two-stage pipeline. In the first stage, we design a multimodal large language model (MLLM) director that produces a blueprint video conditioned on diverse instruction signals, thereby governing high-level semantics such as character motion and emotions. In the second stage, guided by blueprint keyframes, we generate multiple sub-clips in parallel using a first-last frame strategy. This global-to-local framework preserves fine-grained details while faithfully encoding the high-level intent behind multimodal instructions. Our parallel architecture also enables fast and stable generation of long-duration videos, making it suitable for real-world applications such as digital human livestreaming and vlogging. To comprehensively evaluate our method, we construct a benchmark of 375 curated samples covering diverse instructions and challenging scenarios. Extensive experiments demonstrate that Kling-Avatar is capable of generating vivid, fluent, long-duration videos at up to 1080p and 48 fps, achieving superior performance in lip synchronization accuracy, emotion and dynamic expressiveness, instruction controllability, identity preservation, and cross-domain generalization. These results establish Kling-Avatar as a new benchmark for semantically grounded, high-fidelity audio-driven avatar synthesis.

Chuanyu Tang, Yilong Chen, Zhenyu Zhang et al.

Low-Rank Adaptation (LoRA) drives research to align its performance with full fine-tuning. However, significant challenges remain: (1) Simply increasing the rank size of LoRA does not effectively capture high-rank information, which leads to a performance bottleneck.(2) MoE-style LoRA methods substantially increase parameters and inference latency, contradicting the goals of efficient fine-tuning and ease of application. To address these challenges, we introduce Mixture of Ranks (MoR), which learns rank-specific information for different tasks based on input and efficiently integrates multi-rank information. We firstly propose a new framework that equates the integration of multiple LoRAs to expanding the rank of LoRA. Moreover, we hypothesize that low-rank LoRA already captures sufficient intrinsic information, and MoR can derive high-rank information through mathematical transformations of the low-rank components. Thus, MoR can reduces the learning difficulty of LoRA and enhances its multi-task capabilities. MoR achieves impressive results, with MoR delivering a 1.31\% performance improvement while using only 93.93\% of the parameters compared to baseline methods.

Yu Zhang, Rui Yu, Zhipeng Yao et al.

The Mean Square Error (MSE) is commonly utilized to estimate the solution of the optimal value function in the vast majority of offline reinforcement learning (RL) models and has achieved outstanding performance. However, we find that its principle can lead to overestimation phenomenon for the value function. In this paper, we first theoretically analyze overestimation phenomenon led by MSE and provide the theoretical upper bound of the overestimated error. Furthermore, to address it, we propose a novel Bellman underestimated operator to counteract overestimation phenomenon and then prove its contraction characteristics. At last, we propose the offline RL algorithm based on underestimated operator and diffusion policy model. Extensive experimental results on D4RL tasks show that our method can outperform state-of-the-art offline RL algorithms, which demonstrates that our theoretical analysis and underestimation way are effective for offline RL tasks.

Zefeng Zhang, Jiawei Sheng, Chuang Zhang et al.

Multimodal Entity Linking (MEL) aims to link ambiguous mentions in multimodal contexts to entities in a multimodal knowledge graph. A pivotal challenge is to fully leverage multi-element correlations between mentions and entities to bridge modality gap and enable fine-grained semantic matching. Existing methods attempt several local correlative mechanisms, relying heavily on the automatically learned attention weights, which may over-concentrate on partial correlations. To mitigate this issue, we formulate the correlation assignment problem as an optimal transport (OT) problem, and propose a novel MEL framework, namely OT-MEL, with OT-guided correlation assignment. Thereby, we exploit the correlation between multimodal features to enhance multimodal fusion, and the correlation between mentions and entities to enhance fine-grained matching. To accelerate model prediction, we further leverage knowledge distillation to transfer OT assignment knowledge to attention mechanism. Experimental results show that our model significantly outperforms previous state-of-the-art baselines and confirm the effectiveness of the OT-guided correlation assignment.

Zhixin Li, Wenyuan Zhang, Jie Shan

Building models are conventionally reconstructed by building roof points planar segmentation and then using a topology graph to group the planes together. Roof edges and vertices are then mathematically represented by intersecting segmented planes. Technically, such solution is based on sequential local fitting, i.e., the entire data of one building are not simultaneously participating in determining the building model. As a consequence, the solution is lack of topological integrity and geometric rigor. Fundamentally different from this traditional approach, we propose a holistic parametric reconstruction method which means taking into consideration the entire point clouds of one building simultaneously. In our work, building models are reconstructed from predefined parametric (roof) primitives. We first use a well-designed deep neural network to segment and identify primitives in the given building point clouds. A holistic optimization strategy is then introduced to simultaneously determine the parameters of a segmented primitive. In the last step, the optimal parameters are used to generate a watertight building model in CityGML format. The airborne LiDAR dataset RoofN3D with predefined roof types is used for our test. It is shown that PointNet++ applied to the entire dataset can achieve an accuracy of 83% for primitive classification. For a subset of 910 buildings in RoofN3D, the holistic approach is then used to determine the parameters of primitives and reconstruct the buildings. The achieved overall quality of reconstruction is 0.08 meters for point-surface-distance or 0.7 times RMSE of the input LiDAR points. The study demonstrates the efficiency and capability of the proposed approach and its potential to handle large scale urban point clouds.

Jiahua Xu, Ziyuan Luo, Wei Zhou et al.

Objective quality assessment of stereoscopic panoramic images becomes a challenging problem owing to the rapid growth of 360-degree contents. Different from traditional 2D image quality assessment (IQA), more complex aspects are involved in 3D omnidirectional IQA, especially unlimited field of view (FoV) and extra depth perception, which brings difficulty to evaluate the quality of experience (QoE) of 3D omnidirectional images. In this paper, we propose a multi-viewport based fullreference stereo 360 IQA model. Due to the freely changeable viewports when browsing in the head-mounted display (HMD), our proposed approach processes the image inside FoV rather than the projected one such as equirectangular projection (ERP). In addition, since overall QoE depends on both image quality and depth perception, we utilize the features estimated by the difference map between left and right views which can reflect disparity. The depth perception features along with binocular image qualities are employed to further predict the overall QoE of 3D 360 images. The experimental results on our public Stereoscopic OmnidirectionaL Image quality assessment Database (SOLID) show that the proposed method achieves a significant improvement over some well-known IQA metrics and can accurately reflect the overall QoE of perceived images.