Chenyu Zhang

Vipul Gupta, Zhuowan Li, Adam Kortylewski et al.

While Visual Question Answering (VQA) has progressed rapidly, previous works raise concerns about robustness of current VQA models. In this work, we study the robustness of VQA models from a novel perspective: visual context. We suggest that the models over-rely on the visual context, i.e., irrelevant objects in the image, to make predictions. To diagnose the model's reliance on visual context and measure their robustness, we propose a simple yet effective perturbation technique, SwapMix. SwapMix perturbs the visual context by swapping features of irrelevant context objects with features from other objects in the dataset. Using SwapMix we are able to change answers to more than 45 % of the questions for a representative VQA model. Additionally, we train the models with perfect sight and find that the context over-reliance highly depends on the quality of visual representations. In addition to diagnosing, SwapMix can also be applied as a data augmentation strategy during training in order to regularize the context over-reliance. By swapping the context object features, the model reliance on context can be suppressed effectively. Two representative VQA models are studied using SwapMix: a co-attention model MCAN and a large-scale pretrained model LXMERT. Our experiments on the popular GQA dataset show the effectiveness of SwapMix for both diagnosing model robustness and regularizing the over-reliance on visual context. The code for our method is available at https://github.com/vipulgupta1011/swapmix

Chenyu Zhang, Yiwen Liu, Yin Sun et al.

Sales lead conversion in high-stakes domains (e.g., automotive, real estate) differs fundamentally from e-commerce recommendation due to prolonged decision cycles and multi-stage funnels. Traditional lead scoring methods rule-based scorecards, machine learning, or pointwise CTR models face severe challenges: sparse supervision, a semantic gap in unstructured CRM logs, and inability to capture relative lead priority. While Large Language Models(LLMs) offer superior semantic understanding of customer interactions, general-purpose LLMs are ill-suited for lead ranking: they generate text rather than comparable scores, and lack alignment with the hierarchical priorities of sales funnels. We introduce an LLM-based discriminative framework for sales lead scoring, which supports joint modeling of structured CRM features and unstructured customer interactions. On top of this framework, we propose HPRO (Hierarchical Preference Ranking Optimization), which augments sales lead scoring with a hierarchical preference ranking objective. HPRO employs a margin-aware Bradley-Terry formulation to transform sparse binary labels into dense, funnel-aware preference pairs, enabling lead scoring to leverage both pointwise and pairwise supervision. Experiments on large-scale data from a leading NEV brand demonstrate state-of-the-art classification (AUC 0.8161) and ranking performance (+39.7% precision among top-ranked leads). A 132-day online A/B test validates 9.5% sales volume uplift, confirming real-world commercial impact.

Chenyu Zhang, Benjamin Van Durme, Zhuowan Li et al.

Our commonsense knowledge about objects includes their typical visual attributes; we know that bananas are typically yellow or green, and not purple. Text and image corpora, being subject to reporting bias, represent this world-knowledge to varying degrees of faithfulness. In this paper, we investigate to what degree unimodal (language-only) and multimodal (image and language) models capture a broad range of visually salient attributes. To that end, we create the Visual Commonsense Tests (ViComTe) dataset covering 5 property types (color, shape, material, size, and visual co-occurrence) for over 5000 subjects. We validate this dataset by showing that our grounded color data correlates much better than ungrounded text-only data with crowdsourced color judgments provided by Paik et al. (2021). We then use our dataset to evaluate pretrained unimodal models and multimodal models. Our results indicate that multimodal models better reconstruct attribute distributions, but are still subject to reporting bias. Moreover, increasing model size does not enhance performance, suggesting that the key to visual commonsense lies in the data.

Chenyu Zhang, Mingwang Hu, Wenhui Li et al.

Recently, the text-to-image diffusion model has gained considerable attention from the community due to its exceptional image generation capability. A representative model, Stable Diffusion, amassed more than 10 million users within just two months of its release. This surge in popularity has facilitated studies on the robustness and safety of the model, leading to the proposal of various adversarial attack methods. Simultaneously, there has been a marked increase in research focused on defense methods to improve the robustness and safety of these models. In this survey, we provide a comprehensive review of the literature on adversarial attacks and defenses targeting text-to-image diffusion models. We begin with an overview of text-to-image diffusion models, followed by an introduction to a taxonomy of adversarial attacks and an in-depth review of existing attack methods. We then present a detailed analysis of current defense methods that improve model robustness and safety. Finally, we discuss ongoing challenges and explore promising future research directions. For a complete list of the adversarial attack and defense methods covered in this survey, please refer to our curated repository at https://github.com/datar001/Awesome-AD-on-T2IDM.

Chenyu Zhang, Xu Chen, Xuan Di · mit

Mean field games (MFGs) model interactions in large-population multi-agent systems through population distributions. Traditional learning methods for MFGs are based on fixed-point iteration (FPI), where policy updates and induced population distributions are computed separately and sequentially. However, FPI-type methods may suffer from inefficiency and instability due to potential oscillations caused by this forward-backward procedure. In this work, we propose a novel perspective that treats the policy and population as a unified parameter controlling the game dynamics. By applying stochastic parameter approximation to this unified parameter, we develop SemiSGD, a simple stochastic gradient descent (SGD)-type method, where an agent updates its policy and population estimates simultaneously and fully asynchronously. Building on this perspective, we further apply linear function approximation (LFA) to the unified parameter, resulting in the first population-aware LFA (PA-LFA) for learning MFGs on continuous state-action spaces. A comprehensive finite-time convergence analysis is provided for SemiSGD with PA-LFA, including its convergence to the equilibrium for linear MFGs -- a class of MFGs with a linear structure concerning the population -- under the standard contractivity condition, and to a neighborhood of the equilibrium under a more practical condition. We also characterize the approximation error for non-linear MFGs. We validate our theoretical findings with six experiments on three MFGs.

DeepSeek-AI, Daya Guo, Dejian Yang et al. · stanford, tsinghua

We introduce our first-generation reasoning models, DeepSeek-R1-Zero and DeepSeek-R1. DeepSeek-R1-Zero, a model trained via large-scale reinforcement learning (RL) without supervised fine-tuning (SFT) as a preliminary step, demonstrates remarkable reasoning capabilities. Through RL, DeepSeek-R1-Zero naturally emerges with numerous powerful and intriguing reasoning behaviors. However, it encounters challenges such as poor readability, and language mixing. To address these issues and further enhance reasoning performance, we introduce DeepSeek-R1, which incorporates multi-stage training and cold-start data before RL. DeepSeek-R1 achieves performance comparable to OpenAI-o1-1217 on reasoning tasks. To support the research community, we open-source DeepSeek-R1-Zero, DeepSeek-R1, and six dense models (1.5B, 7B, 8B, 14B, 32B, 70B) distilled from DeepSeek-R1 based on Qwen and Llama.

Chenyu Zhang, Muhammad Monjurul Karim, Ruwen Qin

The vast network of bridges in the United States raises a high requirement for maintenance and rehabilitation. The massive cost of manual visual inspection to assess bridge conditions is a burden to some extent. Advanced robots have been leveraged to automate inspection data collection. Automating the segmentations of multiclass elements and surface defects on the elements in the large volume of inspection image data would facilitate an efficient and effective assessment of the bridge condition. Training separate single-task networks for element parsing (i.e., semantic segmentation of multiclass elements) and defect segmentation fails to incorporate the close connection between these two tasks. Both recognizable structural elements and apparent surface defects are present in the inspection images. This paper is motivated to develop a multitask deep learning model that fully utilizes such interdependence between bridge elements and defects to boost the model's task performance and generalization. Furthermore, the study investigated the effectiveness of the proposed model designs for improving task performance, including feature decomposition, cross-talk sharing, and multi-objective loss function. A dataset with pixel-level labels of bridge elements and corrosion was developed for model training and testing. Quantitative and qualitative results from evaluating the developed multitask deep model demonstrate its advantages over the single-task-based model not only in performance (2.59% higher mIoU on bridge parsing and 1.65% on corrosion segmentation) but also in computational time and implementation capability.

Chenyu Zhang, Xinchen Lyu, Chenshan Ren et al.

Positional encoding plays a pivotal role in determin?ing the extrapolation and generalization performance of wireless foundation models for channel state information (CSI) modeling, latent characterization, and task-specific prediction. However, existing CSI models inherit static or one-dimensional positional priors from natural language and vision architectures, which fundamentally misalign with the intrinsic physics of wireless channels by lacking explicit relative decay, collapsing the 3D spatio-temporal-frequency structure, and remaining scenario?rigid. This paper proposes Adaptive 3D-RoPE, a physics-aligned rotary positional encoding that establishes the structural corner?stone for wireless foundation models. The framework integrates a learnable, axis-decoupled 3D frequency bank to explicitly disentangle multi-dimensional phase dependencies, coupled with a lightweight channel-conditioned controller that dynamically modulates the prior via compact global CSI descriptors. This sample-adaptive mechanism transforms positional encoding from a static transformer component into a dynamic, coherence-aware inductive bias to resolve heterogeneous channel physics. Extensive experiments across 100 datasets demonstrate the superiority of the proposed scheme in both scale extrapolation and zero-shot generalization. Compared to the state-of-the-art, our method achieves up to a 10.7 dB reduction in normalized mean square error (NMSE) under 8 times antenna scale extrapolation. Given the same CSI input scales, our method can also improve zero-shot NMSE by 1.07 dB across unseen mobility scenarios and 0.90 dB in low-frequency-to-millimeter-wave tasks.

Chenyu Zhang, Muhammad Monjurul Karim, Zhaozheng Yin et al.

Aerial robots such as drones have been leveraged to perform bridge inspections. Inspection images with both recognizable structural elements and apparent surface defects can be collected by onboard cameras to provide valuable information for the condition assessment. This article aims to determine a suitable deep neural network (DNN) for parsing multiclass bridge elements in inspection images. An extensive set of quantitative evaluations along with qualitative examples show that High-Resolution Net (HRNet) possesses the desired ability. With data augmentation and a training sample of 130 images, a pre-trained HRNet is efficiently transferred to the task of structural element parsing and has achieved a 92.67% mean F1-score and 86.33% mean IoU.

Zhibo Jin, Jiayu Zhang, Zhiyu Zhu et al.

In recent times, the swift evolution of adversarial attacks has captured widespread attention, particularly concerning their transferability and other performance attributes. These techniques are primarily executed at the sample level, frequently overlooking the intrinsic parameters of models. Such neglect suggests that the perturbations introduced in adversarial samples might have the potential for further reduction. Given the essence of adversarial attacks is to impair model integrity with minimal noise on original samples, exploring avenues to maximize the utility of such perturbations is imperative. Against this backdrop, we have delved into the complexities of adversarial attack algorithms, dissecting the adversarial process into two critical phases: the Directional Supervision Process (DSP) and the Directional Optimization Process (DOP). While DSP determines the direction of updates based on the current samples and model parameters, it has been observed that existing model parameters may not always be conducive to adversarial attacks. The impact of models on adversarial efficacy is often overlooked in current research, leading to the neglect of DSP. We propose that under certain conditions, fine-tuning model parameters can significantly enhance the quality of DSP. For the first time, we propose that under certain conditions, fine-tuning model parameters can significantly improve the quality of the DSP. We provide, for the first time, rigorous mathematical definitions and proofs for these conditions, and introduce multiple methods for fine-tuning model parameters within DSP. Our extensive experiments substantiate the effectiveness of the proposed P3A method. Our code is accessible at: https://anonymous.4open.science/r/P3A-A12C/

Yufei Shi, Weilong Yan, Naixuan Huang et al.

Existing approaches for digital short-drama production typically rely on one-shot LLM generated scripts and loosely coupled pipelines, which fail to satisfy three key requirements of short-drama generation: (1) narrative pacing, resulting in weak hooks, insufficient escalation, and unattractive endings; (2) spatial consistency, leading to drifting scene layouts and inconsistent character positions across clips; and (3) production-level quality control, requiring extensive manual review and correction across script and visual stages. We present One Sentence, One Drama, a hierarchical multi-agent framework that transforms a user's single-sentence idea into a fully produced short drama through structured intermediate modules and iterative refinement. Our approach is built upon three key components: (1) a multi-agent debate-based story generation module that enforces short-drama pacing and narrative coherence; (2) a 3D-grounded first-frame generation mechanism that establishes a shared spatial reference for consistent character positioning and scene layout across clips; and (3) multi-stage reviewer loops that perform comprehensive error detection and targeted revision across script, visual, and video generation stages. We also introduce scene-level BGM matching and scene transition planning to improve the audience's immersive experience. To systematically evaluate this task, we introduce Short-Drama-Bench, a benchmark that extends standard video quality metrics with short-drama-specific criteria. Experimental results demonstrate that our method significantly outperforms existing pipelines in narrative quality, cross-clip consistency, and overall viewing experience.

Zhibo Jin, Jiayu Zhang, Zhiyu Zhu et al.

Transferable adversarial attacks pose significant threats to deep neural networks, particularly in black-box scenarios where internal model information is inaccessible. Studying adversarial attack methods helps advance the performance of defense mechanisms and explore model vulnerabilities. These methods can uncover and exploit weaknesses in models, promoting the development of more robust architectures. However, current methods for transferable attacks often come with substantial computational costs, limiting their deployment and application, especially in edge computing scenarios. Adversarial generative models, such as Generative Adversarial Networks (GANs), are characterized by their ability to generate samples without the need for retraining after an initial training phase. GE-AdvGAN, a recent method for transferable adversarial attacks, is based on this principle. In this paper, we propose a novel general framework for gradient editing-based transferable attacks, named GE-AdvGAN+, which integrates nearly all mainstream attack methods to enhance transferability while significantly reducing computational resource consumption. Our experiments demonstrate the compatibility and effectiveness of our framework. Compared to the baseline AdvGAN, our best-performing method, GE-AdvGAN++, achieves an average ASR improvement of 47.8. Additionally, it surpasses the latest competing algorithm, GE-AdvGAN, with an average ASR increase of 5.9. The framework also exhibits enhanced computational efficiency, achieving 2217.7 FPS, outperforming traditional methods such as BIM and MI-FGSM. The implementation code for our GE-AdvGAN+ framework is available at https://github.com/GEAdvGANP

Chenyu Zhang, Xinchen Lyu, Chenshan Ren et al.

Wireless foundation models promise transformative capabilities for channel state information (CSI) processing across diverse 6G network applications, yet face fundamental challenges due to the inherent dual heterogeneity of CSI across both scale and scenario dimensions. However, current pretraining approaches either constrain inputs to fixed dimensions or isolate training by scale, limiting the generalization and scalability of wireless foundation models. In this paper, we propose HeterCSI, a channel-adaptive pretraining framework that reconciles training efficiency with robust cross-scenario generalization via a new understanding of gradient dynamics in heterogeneous CSI pretraining. Our key insight reveals that CSI scale heterogeneity primarily causes destructive gradient interference, while scenario diversity actually promotes constructive gradient alignment when properly managed. Specifically, we formulate heterogeneous CSI batch construction as a partitioning optimization problem that minimizes zero-padding overhead while preserving scenario diversity. To solve this, we develop a scale-aware adaptive batching strategy that aligns CSI samples of similar scales, and design a double-masking mechanism to isolate valid signals from padding artifacts. Extensive experiments on 12 datasets demonstrate that HeterCSI establishes a generalized foundation model without scenario-specific finetuning, achieving superior average performance over full-shot baselines. Compared to the state-of-the-art zero-shot benchmark WiFo, it reduces NMSE by 7.19 dB, 4.08 dB, and 5.27 dB for CSI reconstruction, time-domain, and frequency-domain prediction, respectively. The proposed HeterCSI framework also reduces training latency by 53% compared to existing approaches while improving generalization performance by 1.53 dB on average.

Ziye Chen, Hongbin Lin, Chenyu Zhang et al.

Zeroth-order (ZO) optimization enables large-language-model fine-tuning without storing backpropagation activations, while LoRA supplies compact trainable adapters. Combining them creates a rank paradox: increasing LoRA rank improves adapter capacity, but standard two-point ZO either perturbs a rank-dependent number of coordinates or, under atomwise updates, can make the finite-difference signal unobservable. This paper shows that the bottleneck is a measurement-topology problem rather than a need for an external subspace. LoRA already decomposes into matched rank-$1$ atoms, each a complete factor-coordinate block of dimension $d_\text{out}+d_\text{in}$. Querying one atom per step keeps the stored adapter rank $r$ while removing $r$ from the single-query perturbation dimension. The naive atomwise query is still miscalibrated: if it inherits canonical LoRA scaling $α/r$, the active finite-difference signal shrinks as $1/r$ and the active finite-difference signal-to-noise ratio (FD-SNR) as $1/r^2$, producing directional collapse under a fixed residual evaluation-noise floor. AR1-ZO pairs alternating rank-$1$ atom queries with topology-aware scaling $γ=αr$, restoring rank-invariant active signal without auxiliary bases, activation hooks, curvature estimates, or extra forward queries. Theory proves atom minimality, rank-independent active query dimension, directional collapse and restoration, and the remaining rank dependence as an amortized coverage cost. Experiments on OPT and Qwen3 models validate the signal mechanism and show that AR1-ZO makes high-rank LoRA effective among matched-budget ZO methods under the standard two-forward-pass query budget.

Minsol Michelle Kim, Daniel M. Low, David Lafond et al.

Breaking negative mental health cycles, including rumination and recurring regrets, requires reflection that translates awareness into behavioral change. Grounded in the Transtheoretical Model (TTM) and Gross's Emotion Regulation (ER) Process Model, we examine how Technologies Supporting Self-Reflection (TSR) bridge reflection and action. In a 15-day in-the-wild study (N = 20), participants used a voice-based journaling system to capture regrets and wishes and engaged in WhatIf-Planning, a novel structured reflection module integrating counterfactual thinking with if-then planning. Participants were randomized to either a free-form condition or a Gross-guided condition, which maps the five processes of Gross's ER model into explicit journaling prompts. We contribute: (1) a unified reflection-to-action TSR system that operationalizes the Preparation stage of TTM to bridge Contemplation and Action, and (2) triangulated empirical evidence from an in-the-wild journaling study that first operationalizes Gross's Process Model, revealing effects on coping flexibility and emotion regulation in daily life. Results show significant pre-post improvements in coping flexibility, indicating adaptive self-regulation across conditions, with the Gross-guided group generating more counterfactual alternatives, articulating concrete if-then action plans, and implementing more plans for self-driven change.

Chenyu Zhang, Daniil Cherniavskii, Antonios Tragoudaras et al.

Recent advances in image and video generation raise hopes that these models possess world modeling capabilities, the ability to generate realistic, physically plausible videos. This could revolutionize applications in robotics, autonomous driving, and scientific simulation. However, before treating these models as world models, we must ask: Do they adhere to physical conservation laws? To answer this, we introduce Morpheus, a benchmark for evaluating video generation models on physical reasoning. It features 80 real-world videos capturing physical phenomena, guided by conservation laws. Since artificial generations lack ground truth, we assess physical plausibility using physics-informed metrics evaluated with respect to infallible conservation laws known per physical setting, leveraging advances in physics-informed neural networks and vision-language foundation models. Our findings reveal that even with advanced prompting and video conditioning, current models struggle to encode physical principles despite generating aesthetically pleasing videos. All data, leaderboard, and code are open-sourced at our project page.

Nguyen Huu Bao Long, Chenyu Zhang, Yuzhi Shi et al.

Vision Transformers with various attention modules have demonstrated superior performance on vision tasks. While using sparsity-adaptive attention, such as in DAT, has yielded strong results in image classification, the key-value pairs selected by deformable points lack semantic relevance when fine-tuning for semantic segmentation tasks. The query-aware sparsity attention in BiFormer seeks to focus each query on top-k routed regions. However, during attention calculation, the selected key-value pairs are influenced by too many irrelevant queries, reducing attention on the more important ones. To address these issues, we propose the Deformable Bi-level Routing Attention (DBRA) module, which optimizes the selection of key-value pairs using agent queries and enhances the interpretability of queries in attention maps. Based on this, we introduce the Deformable Bi-level Routing Attention Transformer (DeBiFormer), a novel general-purpose vision transformer built with the DBRA module. DeBiFormer has been validated on various computer vision tasks, including image classification, object detection, and semantic segmentation, providing strong evidence of its effectiveness.Code is available at {https://github.com/maclong01/DeBiFormer}

Chenyu Zhang, Kunlun Xu, Zichen Liu et al.

Vision-language models (VLMs) encounter considerable challenges when adapting to domain shifts stemming from changes in data distribution. Test-time adaptation (TTA) has emerged as a promising approach to enhance VLM performance under such conditions. In practice, test data often arrives in batches, leading to increasing interest in the transductive TTA setting. However, existing TTA methods primarily focus on individual test samples, overlooking crucial cross-sample correlations within a batch. While recent ViT-based TTA methods have introduced batch-level adaptation, they remain suboptimal for VLMs due to inadequate integration of the text modality. To address these limitations, we propose a novel transductive TTA framework, Supportive Clique-based Attribute Prompting (SCAP), which effectively combines visual and textual information to enhance adaptation by generating fine-grained attribute prompts across test batches. SCAP first forms supportive cliques of test samples in an unsupervised manner based on visual similarity and learns an attribute prompt for each clique, capturing shared attributes critical for adaptation. For each test sample, SCAP aggregates attribute prompts from its associated cliques, providing enriched contextual information. To ensure adaptability over time, we incorporate a retention module that dynamically updates attribute prompts and their associated attributes as new data arrives. Comprehensive experiments across multiple benchmarks demonstrate that SCAP outperforms existing state-of-the-art methods, significantly advancing VLM generalization under domain shifts. Our code is available at https://github.com/zhoujiahuan1991/CVPR2025-SCAP.

Chenxin Tao, Shiqian Su, Xizhou Zhu et al.

The rapid advance of Large Language Models (LLMs) has catalyzed the development of Vision-Language Models (VLMs). Monolithic VLMs, which avoid modality-specific encoders, offer a promising alternative to the compositional ones but face the challenge of inferior performance. Most existing monolithic VLMs require tuning pre-trained LLMs to acquire vision abilities, which may degrade their language capabilities. To address this dilemma, this paper presents a novel high-performance monolithic VLM named HoVLE. We note that LLMs have been shown capable of interpreting images, when image embeddings are aligned with text embeddings. The challenge for current monolithic VLMs actually lies in the lack of a holistic embedding module for both vision and language inputs. Therefore, HoVLE introduces a holistic embedding module that converts visual and textual inputs into a shared space, allowing LLMs to process images in the same way as texts. Furthermore, a multi-stage training strategy is carefully designed to empower the holistic embedding module. It is first trained to distill visual features from a pre-trained vision encoder and text embeddings from the LLM, enabling large-scale training with unpaired random images and text tokens. The whole model further undergoes next-token prediction on multi-modal data to align the embeddings. Finally, an instruction-tuning stage is incorporated. Our experiments show that HoVLE achieves performance close to leading compositional models on various benchmarks, outperforming previous monolithic models by a large margin. Model available at https://huggingface.co/OpenGVLab/HoVLE.

Chenyu Zhang, Lanjun Wang, Yiwen Ma et al.

Text-to-Image(T2I) models typically deploy safety filters to prevent the generation of sensitive images. Unfortunately, recent jailbreaking attack methods manually design prompts for the LLM to generate adversarial prompts, which effectively bypass safety filters while producing sensitive images, exposing safety vulnerabilities of T2I models. However, due to the LLM's limited understanding of the T2I model and its safety filters, existing methods require numerous queries to achieve a successful attack, limiting their practical applicability. To address this issue, we propose Reason2Attack(R2A), which aims to enhance the LLM's reasoning capabilities in generating adversarial prompts by incorporating the jailbreaking attack into the post-training process of the LLM. Specifically, we first propose a CoT example synthesis pipeline based on Frame Semantics, which generates adversarial prompts by identifying related terms and corresponding context illustrations. Using CoT examples generated by the pipeline, we fine-tune the LLM to understand the reasoning path and format the output structure. Subsequently, we incorporate the jailbreaking attack task into the reinforcement learning process of the LLM and design an attack process reward that considers prompt length, prompt stealthiness, and prompt effectiveness, aiming to further enhance reasoning accuracy. Extensive experiments on various T2I models show that R2A achieves a better attack success ratio while requiring fewer queries than baselines. Moreover, our adversarial prompts demonstrate strong attack transferability across both open-source and commercial T2I models.

Chenyu Zhang, Lanjun Wang, Yiwen Ma et al.

Text-to-image (T2I) models commonly incorporate defense mechanisms to prevent the generation of sensitive images. Unfortunately, recent jailbreak attacks have shown that adversarial prompts can effectively bypass these mechanisms and induce T2I models to produce sensitive content, revealing critical safety vulnerabilities. However, existing attack methods implicitly assume that the attacker knows the type of deployed defenses, which limits their effectiveness against unknown or diverse defense mechanisms. In this work, we reveal an underexplored vulnerability of T2I models to metaphor-based jailbreak attacks (MJA), which aims to attack diverse defense mechanisms without prior knowledge of their type by generating metaphor-based adversarial prompts. Specifically, MJA consists of two modules: an LLM-based multi-agent generation module (LMAG) and an adversarial prompt optimization module (APO). LMAG decomposes the generation of metaphor-based adversarial prompts into three subtasks: metaphor retrieval, context matching, and adversarial prompt generation. Subsequently, LMAG coordinates three LLM-based agents to generate diverse adversarial prompts by exploring various metaphors and contexts. To enhance attack efficiency, APO first trains a surrogate model to predict the attack results of adversarial prompts and then designs an acquisition strategy to adaptively identify optimal adversarial prompts. Extensive experiments on T2I models with various external and internal defense mechanisms demonstrate that MJA achieves stronger attack performance while using fewer queries, compared with six baseline methods. Additionally, we provide an in-depth vulnerability analysis suggesting that metaphor-based adversarial prompts evade safety mechanisms by inducing semantic ambiguity, while sensitive images arise from the model's probabilistic interpretation of concealed semantics.

Shuai Guo, Houqiang Zhong, Qiuwen Wang et al.

Recently, NVS in human-object interaction scenes has received increasing attention. Existing human-object interaction datasets mainly consist of static data with limited views, offering only RGB images or videos, mostly containing interactions between a single person and objects. Moreover, these datasets exhibit complexities in lighting environments, poor synchronization, and low resolution, hindering high-quality human-object interaction studies. In this paper, we introduce a new people-object interaction dataset that comprises 38 series of 30-view multi-person or single-person RGB-D video sequences, accompanied by camera parameters, foreground masks, SMPL models, some point clouds, and mesh files. Video sequences are captured by 30 Kinect Azures, uniformly surrounding the scene, each in 4K resolution 25 FPS, and lasting for 1$\sim$19 seconds. Meanwhile, we evaluate some SOTA NVS models on our dataset to establish the NVS benchmarks. We hope our work can inspire further research in humanobject interaction.

Yuda Bi, Chenyu Zhang, Qiheng Wang et al.

Grokking -- the delayed onset of generalization after early memorization -- is often described with phase-transition language, but that claim has lacked falsifiable finite-size inputs. Here we supply those inputs by treating the group order $p$ of $\mathbb{Z}_p$ as an admissible extensive variable and a held-out spectral head-tail contrast as a representation-level order parameter, then apply a condensed-matter-style diagnostic chain to coarse-grid sweeps and a dense near-critical addition audit. Binder-like crossings reveal a shared finite-size boundary, and susceptibility comparison strongly disfavors a smooth-crossover interpretation ($Δ\mathrm{AIC}=16.8$ in the near-critical audit). Phase-transition language in grokking can therefore be tested as a quantitative finite-size claim rather than invoked as analogy alone, although the transition order remains unresolved at present.

Ruidong Chen, Honglin Guo, Lanjun Wang et al.

Recent advances in text-to-image diffusion models enable photorealistic image generation, but they also risk producing malicious content, such as NSFW images. To mitigate risk, concept erasure methods are studied to facilitate the model to unlearn specific concepts. However, current studies struggle to fully erase malicious concepts implicitly embedded in prompts (e.g., metaphorical expressions or adversarial prompts) while preserving the model's normal generation capability. To address this challenge, our study proposes TRCE, using a two-stage concept erasure strategy to achieve an effective trade-off between reliable erasure and knowledge preservation. Firstly, TRCE starts by erasing the malicious semantics implicitly embedded in textual prompts. By identifying a critical mapping objective(i.e., the [EoT] embedding), we optimize the cross-attention layers to map malicious prompts to contextually similar prompts but with safe concepts. This step prevents the model from being overly influenced by malicious semantics during the denoising process. Following this, considering the deterministic properties of the sampling trajectory of the diffusion model, TRCE further steers the early denoising prediction toward the safe direction and away from the unsafe one through contrastive learning, thus further avoiding the generation of malicious content. Finally, we conduct comprehensive evaluations of TRCE on multiple malicious concept erasure benchmarks, and the results demonstrate its effectiveness in erasing malicious concepts while better preserving the model's original generation ability. The code is available at: http://github.com/ddgoodgood/TRCE. CAUTION: This paper includes model-generated content that may contain offensive material.

Chenyu Zhang, Zhaozheng Yin, Ruwen Qin

Efficiently monitoring the condition of civil infrastructure requires automating the structural condition assessment in visual inspection. This paper proposes an Attention-Enhanced Co-Interactive Fusion Network (AECIF-Net) for automatic structural condition assessment in visual bridge inspection. AECIF-Net can simultaneously parse structural elements and segment surface defects on the elements in inspection images. It integrates two task-specific relearning subnets to extract task-specific features from an overall feature embedding. A co-interactive feature fusion module further captures the spatial correlation and facilitates information sharing between tasks. Experimental results demonstrate that the proposed AECIF-Net outperforms the current state-of-the-art approaches, achieving promising performance with 92.11% mIoU for element segmentation and 87.16% mIoU for corrosion segmentation on the test set of the new benchmark dataset Steel Bridge Condition Inspection Visual (SBCIV). An ablation study verifies the merits of the designs for AECIF-Net, and a case study demonstrates its capability to automate structural condition assessment.

Chenyu Zhang, Wenxue Guan, Xiaodan Xing et al.

Whole heart segmentation (WHS) supports cardiovascular disease (CVD) diagnosis, disease monitoring, treatment planning, and prognosis. Deep learning has become the most widely used method for WHS applications in recent years. However, segmentation of whole-heart structures faces numerous challenges including heart shape variability during the cardiac cycle, clinical artifacts like motion and poor contrast-to-noise ratio, domain shifts in multi-center data, and the distinct modalities of CT and MRI. To address these limitations and improve segmentation quality, this paper introduces a new topology-preserving module that is integrated into deep neural networks. The implementation achieves anatomically plausible segmentation by using learned topology-preserving fields, which are based entirely on 3D convolution and are therefore very effective for 3D voxel data. We incorporate natural constraints between structures into the end-to-end training and enrich the feature representation of the neural network. The effectiveness of the proposed method is validated on an open-source medical heart dataset, specifically using the WHS++ data. The results demonstrate that the architecture performs exceptionally well, achieving a Dice coefficient of 0.939 during testing. This indicates full topology preservation for individual structures and significantly outperforms other baselines in preserving the overall scene topology.

Chenyu Zhang, Lanjun Wang, Yueyang Cheng et al.

The rise of text-to-image (T2I) models has increasingly raised concerns regarding the generation of risky content, such as sexual, violent, and copyright-protected images, highlighting the need for effective safeguards within the models themselves. Although existing methods have been proposed to eliminate risky concepts from T2I models, they are primarily developed for earlier U-Net architectures, leaving the state-of-the-art Diffusion-Transformer-based T2I models inadequately protected. This gap stems from a fundamental architectural shift: Diffusion Transformers (DiTs) entangle semantic injection and visual synthesis via joint attention, which makes it difficult to isolate and erase risky content within the generation. To bridge this gap, we investigate how semantic concepts are represented in DiTs and discover that attention heads exhibit concept-specific sensitivity. This property enables both the detection and suppression of risky content. Building on this discovery, we propose AHV-D\&S, a training-free inference-time safeguard for image generation in DiTs. Specifically, AHV-D\&S quantifies each textual token's sensitivity across all attention heads as an Attention Head Vector (AHV), which serves as a discriminative signature for detecting risky generation tendencies. In the inference stage, we propose a momentum-based strategy to dynamically track token-wise AHVs across denoising steps, and a sensitivity-guided adaptive suppression strategy that suppresses the attention weights of identified risky tokens based on head-specific risk scores. Extensive experiments demonstrate that AHV-D\&S effectively suppresses sexual, copyrighted-style, and various harmful content while preserving visual quality, and further exhibits strong robustness against adversarial prompts and transferability across different DiT-based T2I models.

Ke Li, Chenyu Zhang, Yuxin Ding et al.

Driving scenes are inherently heterogeneous and dynamic. Multi-attribute scene identification, as a high-level visual perception capability, provides autonomous vehicles (AVs) with essential contextual awareness to understand, reason through, and interact with complex driving environments. Although scene identification is best modeled as a multi-label classification problem via multitask learning, it faces two major challenges: the difficulty of acquiring balanced, comprehensively annotated datasets and the need to re-annotate all training data when new attributes emerge. To address these challenges, this paper introduces a novel deep learning method that integrates Knowledge Acquisition and Accumulation (KAA) with Consistency-based Active Learning (CAL). KAA leverages monotask learning on heterogeneous single-label datasets to build a knowledge foundation, while CAL bridges the gap between single- and multi-label data, adapting the foundation model for multi-label scene classification. An ablation study on the newly developed Driving Scene Identification (DSI) dataset demonstrates a 56.1% improvement over an ImageNet-pretrained baseline. Moreover, KAA-CAL outperforms state-of-the-art multi-label classification methods on the BDD100K and HSD datasets, achieving this with 85% less data and even recognizing attributes unseen during foundation model training. The DSI dataset and KAA-CAL implementation code are publicly available at https://github.com/KELISBU/KAA-CAL .

Chenyu Zhang, Tairen Zhang, Lanjun Wang et al.

Using risky text prompts, such as pornography and violent prompts, to test the safety of text-to-image (T2I) models is a critical task. However, existing risky prompt datasets are limited in three key areas: 1) limited risky categories, 2) coarse-grained annotation, and 3) low effectiveness. To address these limitations, we introduce T2I-RiskyPrompt, a comprehensive benchmark designed for evaluating safety-related tasks in T2I models. Specifically, we first develop a hierarchical risk taxonomy, which consists of 6 primary categories and 14 fine-grained subcategories. Building upon this taxonomy, we construct a pipeline to collect and annotate risky prompts. Finally, we obtain 6,432 effective risky prompts, where each prompt is annotated with both hierarchical category labels and detailed risk reasons. Moreover, to facilitate the evaluation, we propose a reason-driven risky image detection method that explicitly aligns the MLLM with safety annotations. Based on T2I-RiskyPrompt, we conduct a comprehensive evaluation of eight T2I models, nine defense methods, five safety filters, and five attack strategies, offering nine key insights into the strengths and limitations of T2I model safety. Finally, we discuss potential applications of T2I-RiskyPrompt across various research fields. The dataset and code are provided in https://github.com/datar001/T2I-RiskyPrompt.

Chenyu Zhang, Lanjun Wang, Anan Liu

Recent developments in text-to-image models, particularly Stable Diffusion, have marked significant achievements in various applications. With these advancements, there are growing safety concerns about the vulnerability of the model that malicious entities exploit to generate targeted harmful images. However, the existing methods in the vulnerability of the model mainly evaluate the alignment between the prompt and generated images, but fall short in revealing the vulnerability associated with targeted image generation. In this study, we formulate the problem of targeted adversarial attack on Stable Diffusion and propose a framework to generate adversarial prompts. Specifically, we design a gradient-based embedding optimization method to craft reliable adversarial prompts that guide stable diffusion to generate specific images. Furthermore, after obtaining successful adversarial prompts, we reveal the mechanisms that cause the vulnerability of the model. Extensive experiments on two targeted attack tasks demonstrate the effectiveness of our method in targeted attacks. The code can be obtained in https://github.com/datar001/Revealing-Vulnerabilities-in-Stable-Diffusion-via-Targeted-Attacks.

DeepSeek-AI, Aixin Liu, Bei Feng et al. · stanford, tsinghua

We present DeepSeek-V3, a strong Mixture-of-Experts (MoE) language model with 671B total parameters with 37B activated for each token. To achieve efficient inference and cost-effective training, DeepSeek-V3 adopts Multi-head Latent Attention (MLA) and DeepSeekMoE architectures, which were thoroughly validated in DeepSeek-V2. Furthermore, DeepSeek-V3 pioneers an auxiliary-loss-free strategy for load balancing and sets a multi-token prediction training objective for stronger performance. We pre-train DeepSeek-V3 on 14.8 trillion diverse and high-quality tokens, followed by Supervised Fine-Tuning and Reinforcement Learning stages to fully harness its capabilities. Comprehensive evaluations reveal that DeepSeek-V3 outperforms other open-source models and achieves performance comparable to leading closed-source models. Despite its excellent performance, DeepSeek-V3 requires only 2.788M H800 GPU hours for its full training. In addition, its training process is remarkably stable. Throughout the entire training process, we did not experience any irrecoverable loss spikes or perform any rollbacks. The model checkpoints are available at https://github.com/deepseek-ai/DeepSeek-V3.

Chenyu Zhang, Minsol Kim, Shohreh Ghorbani et al.

Despite rapid growth in multimodal large language models (MLLMs), their reasoning traces remain opaque: it is often unclear which modality drives a prediction, how conflicts are resolved, or when one stream dominates. In this paper, we introduce modality sabotage, a diagnostic failure mode in which a high-confidence unimodal error overrides other evidence and misleads the fused result. To analyze such dynamics, we propose a lightweight, model-agnostic evaluation layer that treats each modality as an agent, producing candidate labels and a brief self-assessment used for auditing. A simple fusion mechanism aggregates these outputs, exposing contributors (modalities supporting correct outcomes) and saboteurs (modalities that mislead). Applying our diagnostic layer in a case study on multimodal emotion recognition benchmarks with foundation models revealed systematic reliability profiles, providing insight into whether failures may arise from dataset artifacts or model limitations. More broadly, our framework offers a diagnostic scaffold for multimodal reasoning, supporting principled auditing of fusion dynamics and informing possible interventions.

Chenyu Zhang, Han Wang, Aritra Mitra et al. · mit

Federated reinforcement learning (FRL) has emerged as a promising paradigm for reducing the sample complexity of reinforcement learning tasks by exploiting information from different agents. However, when each agent interacts with a potentially different environment, little to nothing is known theoretically about the non-asymptotic performance of FRL algorithms. The lack of such results can be attributed to various technical challenges and their intricate interplay: Markovian sampling, linear function approximation, multiple local updates to save communication, heterogeneity in the reward functions and transition kernels of the agents' MDPs, and continuous state-action spaces. Moreover, in the on-policy setting, the behavior policies vary with time, further complicating the analysis. In response, we introduce FedSARSA, a novel federated on-policy reinforcement learning scheme, equipped with linear function approximation, to address these challenges and provide a comprehensive finite-time error analysis. Notably, we establish that FedSARSA converges to a policy that is near-optimal for all agents, with the extent of near-optimality proportional to the level of heterogeneity. Furthermore, we prove that FedSARSA leverages agent collaboration to enable linear speedups as the number of agents increases, which holds for both fixed and adaptive step-size configurations.

Yihua Shao, Yan Gu, Siyu Chen et al.

Large language models (LLMs) show impressive performance in solving complex language tasks. However, its large number of parameters presents significant challenges for the deployment. So, compressing LLMs to low bits can enable to deploy on resource-constrained devices. To address this problem, we propose gradient-aware weight quantization (GWQ), the first quantization approach for low-bit weight quantization that leverages gradients to localize outliers, requiring only a minimal amount of calibration data for outlier detection. GWQ retains the top 1\% outliers preferentially at FP16 precision, while the remaining non-outlier weights are stored in a low-bit. We widely evaluate GWQ on different task include language modeling, grounding detection, massive multitask language understanding and vision-language question and answering. Results show that models quantified by GWQ performs better than other quantization method. During quantization process, GWQ only need one calibration set to realize effective quant. Also, GWQ achieves 1.2x inference speedup in comparison to the original model and effectively reduces the inference memory.

Yihua Shao, Minxi Yan, Yang Liu et al.

Low-rank Adaptation (LoRA) has demonstrated remarkable capabilities for task specific fine-tuning. However, in scenarios that involve multiple tasks, training a separate LoRA model for each one results in considerable inefficiency in terms of storage and inference. Moreover, existing parameter generation methods fail to capture the correlations among these tasks, making multi-task LoRA parameter generation challenging. To address these limitations, we propose In-Context Meta LoRA (ICM-LoRA), a novel approach that efficiently achieves task-specific customization of large language models (LLMs). Specifically, we use training data from all tasks to train a tailored generator, Conditional Variational Autoencoder (CVAE). CVAE takes task descriptions as inputs and produces task-aware LoRA weights as outputs. These LoRA weights are then merged with LLMs to create task-specialized models without the need for additional fine-tuning. Furthermore, we utilize in-context meta-learning for knowledge enhancement and task mapping, to capture the relationship between tasks and parameter distributions. As a result, our method achieves more accurate LoRA parameter generation for diverse tasks using CVAE. ICM-LoRA enables more accurate LoRA parameter reconstruction than current parameter reconstruction methods and is useful for implementing task-specific enhancements of LoRA parameters. At the same time, our method occupies 283MB, only 1\% storage compared with the original LoRA.

Fuzhong Zhou, Chenyu Zhang, Xu Chen et al. · mit

We propose a discrete time graphon game formulation on continuous state and action spaces using a representative player to study stochastic games with heterogeneous interaction among agents. This formulation admits both philosophical and mathematical advantages, compared to a widely adopted formulation using a continuum of players. We prove the existence and uniqueness of the graphon equilibrium with mild assumptions, and show that this equilibrium can be used to construct an approximate solution for finite player game on networks, which is challenging to analyze and solve due to curse of dimensionality. An online oracle-free learning algorithm is developed to solve the equilibrium numerically, and sample complexity analysis is provided for its convergence.

Ziang Xu, Bin Li, Yang Hu et al.

Accurate 3D mapping in endoscopy enables quantitative, holistic lesion characterization within the gastrointestinal (GI) tract, requiring reliable depth and pose estimation. However, endoscopy systems are monocular, and existing methods relying on synthetic datasets or complex models often lack generalizability in challenging endoscopic conditions. We propose a robust self-supervised monocular depth and pose estimation framework that incorporates a Generative Latent Bank and a Variational Autoencoder (VAE). The Generative Latent Bank leverages extensive depth scenes from natural images to condition the depth network, enhancing realism and robustness of depth predictions through latent feature priors. For pose estimation, we reformulate it within a VAE framework, treating pose transitions as latent variables to regularize scale, stabilize z-axis prominence, and improve x-y sensitivity. This dual refinement pipeline enables accurate depth and pose predictions, effectively addressing the GI tract's complex textures and lighting. Extensive evaluations on SimCol and EndoSLAM datasets confirm our framework's superior performance over published self-supervised methods in endoscopic depth and pose estimation.

Chenyu Zhang, Xu Chen, Xuan Di · mit

Mean field games (MFGs) are a promising framework for modeling the behavior of large-population systems. However, solving MFGs can be challenging due to the coupling of forward population evolution and backward agent dynamics. Typically, obtaining mean field Nash equilibria (MFNE) involves an iterative approach where the forward and backward processes are solved alternately, known as fixed-point iteration (FPI). This method requires fully observed population propagation and agent dynamics over the entire spatial domain, which could be impractical in some real-world scenarios. To overcome this limitation, this paper introduces a novel online single-agent model-free learning scheme, which enables a single agent to learn MFNE using online samples, without prior knowledge of the state-action space, reward function, or transition dynamics. Specifically, the agent updates its policy through the value function (Q), while simultaneously evaluating the mean field state (M), using the same batch of observations. We develop two variants of this learning scheme: off-policy and on-policy QM iteration. We prove that they efficiently approximate FPI, and a sample complexity guarantee is provided. The efficacy of our methods is confirmed by numerical experiments.

Yash Jangir, Yidi Zhang, Kashu Yamazaki et al.

The pursuit of robot generalists - instructable agents capable of performing diverse tasks across diverse environments - demands rigorous and scalable evaluation. Yet real-world testing of robot policies remains fundamentally constrained: it is labor-intensive, slow, unsafe at scale, and difficult to reproduce. Existing simulation benchmarks are similarly limited, as they train and test policies within the same synthetic domains and cannot assess models trained from real-world demonstrations or alternative simulation environments. As policies expand in scope and complexity, these barriers only intensify, since defining "success" in robotics often hinges on nuanced human judgments of execution quality. In this paper, we introduce a new benchmarking framework that overcomes these challenges by shifting VLA evaluation into large-scale simulated environments augmented with online human feedback. Leveraging advances in vision-language models, 2D-to-3D generative modeling, and differentiable rendering, our approach automatically converts video demonstrations from widely used robot datasets into simulated counterparts. Within these digital twins, we assess VLA policies using both automated VLM-guided scoring and scalable human preference judgments collected from crowdworkers, transforming human involvement from tedious scene setup, resetting, and safety supervision into lightweight preference comparisons. To measure robustness, we systematically perturb simulated environments along multiple axes, such as textures and object placements, stress-testing policy generalization under controlled variation. The result is a continuously evolving, reproducible, and scalable benchmark for real-world trained robot manipulation policies, addressing a critical missing capability in today's robotics landscape.

Xiaodong Qu, Andrews Damoah, Joshua Sherwood et al.

Artificial Intelligence (AI) agents have rapidly evolved from specialized, rule-based programs to versatile, learning-driven autonomous systems capable of perception, reasoning, and action in complex environments. The explosion of data, advances in deep learning, reinforcement learning, and multi-agent coordination have accelerated this transformation. Yet, designing and deploying unified AI agents that seamlessly integrate cognition, planning, and interaction remains a grand challenge. In this review, we systematically examine the architectural principles, foundational components, and emergent paradigms that define the landscape of contemporary AI agents. We synthesize insights from cognitive science-inspired models, hierarchical reinforcement learning frameworks, and large language model-based reasoning. Moreover, we discuss the pressing ethical, safety, and interpretability concerns associated with deploying these agents in real-world scenarios. By highlighting major breakthroughs, persistent challenges, and promising research directions, this review aims to guide the next generation of AI agent systems toward more robust, adaptable, and trustworthy autonomous intelligence.

Yiwen Liu, Chenyu Zhang, Junjie Song et al.

As a prominent data modality task, time series forecasting plays a pivotal role in diverse applications. With the remarkable advancements in Large Language Models (LLMs), the adoption of LLMs as the foundational architecture for time series modeling has gained significant attention. Although existing models achieve some success, they rarely both model time and frequency characteristics in a pretraining-finetuning paradigm leading to suboptimal performance in predictions of complex time series, which requires both modeling periodicity and prior pattern knowledge of signals. We propose MoFE-Time, an innovative time series forecasting model that integrates time and frequency domain features within a Mixture of Experts (MoE) network. Moreover, we use the pretraining-finetuning paradigm as our training framework to effectively transfer prior pattern knowledge across pretraining and finetuning datasets with different periodicity distributions. Our method introduces both frequency and time cells as experts after attention modules and leverages the MoE routing mechanism to construct multidimensional sparse representations of input signals. In experiments on six public benchmarks, MoFE-Time has achieved new state-of-the-art performance, reducing MSE and MAE by 6.95% and 6.02% compared to the representative methods Time-MoE. Beyond the existing evaluation benchmarks, we have developed a proprietary dataset, NEV-sales, derived from real-world business scenarios. Our method achieves outstanding results on this dataset, underscoring the effectiveness of the MoFE-Time model in practical commercial applications.

Shota Yamazaki, Chenyu Zhang, Takuya Nanri et al.

End-to-end style autonomous driving models have been developed recently. These models lack interpretability of decision-making process from perception to control of the ego vehicle, resulting in anxiety for passengers. To alleviate it, it is effective to build a model which outputs captions describing future behaviors of the ego vehicle and their reason. However, the existing approaches generate reasoning text that inadequately reflects the future plans of the ego vehicle, because they train models to output captions using momentary control signals as inputs. In this study, we propose a reasoning model that takes future planning trajectories of the ego vehicle as inputs to solve this limitation with the dataset newly collected.

Jiajun Hou, Chenyu Zhang, Rui Meng

Entity Linking (EL), the task of mapping textual entity mentions to their corresponding entries in knowledge bases, constitutes a fundamental component of natural language understanding. Recent advancements in Large Language Models (LLMs) have demonstrated remarkable potential for enhancing EL performance. Prior research has leveraged LLMs to improve entity disambiguation and input representation, yielding significant gains in accuracy and robustness. However, these approaches typically apply LLMs to isolated stages of the EL task, failing to fully integrate their capabilities throughout the entire process. In this work, we introduce DeepEL, a comprehensive framework that incorporates LLMs into every stage of the entity linking task. Furthermore, we identify that disambiguating entities in isolation is insufficient for optimal performance. To address this limitation, we propose a novel self-validation mechanism that utilizes global contextual information, enabling LLMs to rectify their own predictions and better recognize cohesive relationships among entities within the same sentence. Extensive empirical evaluation across ten benchmark datasets demonstrates that DeepEL substantially outperforms existing state-of-the-art methods, achieving an average improvement of 2.6\% in overall F1 score and a remarkable 4% gain on out-of-domain datasets. These results underscore the efficacy of deep LLM integration in advancing the state-of-the-art in entity linking.

Chenyu Zhang, Xiaohang Luo

Formative feedback is widely recognized as one of the most effective drivers of student learning, yet it remains difficult to implement equitably at scale. In large or low-resource courses, instructors often lack the time, staffing, and bandwidth required to review and respond to every student reflection, creating gaps in support precisely where learners would benefit most. This paper presents a theory-grounded system that uses five coordinated role-based LLM agents (Evaluator, Equity Monitor, Metacognitive Coach, Aggregator, and Reflexion Reviewer) to score learner reflections with a shared rubric and to generate short, bias-aware, learner-facing comments. The agents first produce structured rubric scores, then check for potentially biased or exclusionary language, add metacognitive prompts that invite students to think about their own thinking, and finally compose a concise feedback message of at most 120 words. The system includes simple fairness checks that compare scoring error across lower and higher scoring learners, enabling instructors to monitor and bound disparities in accuracy. We evaluate the pipeline in a 12-session AI literacy program with adult learners. In this setting, the system produces rubric scores that approach expert-level agreement, and trained graders rate the AI-generated comments as helpful, empathetic, and well aligned with instructional goals. Taken together, these results show that multi-agent LLM systems can deliver equitable, high-quality formative feedback at a scale and speed that would be impossible for human graders alone. More broadly, the work points toward a future where feedback-rich learning becomes feasible for any course size or context, advancing long-standing goals of equity, access, and instructional capacity in education.

Chenyu Zhang, Navid Azizan · mit

Multi-agent learning faces a fundamental tension: leveraging distributed collaboration without sacrificing the personalization needed for diverse agents. This tension intensifies when aiming for full personalization while adapting to unknown heterogeneity levels -- gaining collaborative speedup when agents are similar, without performance degradation when they are different. Embracing the challenge, we propose personalized collaborative learning (PCL), a novel framework for heterogeneous agents to collaboratively learn personalized solutions with seamless adaptivity. Through carefully designed bias correction and importance correction mechanisms, our method AffPCL robustly handles both environment and objective heterogeneity. We prove that AffPCL reduces sample complexity over independent learning by a factor of $\max\{n^{-1}, δ\}$, where $n$ is the number of agents and $δ\in[0,1]$ measures their heterogeneity. This affinity-based acceleration automatically interpolates between the linear speedup of federated learning in homogeneous settings and the baseline of independent learning, without requiring prior knowledge of the system. Our analysis further reveals that an agent may obtain linear speedup even by collaborating with arbitrarily dissimilar agents, unveiling new insights into personalization and collaboration in the high heterogeneity regime.

Chenyu Zhang, Sharifa Alghowinem, Cynthia Breazeal

While recent studies have examined the leaning impact of large language model (LLM) in educational contexts, the affective dynamics of LLM-mediated tutoring remain insufficiently understood. This work introduces the first ensemble-LLM framework for large-scale affect sensing in tutoring dialogues, advancing the conversation on responsible pathways for integrating generative AI into education by attending to learners' evolving affective states. To achieve this, we analyzed two semesters' worth of 16,986 conversational turns exchanged between PyTutor, an LLM-powered AI tutor, and 261 undergraduate learners across three U.S. institutions. To investigate learners' emotional experiences, we generate zero-shot affect annotations from three frontier LLMs (Gemini, GPT-4o, Claude), including scalar ratings of valence, arousal, and learning-helpfulness, along with free-text emotion labels. These estimates are fused through rank-weighted intra-model pooling and plurality consensus across models to produce robust emotion profiles. Our analysis shows that during interaction with the AI tutor, students typically report mildly positive affect and moderate arousal. Yet learning is not uniformly smooth: confusion and curiosity are frequent companions to problem solving, and frustration, while less common, still surfaces in ways that can derail progress. Emotional states are short-lived--positive moments last slightly longer than neutral or negative ones, but they are fragile and easily disrupted. Encouragingly, negative emotions often resolve quickly, sometimes rebounding directly into positive states. Neutral moments frequently act as turning points, more often steering students upward than downward, suggesting opportunities for tutors to intervene at precisely these junctures.

Yin Sun, Yiwen Liu, Junjie Song et al.

In the area of commercial auto sales system, high-quality lead score sequencing determines the priority of a sale's work and is essential for optimizing the efficiency of the sales system. Since CRM (Customer Relationship Management) system contains plenty of textual interaction features between sales and customers, traditional techniques such as Click Through Rate (CTR) prediction struggle with processing the complex information inherent in natural language features, which limits their effectiveness in sales lead ranking. Bridging this gap is critical for enhancing business intelligence and decision-making. Recently, the emergence of large language models (LLMs) has opened new avenues for improving recommendation systems, this study introduces asLLR (LLM-based Leads Ranking in Auto Sales), which integrates CTR loss and Question Answering (QA) loss within a decoder-only large language model architecture. This integration enables the simultaneous modeling of both tabular and natural language features. To verify the efficacy of asLLR, we constructed an innovative dataset derived from the customer lead pool of a prominent new energy vehicle brand, with 300,000 training samples and 40,000 testing samples. Our experimental results demonstrate that asLLR effectively models intricate patterns in commercial datasets, achieving the AUC of 0.8127, surpassing traditional CTR estimation methods by 0.0231. Moreover, asLLR enhances CTR models when used for extracting text features by 0.0058. In real-world sales scenarios, after rigorous online A/B testing, asLLR increased the sales volume by about 9.5% compared to the traditional method, providing a valuable tool for business intelligence and operational decision-making.

Chenyu Zhang, Yuanhang Wu, Xiaoxi Ma et al.

Detection and tracking are the basic tasks of radar systems. Current joint detection tracking methods, which focus on dynamically adjusting detection thresholds from tracking results, still present challenges in fully utilizing the potential of radar signals. These are mainly reflected in the limited capacity of the constant false-alarm rate model to accurately represent information, the insufficient depiction of complex scenes, and the limited information acquired by the tracker. We introduce the Integrated Detection and Tracking based on radar feature (InDT) method, which comprises a network architecture for radar signal detection and a tracker that leverages detection assistance. The InDT detector extracts feature information from each Range-Doppler (RD) matrix and then returns the target position through the feature enhancement module and the detection head. The InDT tracker adaptively updates the measurement noise covariance of the Kalman filter based on detection confidence. The similarity of target RD features is measured by cosine distance, which enhances the data association process by combining location and feature information. Finally, the efficacy of the proposed method was validated through testing on both simulated data and publicly available datasets.

Chenyu Zhang, Yihao Luo, Yinzhe Wu et al.

Accurate cardiac mesh reconstruction from volumetric data is essential for personalized cardiac modeling and clinical analysis. However, existing deformation-based approaches are prone to topological inconsistencies, particularly membrane penetration, which undermines the anatomical plausibility of the reconstructed mesh. To address this issue, we introduce Topology-Preserving Mesh Loss (TPM Loss), a novel loss function that explicitly enforces topological constraints during mesh deformation. By identifying topology-violating points, TPM Loss ensures spatially consistent reconstructions. Extensive experiments on CT and MRI datasets show that TPM Loss reduces topology violations by up to 93.1% while maintaining high segmentation accuracy (DSC: 89.1%-92.9%) and improving mesh fidelity (Chamfer Distance reduction up to 0.26 mm). These results demonstrate that TPM Loss effectively prevents membrane penetration and significantly improves cardiac mesh quality, enabling more accurate and anatomically consistent cardiac reconstructions.

Ziliang Zhang, Huaming Yu, Danqin Ren et al.

This study presents OceanCastNet (OCN), a machine learning approach for wave forecasting that incorporates wind and wave fields to predict significant wave height, mean wave period, and mean wave direction.We evaluate OCN's performance against the operational ECWAM model using two independent datasets: NDBC buoy and Jason-3 satellite observations. NDBC station validation indicates OCN performs better at 24 stations compared to ECWAM's 10 stations, and Jason-3 satellite validation confirms similar accuracy across 228-hour forecasts. OCN successfully captures wave patterns during extreme weather conditions, demonstrated through Typhoon Goni with prediction errors typically within $\pm$0.5 m. The approach also offers computational efficiency advantages. The results suggest that machine learning approaches can achieve performance comparable to conventional wave forecasting systems for operational wave prediction applications.