Bowen Zheng

Sicong Wang, Ruiting Dong, Yue Liu et al.

Real-world user behavior rarely consists of isolated actions; instead, it often forms intent flows governed by spatiotemporal dependencies. To provide integrated service recommendations, we focus on the task of Generative Spatiotemporal Intent Sequence Recommendation (GSISR), which aims to generate intent sequences that are logically coherent and physically executable within complex spatiotemporal contexts. While LLMs offer strong reasoning potential for GSISR, direct industrial deployment is limited by high inference latency and context-mismatched or physically infeasible plans. To address these challenges, we propose a generative framework, GPlan, that internalizes LLM reasoning into lightweight models through two components. First, to enable reasoning under strict latency constraints, we introduce Progressive Implicit CoT Distillation, which compresses explicit reasoning processes into reserved latent tokens, allowing small models to inherit complex planning logic without generating long reasoning text. Second, to address the disconnect between general knowledge and real-world constraints, we design Spatiotemporal Counterfactual DPO. By aligning the model with counterfactual context-plan pairs, we improve sensitivity to spatiotemporal context and reduce context-mismatched plans. Offline experiments and online A/B testing demonstrate that our approach improves sequence coherence and context responsiveness. Our implementation and the anonymized GSISR dataset are available at https://github.com/alibaba/GPlan.

Tianyi Tang, Yiwen Hu, Bingqian Li et al.

To facilitate the research on large language models (LLMs), this paper presents a comprehensive and unified library, LLMBox, to ease the development, use, and evaluation of LLMs. This library is featured with three main merits: (1) a unified data interface that supports the flexible implementation of various training strategies, (2) a comprehensive evaluation that covers extensive tasks, datasets, and models, and (3) more practical consideration, especially on user-friendliness and efficiency. With our library, users can easily reproduce existing methods, train new models, and conduct comprehensive performance comparisons. To rigorously test LLMBox, we conduct extensive experiments in a diverse coverage of evaluation settings, and experimental results demonstrate the effectiveness and efficiency of our library in supporting various implementations related to LLMs. The detailed introduction and usage guidance can be found at https://github.com/RUCAIBox/LLMBox.

Bowen Zheng, Da-Wei Zhou, Han-Jia Ye et al.

Learning new classes without forgetting is crucial for real-world applications for a classification model. Vision Transformers (ViT) recently achieve remarkable performance in Class Incremental Learning (CIL). Previous works mainly focus on block design and model expansion for ViTs. However, in this paper, we find that when the ViT is incrementally trained, the attention layers gradually lose concentration on local features. We call this interesting phenomenon as \emph{Locality Degradation} in ViTs for CIL. Since the low-level local information is crucial to the transferability of the representation, it is beneficial to preserve the locality in attention layers. In this paper, we encourage the model to preserve more local information as the training procedure goes on and devise a Locality-Preserved Attention (LPA) layer to emphasize the importance of local features. Specifically, we incorporate the local information directly into the vanilla attention and control the initial gradients of the vanilla attention by weighting it with a small initial value. Extensive experiments show that the representations facilitated by LPA capture more low-level general information which is easier to transfer to follow-up tasks. The improved model gets consistently better performance on CIFAR100 and ImageNet100.

Bowen Zheng, Katayun Kamdin, David Shapiro et al.

Iterative ptychographic reconstruction algorithms are widely used for coherent diffractive imaging but can exhibit slow convergence under realistic experimental conditions. We propose a machine learning-augmented approach that accelerates iterative ptychographic reconstruction by introducing a learned fast-forward operator applied during reconstruction. Following an initial warm-up using standard iterations, the fast-forward operator advances the reconstruction toward a more converged state, after which conventional iterative updates are resumed. This strategy preserves the physical consistency and flexibility of established ptychographic solvers while reducing the number of iterations required for convergence. The model is trained on diverse ptychographic datasets and evaluated on experimental data acquired in a different year, demonstrating robustness and temporal generalization. Compared with conventional iterative solvers, the machine learning-augmented method achieves comparable reconstruction quality while converging faster in terms of Poisson negative log-likelihood, yielding over a two-fold reduction in wall-clock time. The approach has been integrated into an existing reconstruction pipeline and deployed in production at a synchrotron beamline, demonstrating practicality for real-time experimental operation.

Xiaoyu Pan, Bowen Zheng, Xinwei Jiang et al.

Optical motion capture (MoCap) is the "gold standard" for accurately capturing full-body motions. To make use of raw MoCap point data, the system labels the points with corresponding body part locations and solves the full-body motions. However, MoCap data often contains mislabeling, occlusion and positional errors, requiring extensive manual correction. To alleviate this burden, we introduce RoMo, a learning-based framework for robustly labeling and solving raw optical motion capture data. In the labeling stage, RoMo employs a divide-and-conquer strategy to break down the complex full-body labeling challenge into manageable subtasks: alignment, full-body segmentation and part-specific labeling. To utilize the temporal continuity of markers, RoMo generates marker tracklets using a K-partite graph-based clustering algorithm, where markers serve as nodes, and edges are formed based on positional and feature similarities. For motion solving, to prevent error accumulation along the kinematic chain, we introduce a hybrid inverse kinematic solver that utilizes joint positions as intermediate representations and adjusts the template skeleton to match estimated joint positions. We demonstrate that RoMo achieves high labeling and solving accuracy across multiple metrics and various datasets. Extensive comparisons show that our method outperforms state-of-the-art research methods. On a real dataset, RoMo improves the F1 score of hand labeling from 0.94 to 0.98, and reduces joint position error of body motion solving by 25%. Furthermore, RoMo can be applied in scenarios where commercial systems are inadequate. The code and data for RoMo are available at https://github.com/non-void/RoMo.

Ruochen Jiao, Xiangguo Liu, Bowen Zheng et al.

Trajectory generation and prediction are two interwoven tasks that play important roles in planner evaluation and decision making for intelligent vehicles. Most existing methods focus on one of the two and are optimized to directly output the final generated/predicted trajectories, which only contain limited information for critical scenario augmentation and safe planning. In this work, we propose a novel behavior-aware Trajectory Autoencoder (TAE) that explicitly models drivers' behavior such as aggressiveness and intention in the latent space, using semi-supervised adversarial autoencoder and domain knowledge in transportation. Our model addresses trajectory generation and prediction in a unified architecture and benefits both tasks: the model can generate diverse, controllable and realistic trajectories to enhance planner optimization in safety-critical and long-tailed scenarios, and it can provide prediction of critical behavior in addition to the final trajectories for decision making. Experimental results demonstrate that our method achieves promising performance on both trajectory generation and prediction.

Bingqian Li, Bowen Zheng, Xiaolei Wang et al.

Large language models (LLMs) have shown great promise in recommender systems, where supervised fine-tuning (SFT) is commonly used for adaptation. Subsequent studies further introduce preference learning to incorporate negative samples into the training process. However, existing methods rely on sequence-level, offline-generated negatives, making them less discriminative and informative when adapting LLMs to recommendation tasks with large negative item spaces. To address these challenges, we propose ILRec, a novel preference fine-tuning framework for LLM-based recommendation, leveraging self-hard negative signals extracted from intermediate layers to improve preference learning. Specifically, we identify self-hard negative tokens from intermediate layers as fine-grained negative supervision that dynamically reflects the model's preference learning process. To effectively integrate these signals into training, we design a two-stage framework comprising cross-layer preference optimization and cross-layer preference distillation, enabling the model to jointly discriminate informative negatives and enhance the quality of negative signals from intermediate layers. In addition, we introduce a lightweight collaborative filtering model to assign token-level rewards for negative signals, mitigating the risk of over-penalizing false negatives. Extensive experiments on three datasets demonstrate ILRec's effectiveness in enhancing the performance of LLM-based recommender systems.

Bowen Zheng, Chenxi Huang, Yuemei Luo

Three-dimensional (3D) reconstruction of head Computed Tomography (CT) images elucidates the intricate spatial relationships of tissue structures, thereby assisting in accurate diagnosis. Nonetheless, securing an optimal head CT scan without deviation is challenging in clinical settings, owing to poor positioning by technicians, patient's physical constraints, or CT scanner tilt angle restrictions. Manual formatting and reconstruction not only introduce subjectivity but also strain time and labor resources. To address these issues, we propose an efficient automatic head CT images 3D reconstruction method, improving accuracy and repeatability, as well as diminishing manual intervention. Our approach employs a deep learning-based object detection algorithm, identifying and evaluating orbitomeatal line landmarks to automatically reformat the images prior to reconstruction. Given the dearth of existing evaluations of object detection algorithms in the context of head CT images, we compared ten methods from both theoretical and experimental perspectives. By exploring their precision, efficiency, and robustness, we singled out the lightweight YOLOv8 as the aptest algorithm for our task, with an mAP of 92.77% and impressive robustness against class imbalance. Our qualitative evaluation of standardized reconstruction results demonstrates the clinical practicability and validity of our method.

Junkai Liu, Yujie Tong, Hui Huang et al.

Legal judgment prediction (LJP), which enables litigants and their lawyers to forecast judgment outcomes and refine litigation strategies, has emerged as a crucial legal NLP task. Existing studies typically utilize legal facts, i.e., facts that have been established by evidence and determined by the judge, to predict the judgment. However, legal facts are often difficult to obtain in the early stages of litigation, significantly limiting the practical applicability of fact-based LJP. To address this limitation, we propose a novel legal NLP task: legal fact prediction (LFP), which takes the evidence submitted by litigants for trial as input to predict legal facts, thereby empowering fact-based LJP technologies to make predictions in the absence of ground-truth legal facts. We also propose the first benchmark dataset, LFPBench, for evaluating the LFP task. Our extensive experiments on LFPBench demonstrate the effectiveness of LFP-empowered LJP and highlight promising research directions for LFP.

Chengjie Jiang, Xiaowen Liu, Bowen Zheng et al.

Infrared and visible image fusion has been developed from vision perception oriented fusion methods to strategies which both consider the vision perception and high-level vision task. However, the existing task-driven methods fail to address the domain gap between semantic and geometric representation. To overcome these issues, we propose a high-level vision task-driven infrared and visible image fusion network via semantic and geometric domain transformation, terms as HSFusion. Specifically, to minimize the gap between semantic and geometric representation, we design two separate domain transformation branches by CycleGAN framework, and each includes two processes: the forward segmentation process and the reverse reconstruction process. CycleGAN is capable of learning domain transformation patterns, and the reconstruction process of CycleGAN is conducted under the constraint of these patterns. Thus, our method can significantly facilitate the integration of semantic and geometric information and further reduces the domain gap. In fusion stage, we integrate the infrared and visible features that extracted from the reconstruction process of two seperate CycleGANs to obtain the fused result. These features, containing varying proportions of semantic and geometric information, can significantly enhance the high level vision tasks. Additionally, we generate masks based on segmentation results to guide the fusion task. These masks can provide semantic priors, and we design adaptive weights for two distinct areas in the masks to facilitate image fusion. Finally, we conducted comparative experiments between our method and eleven other state-of-the-art methods, demonstrating that our approach surpasses others in both visual appeal and semantic segmentation task.

Bowen Zheng, Da-Wei Zhou, Han-Jia Ye et al.

The ability to learn new concepts while preserve the learned knowledge is desirable for learning systems in Class-Incremental Learning (CIL). Recently, feature expansion of the model become a prevalent solution for CIL, where the old features are fixed during the training of the new task while new features are expanded for the new tasks. However, such task-specific features learned from the new task may collide with the old features, leading to misclassification between tasks. Therefore, the expanded model is often encouraged to capture diverse features from the new task, aiming to avoid such collision. However, the existing solution is largely restricted to the samples from the current task, because of the poor accessibility to previous samples. To promote the learning and transferring of diverse features across tasks, we propose a framework called Task-Agnostic Guided Feature Expansion (TagFex). Firstly, it captures task-agnostic features continually with a separate model, providing extra task-agnostic features for subsequent tasks. Secondly, to obtain useful features from the task-agnostic model for the current task, it aggregates the task-agnostic features with the task-specific feature using a merge attention. Then the aggregated feature is transferred back into the task-specific feature for inference, helping the task-specific model capture diverse features. Extensive experiments show the effectiveness and superiority of TagFex on various CIL settings. Code is available at https://github.com/bwnzheng/TagFex_CVPR2025.

Bowen Zheng, Ran Cheng

In the history of knowledge distillation, the focus has once shifted over time from logit-based to feature-based approaches. However, this transition has been revisited with the advent of Decoupled Knowledge Distillation (DKD), which re-emphasizes the importance of logit knowledge through advanced decoupling and weighting strategies. While DKD marks a significant advancement, its underlying mechanisms merit deeper exploration. As a response, we rethink DKD from a predictive distribution perspective. First, we introduce an enhanced version, the Generalized Decoupled Knowledge Distillation (GDKD) loss, which offers a more versatile method for decoupling logits. Then we pay particular attention to the teacher model's predictive distribution and its impact on the gradients of GDKD loss, uncovering two critical insights often overlooked: (1) the partitioning by the top logit considerably improves the interrelationship of non-top logits, and (2) amplifying the focus on the distillation loss of non-top logits enhances the knowledge extraction among them. Utilizing these insights, we further propose a streamlined GDKD algorithm with an efficient partition strategy to handle the multimodality of teacher models' predictive distribution. Our comprehensive experiments conducted on a variety of benchmarks, including CIFAR-100, ImageNet, Tiny-ImageNet, CUB-200-2011, and Cityscapes, demonstrate GDKD's superior performance over both the original DKD and other leading knowledge distillation methods. The code is available at https://github.com/ZaberKo/GDKD.

Bowen Zheng, Yongli Xiang, Ziming Hong et al.

Image-to-Video (I2V) generation models, which condition video generation on reference images, have shown emerging visual instruction-following capability, allowing certain visual cues in reference images to act as implicit control signals for video generation. However, this capability also introduces a previously overlooked risk: adversaries may exploit visual instructions to inject malicious intent through the image modality. In this work, we uncover this risk by proposing Visual Instruction Injection (VII), a training-free and transferable jailbreaking framework that intentionally disguises the malicious intent of unsafe text prompts as benign visual instructions in the safe reference image. Specifically, VII coordinates a Malicious Intent Reprogramming module to distill malicious intent from unsafe text prompts while minimizing their static harmfulness, and a Visual Instruction Grounding module to ground the distilled intent onto a safe input image by rendering visual instructions that preserve semantic consistency with the original unsafe text prompt, thereby inducing harmful content during I2V generation. Empirically, our extensive experiments on four state-of-the-art commercial I2V models (Kling-v2.5-turbo, Gemini Veo-3.1, Seedance-1.5-pro, and PixVerse-V5) demonstrate that VII achieves Attack Success Rates of up to 83.5% while reducing Refusal Rates to near zero, significantly outperforming existing baselines.

Kun Luo, Bowen Zheng, Shidong Lv et al.

Prostate cancer is the second most common cancer in males worldwide, and mpMRI is commonly used for diagnosis. However, interpreting mpMRI is challenging and requires expertise from radiologists. This highlights the urgent need for automated grading in mpMRI. Existing studies lack integration of clinical prior information and suffer from uneven training sample distribution due to prevalence. Therefore, we propose a solution that incorporates prior knowledge, addresses the issue of uneven medical sample distribution, and maintains high interpretability in mpMRI. Firstly, we introduce Prior Knowledge-Based Feature Extraction, which mathematically models the PI-RADS criteria for prostate cancer as diagnostic information into model training. Secondly, we propose Adaptive Recall Feedback Loss to address the extremely imbalanced data problem. This method adjusts the training dynamically based on accuracy and recall in the validation set, resulting in high accuracy and recall simultaneously in the testing set.Thirdly, we design an Enhanced Cascade Prostate Cancer Classifier that classifies prostate cancer into different levels in an interpretable way, which refines the classification results and helps with clinical intervention. Our method is validated through experiments on the PI-CAI dataset and outperforms other methods with a more balanced result in both accuracy and recall rate.

Yi Huang, Bowen Zheng, Yunxi Dong et al.

Metasurface inverse design has become central to realizing complex optical functionality, yet translating target responses into executable, solver-compatible workflows still demands specialized expertise in computational electromagnetics and solver-specific software engineering. Recent large language models (LLMs) offer a complementary route to reducing this workflow-construction burden, but existing language-driven systems remain largely session-bounded and do not preserve reusable workflow knowledge across inverse-design tasks. We present an agentic framework for metasurface inverse design that addresses this limitation through context-level skill evolution. The framework couples a coding agent, evolving skill artifacts, and a deterministic evaluator grounded in physical simulation so that solver-specific strategies can be iteratively refined across tasks without modifying model weights or the underlying physics solver. We evaluate the framework on a benchmark spanning multiple metasurface inverse-design task types, with separate training-aligned and held-out task families. Evolved skills raise in-distribution task success from 38% to 74%, increase criteria pass fraction from 0.510 to 0.870, and reduce average attempts from 4.10 to 2.30. On held-out task families, binary success changes only marginally, but improvements in best margin together with shifts in error composition and agent behavior indicate partial transfer of workflow knowledge. These results suggest that the main value of skill evolution lies in accumulating reusable solver-specific expertise around reliable computational engines, thereby offering a practical path toward more autonomous and accessible metasurface inverse-design workflows.

Xiaoyu Pan, Bowen Zheng, Xinwei Jiang et al.

We present a novel locality-based learning method for cleaning and solving optical motion capture data. Given noisy marker data, we propose a new heterogeneous graph neural network which treats markers and joints as different types of nodes, and uses graph convolution operations to extract the local features of markers and joints and transform them to clean motions. To deal with anomaly markers (e.g. occluded or with big tracking errors), the key insight is that a marker's motion shows strong correlations with the motions of its immediate neighboring markers but less so with other markers, a.k.a. locality, which enables us to efficiently fill missing markers (e.g. due to occlusion). Additionally, we also identify marker outliers due to tracking errors by investigating their acceleration profiles. Finally, we propose a training regime based on representation learning and data augmentation, by training the model on data with masking. The masking schemes aim to mimic the occluded and noisy markers often observed in the real data. Finally, we show that our method achieves high accuracy on multiple metrics across various datasets. Extensive comparison shows our method outperforms state-of-the-art methods in terms of prediction accuracy of occluded marker position error by approximately 20%, which leads to a further error reduction on the reconstructed joint rotations and positions by 30%. The code and data for this paper are available at https://github.com/non-void/LocalMoCap.

Abhimanyu Talwar, Zudi Lin, Donglai Wei et al.

One of the essential tasks in connectomics is the morphology analysis of neurons and organelles like mitochondria to shed light on their biological properties. However, these biological objects often have tangled parts or complex branching patterns, which make it hard to abstract, categorize, and manipulate their morphology. In this paper, we develop a novel topological nomenclature system to name these objects like the appellation for chemical compounds to promote neuroscience analysis based on their skeletal structures. We first convert the volumetric representation into the topology-preserving reduced graph to untangle the objects. Next, we develop nomenclature rules for pyramidal neurons and mitochondria from the reduced graph and finally learn the feature embedding for shape manipulation. In ablation studies, we quantitatively show that graphs generated by our proposed method align with the perception of experts. On 3D shape retrieval and decomposition tasks, we qualitatively demonstrate that the encoded topological nomenclature features achieve better results than state-of-the-art shape descriptors. To advance neuroscience, we will release a 3D segmentation dataset of mitochondria and pyramidal neurons reconstructed from a 100um cube electron microscopy volume with their reduced graph and topological nomenclature annotations. Code is publicly available at https://github.com/donglaiw/ibexHelper.

Bowen Zheng, Weijian Luo, Guang Yang et al.

In this work, we propose Prologue, an approach to bridging the reconstruction-generation gap in autoregressive (AR) image generation. Instead of modifying visual tokens to satisfy both reconstruction and generation, Prologue generates a small set of prologue tokens prepended to the visual token sequence. These prologue tokens are trained exclusively with the AR cross-entropy (CE) loss, while visual tokens remain dedicated to reconstruction. This decoupled design lets us optimize generation through the AR model's true distribution without affecting reconstruction quality, which we further formalize from an ELBO perspective. On ImageNet 256x256, Prologue-Base reduces gFID from 21.01 to 10.75 without classifier-free guidance while keeping reconstruction almost unchanged; Prologue-Large reaches a competitive rFID of 0.99 and gFID of 1.46 using a standard AR model without auxiliary semantic supervision. Interestingly, driven only by AR gradients, prologue tokens exhibit emergent semantic structure: linear probing on 16 prologue tokens reaches 35.88% Top-1, far above the 23.71% of the first 16 tokens from a standard tokenizer; resampling with fixed prologue tokens preserves a similar high-level semantic layout. Our results suggest a new direction: generation quality can be improved by introducing a separate learned generative representation while leaving the original representation intact.

Bowen Zheng, Yihong Luo, Tianyang Hu

Discrete image tokenizers are commonly trained in two stages: first for reconstruction, and then with a prior model fitted to the frozen token sequences. This decoupling leaves the tokenizer unaware of the model that will later generate its tokens. As a result, the learned tokens may preserve image information well but still be difficult for an autoregressive (AR) prior to predict from left to right. We analyze this mismatch using Tripartite Variational Consistency (TVC), which decomposes latent-variable learning into three consistency conditions: conditional-likelihood consistency, prior consistency, and posterior consistency. TVC shows that two-stage training preserves the reconstruction side but leaves prior consistency outside the tokenizer objective: the overall token distribution is fixed before the AR prior participates in training. Motivated by this view, we add a distribution-level prior-matching signal during tokenizer training, while keeping the reconstruction objective unchanged. We optimize this signal with a Wasserstein-gradient-flow update. For hard categorical tokens, the update reduces to a token-level contrast between an auxiliary AR model that tracks the tokenizer's current token distribution and the target AR prior. It requires only forward passes through the two AR models and does not backpropagate through either of them. The resulting tokenizer, wAR-Tok, reduces AR loss and improves generation FID on CIFAR-10 and ImageNet at comparable reconstruction quality.

Bowen Zheng, Weijian Luo, Guang Yang et al.

Most discrete visual tokenizers rely on a default design: every position in the sequence shares the same codebook. Researchers try to scale the codebook size $K$ to get better reconstruction performance. Such a constant-codebook design hits a fundamental information-theoretic limit. We observe that the per-position conditional entropy of the training set decays so quickly along the sequence that, after a few positions, the conditional distribution becomes essentially deterministic. On ImageNet with $K=16384$, this happens within only 2 out of 256 positions, turning the remaining 254 into a memorization problem. We call this phenomenon the Entropy Cliff and formalize it with a simple expression: $t^{*} = \lceil \log_2 N / \log_2 K \rceil$. Interestingly, this phenomenon is not observed in language, as its natural structure keeps the effective entropy per position well below the codebook capacity. To address this, we propose Variable Codebook Size Quantization (VCQ), where the codebook size $K_t$ grows monotonically along the sequence from $K_{\min}=2$ to $K_{\max}$, leaving the loss function, parameter count, and AR training procedure unchanged. With a vanilla autoregressive Transformer and standard next-token prediction, a base version of VCQ reduces gFID w/o CFG from 27.98 to 14.80 on ImageNet $256\times256$ over the baseline. Scaled up, it reaches gFID 1.71 with 684M autoregressive parameters, without any extra training techniques such as semantic regularization or causal alignment. The extreme information bottleneck at $K_{\min}=2$ naturally induces a coarse-to-fine semantic hierarchy: a linear probe on only the first 10 tokens reaches 43.8% top-1 accuracy on ImageNet, compared to 27.1% for uniform codebooks. Ultimately, these results show that what matters is not only the total capacity of the codebook, but also how that capacity is distributed and organized.

Bowen Zheng, Zihan Lin, Enze Liu et al.

Nowadays, reading or writing comments on captivating videos has emerged as a critical part of the viewing experience on online video platforms. However, existing recommender systems primarily focus on users' interaction behaviors with videos, neglecting comment content and interaction in user preference modeling. In this paper, we propose a novel recommendation approach called LSVCR that utilizes user interaction histories with both videos and comments to jointly perform personalized video and comment recommendation. Specifically, our approach comprises two key components: sequential recommendation (SR) model and supplemental large language model (LLM) recommender. The SR model functions as the primary recommendation backbone (retained in deployment) of our method for efficient user preference modeling. Concurrently, we employ a LLM as the supplemental recommender (discarded in deployment) to better capture underlying user preferences derived from heterogeneous interaction behaviors. In order to integrate the strengths of the SR model and the supplemental LLM recommender, we introduce a two-stage training paradigm. The first stage, personalized preference alignment, aims to align the preference representations from both components, thereby enhancing the semantics of the SR model. The second stage, recommendation-oriented fine-tuning, involves fine-tuning the alignment-enhanced SR model according to specific objectives. Extensive experiments in both video and comment recommendation tasks demonstrate the effectiveness of LSVCR. Moreover, online A/B testing on KuaiShou platform verifies the practical benefits of our approach. In particular, we attain a cumulative gain of 4.13% in comment watch time.

Yuyan Ni, Shikun Feng, Haohan Chi et al.

Diffusion-based models have shown great promise in molecular generation but often require a large number of sampling steps to generate valid samples. In this paper, we introduce a novel Straight-Line Diffusion Model (SLDM) to tackle this problem, by formulating the diffusion process to follow a linear trajectory. The proposed process aligns well with the noise sensitivity characteristic of molecular structures and uniformly distributes reconstruction effort across the generative process, thus enhancing learning efficiency and efficacy. Consequently, SLDM achieves state-of-the-art performance on 3D molecule generation benchmarks, delivering a 100-fold improvement in sampling efficiency.

Bowen Zheng, Hongyu Lu, Yu Chen et al.

Recently, generative recommendation has emerged as a promising paradigm, attracting significant research attention. The basic framework involves an item tokenizer, which represents each item as a sequence of codes serving as its identifier, and a generative recommender that predicts the next item by autoregressively generating the target item identifier. However, in existing methods, both the tokenizer and the recommender are typically domain-specific, limiting their ability for effective transfer or adaptation to new domains. To this end, we propose UTGRec, a Universal item Tokenization approach for transferable Generative Recommendation. Specifically, we design a universal item tokenizer for encoding rich item semantics by adapting a multimodal large language model (MLLM). By devising tree-structured codebooks, we discretize content representations into corresponding codes for item tokenization. To effectively learn the universal item tokenizer on multiple domains, we introduce two key techniques in our approach. For raw content reconstruction, we employ dual lightweight decoders to reconstruct item text and images from discrete representations to capture general knowledge embedded in the content. For collaborative knowledge integration, we assume that co-occurring items are similar and integrate collaborative signals through co-occurrence alignment and reconstruction. Finally, we present a joint learning framework to pre-train and adapt the transferable generative recommender across multiple domains. Extensive experiments on four public datasets demonstrate the superiority of UTGRec compared to both traditional and generative recommendation baselines.

Bowen Zheng, Enze Liu, Zhongfu Chen et al.

Generative recommendation autoregressively generates item identifiers to recommend potential items. Existing methods typically adopt a one-to-one mapping strategy, where each item is represented by a single identifier. However, this scheme poses issues, such as suboptimal semantic modeling for low-frequency items and limited diversity in token sequence data. To overcome these limitations, we propose MTGRec, which leverages Multi-identifier item Tokenization to augment token sequence data for Generative Recommender pre-training. Our approach involves two key innovations: multi-identifier item tokenization and curriculum recommender pre-training. For multi-identifier item tokenization, we leverage the RQ-VAE as the tokenizer backbone and treat model checkpoints from adjacent training epochs as semantically relevant tokenizers. This allows each item to be associated with multiple identifiers, enabling a single user interaction sequence to be converted into several token sequences as different data groups. For curriculum recommender pre-training, we introduce a curriculum learning scheme guided by data influence estimation, dynamically adjusting the sampling probability of each data group during recommender pre-training. After pre-training, we fine-tune the model using a single tokenizer to ensure accurate item identification for recommendation. Extensive experiments on three public benchmark datasets demonstrate that MTGRec significantly outperforms both traditional and generative recommendation baselines in terms of effectiveness and scalability.

Bowen Zheng, Tianming Yang

Diffusion distillation is a widely used technique to reduce the sampling cost of diffusion models, yet it often requires extensive training, and the student performance tends to be degraded. Recent studies show that incorporating a GAN objective may alleviate these issues, yet the underlying mechanism remains unclear. In this work, we first identify a key limitation of distillation: mismatched step sizes and parameter numbers between the teacher and the student model lead them to converge to different local minima, rendering direct imitation suboptimal. We further demonstrate that a standalone GAN objective, without relying a distillation loss, overcomes this limitation and is sufficient to convert diffusion models into efficient one-step generators. Based on this finding, we propose that diffusion training may be viewed as a form of generative pre-training, equipping models with capabilities that can be unlocked through lightweight GAN fine-tuning. Supporting this view, we create a one-step generation model by fine-tuning a pre-trained model with 85% of parameters frozen, achieving strong performance with only 0.2M images and near-SOTA results with 5M images. We further present a frequency-domain analysis that may explain the one-step generative capability gained in diffusion training. Overall, our work provides a new perspective for diffusion training, highlighting its role as a powerful generative pre-training process, which can be the basis for building efficient one-step generation models.

Kesha Ou, Chenghao Wu, Xiaolei Wang et al.

The technical foundations of recommender systems have progressed from collaborative filtering to complex neural models and, more recently, large language models. Despite these technological advances, deployed systems often underserve their users by simply presenting a list of items, leaving the burden of exploration, comparison, and synthesis entirely on the user. This paper argues that this traditional "tool-based" paradigm fundamentally limits user experience, as the system acts as a passive filter rather than an active assistant. To address this limitation, we propose a novel deep research paradigm for recommendation, which replaces conventional item lists with comprehensive, user-centric reports. We instantiate this paradigm through RecPilot, a multi-agent framework comprising two core components: a user trajectory simulation agent that autonomously explores the item space, and a self-evolving report generation agent that synthesizes the findings into a coherent, interpretable report tailored to support user decisions. This approach reframes recommendation as a proactive, agent-driven service. Extensive experiments on public datasets demonstrate that RecPilot not only achieves strong performance in modeling user behaviors but also generates highly persuasive reports that substantially reduce user effort in item evaluation, validating the potential of this new interaction paradigm.

Qi Sun, Shengyan Li, Bowen Zheng et al.

Green's function provides an inherent connection between theoretical analysis and numerical methods for elliptic partial differential equations, and general absence of its closed-form expression necessitates surrogate modeling to guide the design of effective solvers. Unfortunately, numerical computation of Green's function remains challenging due to its doubled dimensionality and intrinsic singularity. In this paper, we present a novel singularity-encoded learning approach to resolve these problems in an unsupervised fashion. Our method embeds the Green's function within a one-order higher-dimensional space by encoding its prior estimate as an augmented variable, followed by a neural network parametrization to manage the increased dimensionality. By projecting the trained neural network solution back onto the original domain, our deep surrogate model exploits its spectral bias to accelerate conventional iterative schemes, serving either as a preconditioner or as part of a hybrid solver. The effectiveness of our proposed method is empirically verified through numerical experiments with two and four dimensional Green's functions, achieving satisfactory resolution of singularities and acceleration of iterative solvers.

Tianpei Lu, Bingsheng Zhang, Lekun Peng et al.

With the increasing deployment of generative machine learning models in privacy-sensitive domains such as healthcare and personalized services, ensuring secure inference has become a critical challenge. Secure multi-party computation (MPC) enables privacy-preserving model inference but suffers from high communication and computation overhead. The main bottleneck lies in the expensive secure evaluation of floating-point operations. Quantization offers a promising solution by converting floating-point operations into lower-precision integer computations, significantly reducing overhead. However, existing MPC-based quantized inference methods either rely on public quantization parameters-posing privacy risks-or suffer from inefficiencies, particularly in handling nonlinear functions such as activations and softmax. In this work, we propose a fine-grained, layer-wise quantization scheme and support 1-bit weight fully connected layers in a secure setting. We design a multi-input lookup table protocol to evaluate softmax efficiently and securely. Furthermore, we use dual secret sharing schemes and perform precision conversions via lookup tables, eliminating truncation overhead entirely. Experimental evaluation on BERT-base models demonstrates that our approach achieves up to $8\times$ speedup compared to Lu \emph{et al}. (NDSS 25), $9\times$ speedup compared to Gupta \emph{et al}. (PETS 24) and $22 \times$ speedup compared to Knott \emph{et al}. (NeurIPS 21).

Bowen Zheng, Ming Ma, Zhongqiao Lin et al.

Large language models are computationally expensive due to their deep structures. Prior research has shown that intermediate layers contain sufficient information to generate accurate answers, leading to the development of early-exit algorithms that reduce inference costs by terminating computation at earlier layers. However, these methods often suffer from poor performance due to misalignment between intermediate and output layer representations that lead to decoding inaccuracy. To address these challenges, we propose SPADE (SPace Alignment DEcoding), a novel decoding method that aligns intermediate layer representations with the output layer by propagating a minimally reduced sequence consisting of only the start token and the answer token. We further optimize the early-exit decision-making process by training a linear approximation of SPADE that computes entropy-based confidence metrics. Putting them together, we create a hybrid early-exit algorithm that monitors confidence levels and stops inference at intermediate layers while using SPADE to generate high-quality outputs. This approach significantly reduces inference costs without compromising accuracy, offering a scalable and efficient solution for deploying large language models in real-world applications.

Bowen Zheng

To address the challenges of 3D modeling and structural simulation in industrial environment, such as the difficulty of equipment deployment, and the difficulty of balancing accuracy and real-time performance, this paper proposes an integrated workflow, which integrates high-fidelity 3D reconstruction based on monocular video, finite element simulation analysis, and mixed reality visual display, aiming to build an interactive digital twin system for industrial inspection, equipment maintenance and other scenes. Firstly, the Neuralangelo algorithm based on deep learning is used to reconstruct the 3D mesh model with rich details from the surround-shot video. Then, the QuadRemesh tool of Rhino is used to optimize the initial triangular mesh and generate a structured mesh suitable for finite element analysis. The optimized mesh is further discretized by HyperMesh, and the material parameter setting and stress simulation are carried out in Abaqus to obtain high-precision stress and deformation results. Finally, combined with Unity and Vuforia engine, the real-time superposition and interactive operation of simulation results in the augmented reality environment are realized, which improves users 'intuitive understanding of structural response. Experiments show that the method has good simulation efficiency and visualization effect while maintaining high geometric accuracy. It provides a practical solution for digital modeling, mechanical analysis and interactive display in complex industrial scenes, and lays a foundation for the deep integration of digital twin and mixed reality technology in industrial applications.

Zeqing Wang, Bowen Zheng, Xingyi Yang et al.

Diffusion Transformer (DiT)-based video diffusion models generate high-quality videos at scale but incur prohibitive processing latency and memory costs for long videos. To address this, we propose a novel distributed inference strategy, termed DualParal. The core idea is that, instead of generating an entire video on a single GPU, we parallelize both temporal frames and model layers across GPUs. However, a naive implementation of this division faces a key limitation: since diffusion models require synchronized noise levels across frames, this implementation leads to the serialization of original parallelisms. We leverage a block-wise denoising scheme to handle this. Namely, we process a sequence of frame blocks through the pipeline with progressively decreasing noise levels. Each GPU handles a specific block and layer subset while passing previous results to the next GPU, enabling asynchronous computation and communication. To further optimize performance, we incorporate two key enhancements. Firstly, a feature cache is implemented on each GPU to store and reuse features from the prior block as context, minimizing inter-GPU communication and redundant computation. Secondly, we employ a coordinated noise initialization strategy, ensuring globally consistent temporal dynamics by sharing initial noise patterns across GPUs without extra resource costs. Together, these enable fast, artifact-free, and infinitely long video generation. Applied to the latest diffusion transformer video generator, our method efficiently produces 1,025-frame videos with up to 6.54$\times$ lower latency and 1.48$\times$ lower memory cost on 8$\times$RTX 4090 GPUs.

Bowen Zheng, Ming Ma, Zhongqiao Lin et al.

Large Language Models (LLMs) have demonstrated remarkable abilities, one of the most important being in-context learning (ICL). With ICL, LLMs can derive the underlying rule from a few demonstrations and provide answers that comply with the rule. Previous work hypothesized that the network creates a task vector in specific positions during ICL. The task vector can be computed by averaging across the dataset. It conveys the overall task information and can thus be considered global. Patching the global task vector allows LLMs to achieve zero-shot performance with dummy inputs comparable to few-shot learning. However, we find that such a global task vector does not exist in all tasks, especially in tasks that rely on rules that can only be inferred from multiple demonstrations, such as categorization tasks. Instead, the information provided by each demonstration is first transmitted to its answer position and forms a local task vector associated with the demonstration. In some tasks but not in categorization tasks, all demonstrations' local task vectors converge in later layers, forming the global task vector. We further show that local task vectors encode a high-level abstraction of rules extracted from the demonstrations. Our study provides novel insights into the mechanism underlying ICL in LLMs, demonstrating how ICL may be achieved through an information aggregation mechanism.

Sabrina Drammis, Bowen Zheng, Karthik Srinivasan et al.

A feedforward neural network using rectified linear units constructs a mapping from inputs to outputs by partitioning its input space into a set of convex regions where points within a region share a single affine transformation. In order to understand how neural networks work, when and why they fail, and how they compare to biological intelligence, we need to understand the organization and formation of these regions. Step one is to design and implement algorithms for exact region enumeration in networks beyond toy examples. In this work, we present parallel algorithms for exact enumeration in deep (and shallow) neural networks. Our work has three main contributions: (1) we present a novel algorithm framework and parallel algorithms for region enumeration; (2) we implement one of our algorithms on a variety of network architectures and experimentally show how the number of regions dictates runtime; and (3) we show, using our algorithm's output, how the dimension of a region's affine transformation impacts further partitioning of the region by deeper layers. To our knowledge, we run our implemented algorithm on networks larger than all of the networks used in the existing region enumeration literature. Further, we experimentally demonstrate the importance of parallelism for region enumeration of any reasonably sized network.

Bowen Zheng, Ran Cheng

While off-policy reinforcement learning (RL) algorithms are sample efficient due to gradient-based updates and data reuse in the replay buffer, they struggle with convergence to local optima due to limited exploration. On the other hand, population-based algorithms offer a natural exploration strategy, but their heuristic black-box operators are inefficient. Recent algorithms have integrated these two methods, connecting them through a shared replay buffer. However, the effect of using diverse data from population optimization iterations on off-policy RL algorithms has not been thoroughly investigated. In this paper, we first analyze the use of off-policy RL algorithms in combination with population-based algorithms, showing that the use of population data could introduce an overlooked error and harm performance. To test this, we propose a uniform and scalable training design and conduct experiments on our tailored framework in robot locomotion tasks from the OpenAI gym. Our results substantiate that using population data in off-policy RL can cause instability during training and even degrade performance. To remedy this issue, we further propose a double replay buffer design that provides more on-policy data and show its effectiveness through experiments. Our results offer practical insights for training these hybrid methods.

Xiangguo Liu, Chao Huang, Yixuan Wang et al.

Neural networks have shown great promises in planning, control, and general decision making for learning-enabled cyber-physical systems (LE-CPSs), especially in improving performance under complex scenarios. However, it is very challenging to formally analyze the behavior of neural network based planners for ensuring system safety, which significantly impedes their applications in safety-critical domains such as autonomous driving. In this work, we propose a hierarchical neural network based planner that analyzes the underlying physical scenarios of the system and learns a system-level behavior planning scheme with multiple scenario-specific motion-planning strategies. We then develop an efficient verification method that incorporates overapproximation of the system state reachable set and novel partition and union techniques for formally ensuring system safety under our physics-aware planner. With theoretical analysis, we show that considering the different physical scenarios and building a hierarchical planner based on such analysis may improve system safety and verifiability. We also empirically demonstrate the effectiveness of our approach and its advantage over other baselines in practical case studies of unprotected left turn and highway merging, two common challenging safety-critical tasks in autonomous driving.

Xiangguo Liu, Ruochen Jiao, Bowen Zheng et al.

Neural network-based driving planners have shown great promises in improving task performance of autonomous driving. However, it is critical and yet very challenging to ensure the safety of systems with neural network based components, especially in dense and highly interactive traffic environments. In this work, we propose a safety-driven interactive planning framework for neural network-based lane changing. To prevent over conservative planning, we identify the driving behavior of surrounding vehicles and assess their aggressiveness, and then adapt the planned trajectory for the ego vehicle accordingly in an interactive manner. The ego vehicle can proceed to change lanes if a safe evasion trajectory exists even in the predicted worst case; otherwise, it can stay around the current lateral position or return back to the original lane. We quantitatively demonstrate the effectiveness of our planner design and its advantage over baseline methods through extensive simulations with diverse and comprehensive experimental settings, as well as in real-world scenarios collected by an autonomous vehicle company.

Sensong An, Bowen Zheng, Mikhail Y. Shalaginov et al.

Metasurfaces have provided a novel and promising platform for the realization of compact and large-scale optical devices. The conventional metasurface design approach assumes periodic boundary conditions for each element, which is inaccurate in most cases since the near-field coupling effects between elements will change when surrounded by non-identical structures. In this paper, we propose a deep learning approach to predict the actual electromagnetic (EM) responses of each target meta-atom placed in a large array with near-field coupling effects taken into account. The predicting neural network takes the physical specifications of the target meta-atom and its neighbors as input, and calculates its phase and amplitude in milliseconds. This approach can be applied to explain metasurfaces' performance deterioration caused by mutual coupling and further used to optimize their efficiencies once combined with optimization algorithms. To demonstrate the efficacy of this methodology, we obtain large improvements in efficiency for a beam deflector and a metalens over the conventional design approach. Moreover, we show the correlations between a metasurface's performance and its design errors caused by mutual coupling are not bound to certain specifications (materials, shapes, etc.). As such, we envision that this approach can be readily applied to explore the mutual coupling effects and improve the performance of various metasurface designs.

Sensong An, Bowen Zheng, Mikhail Y. Shalaginov et al.

Metasurfaces have shown promising potentials in shaping optical wavefronts while remaining compact compared to bulky geometric optics devices. Design of meta-atoms, the fundamental building blocks of metasurfaces, relies on trial-and-error method to achieve target electromagnetic responses. This process includes the characterization of an enormous amount of different meta-atom designs with different physical and geometric parameters, which normally demands huge computational resources. In this paper, a deep learning-based metasurface/meta-atom modeling approach is introduced to significantly reduce the characterization time while maintaining accuracy. Based on a convolutional neural network (CNN) structure, the proposed deep learning network is able to model meta-atoms with free-form 2D patterns and different lattice sizes, material refractive indexes and thicknesses. Moreover, the presented approach features the capability to predict meta-atoms' wide spectrum responses in the timescale of milliseconds, which makes it attractive for applications such as fast meta-atom/metasurface on-demand designs and optimizations.

Ying Ju, Fubang Zhao, Shijie Chen et al.

Conversational Question Answering is a challenging task since it requires understanding of conversational history. In this project, we propose a new system RoBERTa + AT +KD, which involves rationale tagging multi-task, adversarial training, knowledge distillation and a linguistic post-process strategy. Our single model achieves 90.4(F1) on the CoQA test set without data augmentation, outperforming the current state-of-the-art single model by 2.6% F1.

Sensong An, Bowen Zheng, Hong Tang et al.

Metasurfaces have enabled precise electromagnetic wave manipulation with strong potential to obtain unprecedented functionalities and multifunctional behavior in flat optical devices. These advantages in precision and functionality come at the cost of tremendous difficulty in finding individual meta-atom structures based on specific requirements (commonly formulated in terms of electromagnetic responses), which makes the design of multifunctional metasurfaces a key challenge in this field. In this paper, we present a Generative Adversarial Networks (GAN) that can tackle this problem and generate meta-atom/metasurface designs to meet multifunctional design goals. Unlike conventional trial-and-error or iterative optimization design methods, this new methodology produces on-demand free-form structures involving only a single design iteration. More importantly, the network structure and the robust training process are independent of the complexity of design objectives, making this approach ideal for multifunctional device design. Additionally, the ability of the network to generate distinct classes of structures with similar electromagnetic responses but different physical features could provide added latitude to accommodate other considerations such as fabrication constraints and tolerances. We demonstrate the network's ability to produce a variety of multifunctional metasurface designs by presenting a bifocal metalens, a polarization-multiplexed beam deflector, a polarization-multiplexed metalens and a polarization-independent metalens.

Sensong An, Clayton Fowler, Bowen Zheng et al.

Metasurfaces have become a promising means for manipulating optical wavefronts in flat and high-performance optical devices. Conventional metasurface device design relies on trial-and-error methods to obtain target electromagnetic (EM) response, an approach that demands significant efforts to investigate the enormous number of possible meta-atom structures. In this paper, a deep neural network approach is introduced that significantly improves on both speed and accuracy compared to techniques currently used to assemble metasurface-based devices. Our neural network approach overcomes three key challenges that have limited previous neural-network-based design schemes: input/output vector dimensional mismatch, accurate EM-wave phase prediction, as well as adaptation to 3-D dielectric structures, and can be generically applied to a wide variety of metasurface device designs across the entire electromagnetic spectrum. Using this new methodology, examples of neural networks capable of producing on-demand designs for meta-atoms, metasurface filters, and phase-change reconfigurable metasurfaces are demonstrated.

Harshil Shah, Bowen Zheng, David Barber

We introduce the Attentive Unsupervised Text (W)riter (AUTR), which is a word level generative model for natural language. It uses a recurrent neural network with a dynamic attention and canvas memory mechanism to iteratively construct sentences. By viewing the state of the memory at intermediate stages and where the model is placing its attention, we gain insight into how it constructs sentences. We demonstrate that AUTR learns a meaningful latent representation for each sentence, and achieves competitive log-likelihood lower bounds whilst being computationally efficient. It is effective at generating and reconstructing sentences, as well as imputing missing words.