Ziyang Chen

Yiwen Ye, Yutong Xie, Jianpeng Zhang et al.

The universal model emerges as a promising trend for medical image segmentation, paving up the way to build medical imaging large model (MILM). One popular strategy to build universal models is to encode each task as a one-hot vector and generate dynamic convolutional layers at the end of the decoder to extract the interested target. Although successful, it ignores the correlations among tasks and meanwhile is too late to make the model 'aware' of the ongoing task. To address both issues, we propose a prompt-driven Universal Segmentation model (UniSeg) for multi-task medical image segmentation using diverse modalities and domains. We first devise a learnable universal prompt to describe the correlations among all tasks and then convert this prompt and image features into a task-specific prompt, which is fed to the decoder as a part of its input. Thus, we make the model 'aware' of the ongoing task early and boost the task-specific training of the whole decoder. Our results indicate that the proposed UniSeg outperforms other universal models and single-task models on 11 upstream tasks. Moreover, UniSeg also beats other pre-trained models on two downstream datasets, providing the community with a high-quality pre-trained model for 3D medical image segmentation. Code and model are available at https://github.com/yeerwen/UniSeg.

Dongfang Li, Zetian Sun, Xinshuo Hu et al.

Open-domain generative systems have gained significant attention in the field of conversational AI (e.g., generative search engines). This paper presents a comprehensive review of the attribution mechanisms employed by these systems, particularly large language models. Though attribution or citation improve the factuality and verifiability, issues like ambiguous knowledge reservoirs, inherent biases, and the drawbacks of excessive attribution can hinder the effectiveness of these systems. The aim of this survey is to provide valuable insights for researchers, aiding in the refinement of attribution methodologies to enhance the reliability and veracity of responses generated by open-domain generative systems. We believe that this field is still in its early stages; hence, we maintain a repository to keep track of ongoing studies at https://github.com/HITsz-TMG/awesome-llm-attributions.

Yiwen Ye, Yutong Xie, Jianpeng Zhang et al.

Self-supervised learning is an efficient pre-training method for medical image analysis. However, current research is mostly confined to specific-modality data pre-training, consuming considerable time and resources without achieving universality across different modalities. A straightforward solution is combining all modality data for joint self-supervised pre-training, which poses practical challenges. Firstly, our experiments reveal conflicts in representation learning as the number of modalities increases. Secondly, multi-modal data collected in advance cannot cover all real-world scenarios. In this paper, we reconsider versatile self-supervised learning from the perspective of continual learning and propose MedCoSS, a continuous self-supervised learning approach for multi-modal medical data. Unlike joint self-supervised learning, MedCoSS assigns different modality data to different training stages, forming a multi-stage pre-training process. To balance modal conflicts and prevent catastrophic forgetting, we propose a rehearsal-based continual learning method. We introduce the k-means sampling strategy to retain data from previous modalities and rehearse it when learning new modalities. Instead of executing the pretext task on buffer data, a feature distillation strategy and an intra-modal mixup strategy are applied to these data for knowledge retention. We conduct continuous self-supervised pre-training on a large-scale multi-modal unlabeled dataset, including clinical reports, X-rays, CT scans, MRI scans, and pathological images. Experimental results demonstrate MedCoSS's exceptional generalization ability across nine downstream datasets and its significant scalability in integrating new modality data. Code and pre-trained weight are available at https://github.com/yeerwen/MedCoSS.

Ziyang Chen, Dongfang Li, Xiang Zhao et al.

In this study, we address the challenge of enhancing temporal knowledge reasoning in Large Language Models (LLMs). LLMs often struggle with this task, leading to the generation of inaccurate or misleading responses. This issue mainly arises from their limited ability to handle evolving factual knowledge and complex temporal logic. To overcome these limitations, we propose Abstract Reasoning Induction (ARI) framework, which divides temporal reasoning into two distinct phases: Knowledge-agnostic and Knowledge-based. This framework offers factual knowledge support to LLMs while minimizing the incorporation of extraneous noisy data. Concurrently, informed by the principles of constructivism, ARI provides LLMs the capability to engage in proactive, self-directed learning from both correct and incorrect historical reasoning samples. By teaching LLMs to actively construct knowledge and methods, it can significantly boosting their temporal reasoning abilities. Our approach achieves remarkable improvements, with relative gains of 29.7% and 9.27% on two temporal QA datasets, underscoring its efficacy in advancing temporal reasoning in LLMs. The code can be found at https://github.com/czy1999/ARI-QA

Ziyang Chen, Shaoguang Wang, Weiyu Guo et al.

Vision-Language-Action (VLA) models have achieved remarkable success in language-conditioned robotic manipulation. However, deploying these models in open-ended environments requires continuously acquiring novel skills, a process that inevitably triggers severe catastrophic forgetting of previously learned behaviors. While experience replay (ER) serves as a standard mitigating strategy, naive uniform sampling fundamentally misaligns with the temporal characteristics of manipulation trajectories. It systematically under-samples brief but causally critical sub-skills, leading to phase starvation, and completely overlooks the varying degrees of forgetting across historical tasks. To overcome these limitations, we introduce PHASER, an architecture-agnostic continual learning framework. PHASER employs a phase-centric capacity allocation to guarantee equal memory support for all sub-skills, coupled with a multi-modal interference routing strategy that dynamically prioritizes historical phases at high risk of forgetting. Furthermore, to enable fully autonomous lifelong adaptation, we integrate Auto-PC, a lightweight pipeline combining unsupervised action-signal change-point detection with VLM-based semantic verification to extract temporal boundaries without intensive manual supervision. Evaluated across three VLA backbones on LIBERO continual learning suites, PHASER yields substantial empirical improvements, increasing Average Success Rate (ASR) by up to 31% over matched-budget ER and achieving an 87.8% final ASR on the LIBERO-Goal CL setting.

Ziyang Chen, Yongsheng Pan, Yiwen Ye et al.

Distribution shift widely exists in medical images acquired from different medical centres and poses a significant obstacle to deploying the pre-trained semantic segmentation model in real-world applications. Test-time adaptation has proven its effectiveness in tackling the cross-domain distribution shift during inference. However, most existing methods achieve adaptation by updating the pre-trained models, rendering them susceptible to error accumulation and catastrophic forgetting when encountering a series of distribution shifts (i.e., under the continual test-time adaptation setup). To overcome these challenges caused by updating the models, in this paper, we freeze the pre-trained model and propose the Visual Prompt-based Test-Time Adaptation (VPTTA) method to train a specific prompt for each test image to align the statistics in the batch normalization layers. Specifically, we present the low-frequency prompt, which is lightweight with only a few parameters and can be effectively trained in a single iteration. To enhance prompt initialization, we equip VPTTA with a memory bank to benefit the current prompt from previous ones. Additionally, we design a warm-up mechanism, which mixes source and target statistics to construct warm-up statistics, thereby facilitating the training process. Extensive experiments demonstrate the superiority of our VPTTA over other state-of-the-art methods on two medical image segmentation benchmark tasks. The code and weights of pre-trained source models are available at https://github.com/Chen-Ziyang/VPTTA.

Yuexi Du, Ziyang Chen, Justin Salamon et al.

The sound effects that designers add to videos are designed to convey a particular artistic effect and, thus, may be quite different from a scene's true sound. Inspired by the challenges of creating a soundtrack for a video that differs from its true sound, but that nonetheless matches the actions occurring on screen, we propose the problem of conditional Foley. We present the following contributions to address this problem. First, we propose a pretext task for training our model to predict sound for an input video clip using a conditional audio-visual clip sampled from another time within the same source video. Second, we propose a model for generating a soundtrack for a silent input video, given a user-supplied example that specifies what the video should "sound like". We show through human studies and automated evaluation metrics that our model successfully generates sound from video, while varying its output according to the content of a supplied example. Project site: https://xypb.github.io/CondFoleyGen/

Nuo Chen, Yan Wang, Haiyun Jiang et al.

In recent years, Dialogue-style Large Language Models (LLMs) such as ChatGPT and GPT4 have demonstrated immense potential in constructing open-domain dialogue agents. However, aligning these agents with specific characters or individuals remains a considerable challenge due to the complexities of character representation and the lack of comprehensive annotations. In this paper, we introduce the Harry Potter Dialogue (HPD) dataset, designed to advance the study of dialogue agents and character alignment. The dataset encompasses all dialogue sessions (in both English and Chinese) from the Harry Potter series and is annotated with vital background information, including dialogue scenes, speakers, character relationships, and attributes. These extensive annotations may empower LLMs to unlock character-driven dialogue capabilities. Furthermore, it can serve as a universal benchmark for evaluating how well can a LLM aligning with a specific character. We benchmark LLMs on HPD using both fine-tuning and in-context learning settings. Evaluation results reveal that although there is substantial room for improvement in generating high-quality, character-aligned responses, the proposed dataset is valuable in guiding models toward responses that better align with the character of Harry Potter.

Chao Feng, Ziyang Chen, Andrew Owens

Manipulated videos often contain subtle inconsistencies between their visual and audio signals. We propose a video forensics method, based on anomaly detection, that can identify these inconsistencies, and that can be trained solely using real, unlabeled data. We train an autoregressive model to generate sequences of audio-visual features, using feature sets that capture the temporal synchronization between video frames and sound. At test time, we then flag videos that the model assigns low probability. Despite being trained entirely on real videos, our model obtains strong performance on the task of detecting manipulated speech videos. Project site: https://cfeng16.github.io/audio-visual-forensics

Ziyang Chen, Yiwen Ye, Yongsheng Pan et al.

Although recent years have witnessed significant advancements in medical image segmentation, the pervasive issue of domain shift among medical images from diverse centres hinders the effective deployment of pre-trained models. Many Test-time Adaptation (TTA) methods have been proposed to address this issue by fine-tuning pre-trained models with test data during inference. These methods, however, often suffer from less-satisfactory optimization due to suboptimal optimization direction (dictated by the gradient) and fixed step-size (predicated on the learning rate). In this paper, we propose the Gradient alignment-based Test-time adaptation (GraTa) method to improve both the gradient direction and learning rate in the optimization procedure. Unlike conventional TTA methods, which primarily optimize the pseudo gradient derived from a self-supervised objective, our method incorporates an auxiliary gradient with the pseudo one to facilitate gradient alignment. Such gradient alignment enables the model to excavate the similarities between different gradients and correct the gradient direction to approximate the empirical gradient related to the current segmentation task. Additionally, we design a dynamic learning rate based on the cosine similarity between the pseudo and auxiliary gradients, thereby empowering the adaptive fine-tuning of pre-trained models on diverse test data. Extensive experiments establish the effectiveness of the proposed gradient alignment and dynamic learning rate and substantiate the superiority of our GraTa method over other state-of-the-art TTA methods on a benchmark medical image segmentation task. The code and weights of pre-trained source models are available at https://github.com/Chen-Ziyang/GraTa.

Xixi Hu, Ziyang Chen, Andrew Owens

We present a method for simultaneously localizing multiple sound sources within a visual scene. This task requires a model to both group a sound mixture into individual sources, and to associate them with a visual signal. Our method jointly solves both tasks at once, using a formulation inspired by the contrastive random walk of Jabri et al. We create a graph in which images and separated sounds correspond to nodes, and train a random walker to transition between nodes from different modalities with high return probability. The transition probabilities for this walk are determined by an audio-visual similarity metric that is learned by our model. We show through experiments with musical instruments and human speech that our model can successfully localize multiple sounds, outperforming other self-supervised methods. Project site: https://hxixixh.github.io/mix-and-localize

Minxuan Lv, Tiehua Mei, Tanlong Du et al.

We present GoLongRL, a fully open-source, capability-oriented post-training recipe for long-context reinforcement learning with verifiable rewards (RLVR). Existing long-context RL methods often treat data construction as a matter of designing increasingly complex retrieval paths, leading to homogeneous task coverage and reward formulations that inadequately reflect practical long-context requirements. Our work offers two contributions. (1) Capability-oriented data construction with full open release. We openly release a dataset of 23K RLVR samples, the complete construction pipeline, and all training code. Guided by a taxonomy of long-context capabilities, the dataset spans 9 task types, each paired with its natural evaluation metric. It comprises curated open-source samples from established corpora and synthetic samples whose QA pairs are generated from real source documents such as books, academic papers, and multi-turn dialogues. Under the same vanilla GRPO setup, our dataset alone outperforms the closed-source QwenLong-L1.5 dataset. Moreover, our Qwen3-30B-A3B model trained on this data delivers long-context performance comparable to DeepSeek-R1-0528 and Qwen3-235B-A22B-Thinking-2507, suggesting that broader coverage and greater reward diversity substantially benefit long-context capability improvement. (2) TMN-Reweight for heterogeneous multitask optimization. To address optimization challenges from heterogeneous rewards, we propose TMN-Reweight, which combines task-level mean normalization for cross-task reward scale alignment with difficulty-adaptive weighting for more reliable advantage estimation. TMN-Reweight further improves average performance over vanilla GRPO, with general capabilities preserved or improved across reported evaluations.

Ziyang Chen, David F. Fouhey, Andrew Owens

Sounds reach one microphone in a stereo pair sooner than the other, resulting in an interaural time delay that conveys their directions. Estimating a sound's time delay requires finding correspondences between the signals recorded by each microphone. We propose to learn these correspondences through self-supervision, drawing on recent techniques from visual tracking. We adapt the contrastive random walk of Jabri et al. to learn a cycle-consistent representation from unlabeled stereo sounds, resulting in a model that performs on par with supervised methods on "in the wild" internet recordings. We also propose a multimodal contrastive learning model that solves a visually-guided localization task: estimating the time delay for a particular person in a multi-speaker mixture, given a visual representation of their face. Project site: https://ificl.github.io/stereocrw/

Ziyang Chen, Shengyi Qian, Andrew Owens

The images and sounds that we perceive undergo subtle but geometrically consistent changes as we rotate our heads. In this paper, we use these cues to solve a problem we call Sound Localization from Motion (SLfM): jointly estimating camera rotation and localizing sound sources. We learn to solve these tasks solely through self-supervision. A visual model predicts camera rotation from a pair of images, while an audio model predicts the direction of sound sources from binaural sounds. We train these models to generate predictions that agree with one another. At test time, the models can be deployed independently. To obtain a feature representation that is well-suited to solving this challenging problem, we also propose a method for learning an audio-visual representation through cross-view binauralization: estimating binaural sound from one view, given images and sound from another. Our model can successfully estimate accurate rotations on both real and synthetic scenes, and localize sound sources with accuracy competitive with state-of-the-art self-supervised approaches. Project site: https://ificl.github.io/SLfM/

Open-H-Embodiment Consortium, Nigel Nelson, Juo-Tung Chen et al.

Autonomous medical robots hold promise to improve patient outcomes, reduce provider workload, democratize access to care, and enable superhuman precision. However, autonomous medical robotics has been limited by a fundamental data problem: existing medical robotic datasets are small, single-embodiment, and rarely shared openly, restricting the development of foundation models that the field needs to advance. We introduce Open-H-Embodiment, the largest open dataset of medical robotic video with synchronized kinematics to date, spanning more than 49 institutions and multiple robotic platforms including the CMR Versius, Intuitive Surgical's da Vinci, da Vinci Research Kit (dVRK), Rob Surgical BiTrack, Virtual Incision's MIRA, Moon Surgical Maestro, and a variety of custom systems, spanning surgical manipulation, robotic ultrasound, and endoscopy procedures. We demonstrate the research enabled by this dataset through two foundation models. GR00T-H is the first open foundation vision-language-action model for medical robotics, which is the only evaluated model to achieve full end-to-end task completion on a structured suturing benchmark (25% of trials vs. 0% for all others) and achieves 64% average success across a 29-step ex vivo suturing sequence. We also train Cosmos-H-Surgical-Simulator, the first action-conditioned world model to enable multi-embodiment surgical simulation from a single checkpoint, spanning nine robotic platforms and supporting in silico policy evaluation and synthetic data generation for the medical domain. These results suggest that open, large-scale medical robot data collection can serve as critical infrastructure for the research community, enabling advances in robot learning, world modeling, and beyond.

Ziyang Chen, Yongsheng Pan, Yong Xia

Glaucoma is one of the leading causes of irreversible blindness. Segmentation of optic disc (OD) and optic cup (OC) on fundus images is a crucial step in glaucoma screening. Although many deep learning models have been constructed for this task, it remains challenging to train an OD/OC segmentation model that could be deployed successfully to different healthcare centers. The difficulties mainly comes from the domain shift issue, i.e., the fundus images collected at these centers usually vary greatly in the tone, contrast, and brightness. To address this issue, in this paper, we propose a novel unsupervised domain adaptation (UDA) method called Reconstruction-driven Dynamic Refinement Network (RDR-Net), where we employ a due-path segmentation backbone for simultaneous edge detection and region prediction and design three modules to alleviate the domain gap. The reconstruction alignment (RA) module uses a variational auto-encoder (VAE) to reconstruct the input image and thus boosts the image representation ability of the network in a self-supervised way. It also uses a style-consistency constraint to force the network to retain more domain-invariant information. The low-level feature refinement (LFR) module employs input-specific dynamic convolutions to suppress the domain-variant information in the obtained low-level features. The prediction-map alignment (PMA) module elaborates the entropy-driven adversarial learning to encourage the network to generate source-like boundaries and regions. We evaluated our RDR-Net against state-of-the-art solutions on four public fundus image datasets. Our results indicate that RDR-Net is superior to competing models in both segmentation performance and generalization ability

Chanwoo Park, Ziyang Chen, Asuman Ozdaglar et al.

Large language models (LLMs) are increasingly deployed as "agents" for decision-making (DM) in interactive and dynamic environments. Yet, since they were not originally designed for DM, recent studies show that LLMs can struggle even in basic online DM problems, failing to achieve low regret or an effective exploration-exploitation tradeoff. To address this, we introduce Iterative Regret-Minimization Fine-Tuning (Iterative RMFT), a post-training procedure that repeatedly distills low-regret decision trajectories back into the base model. At each iteration, the model rolls out multiple decision trajectories, selects the k-lowest regret ones, and fine-tunes itself on them. Unlike prior methods that (a) distill action sequences from known DM algorithms or (b) rely on manually crafted chain-of-thought templates, our approach leverages the regret metric to elicit the model's own DM ability and reasoning rationales. This reliance on model-generated reasoning avoids rigid output engineering and provides more flexible, natural-language training signals. Empirical results show that Iterative RMFT improves LLMs' DM performance across diverse models - from Transformers with numerical input/output, to open-weight LLMs, and advanced closed-weight models like GPT-4o mini. Its flexibility in output and reasoning formats enables generalization across tasks with varying horizons, action spaces, reward processes, and natural-language contexts. Finally, we provide theoretical insight showing that a single-layer Transformer under this paradigm can act as a no-regret learner in a simplified setting. Overall, Iterative RMFT offers a principled and general post-training framework for enhancing LLMs' decision-making capabilities.

Qingyuan Zeng, Ziyang Chen, Pengxiang Cai et al.

Biomedical discovery often requires connecting broad biomedical knowledge with specific experimental or clinical data. Background knowledge suggests relevant mechanisms but is usually too general to map directly onto dataset variables, while data-driven patterns can be dataset-specific and hard to interpret mechanistically. We study this missing link as knowledge contextualization: transforming broad biomedical knowledge into evidence-supported, scenario-grounded propositions that domain experts can inspect, replay, and validate. We propose SCENE, a bi-level multi-agent framework that treats knowledge contextualization as iterative search. The upper level converts broad knowledge into search directions and grounds them in the dataset schema. The lower level executes these directions through multi-objective optimization to identify concrete propositions that balance evidential strength and data support. Feedback between the two levels progressively refines the search. We evaluate SCENE in two settings: discovering patient subgroups with heterogeneous treatment benefits in clinical trial scenarios, and identifying context-specific biological responses in LINCS L1000 studies. In clinical trials, SCENE discovers specific, well-supported subgroups and outperforms existing baselines. In L1000 studies, SCENE identifies perturbational contexts with strong target-response matching and high positive rates. These results show that SCENE bridges broad knowledge and scenario-specific evidence, producing traceable, inspectable hypotheses for follow-up validation.

Qingyuan Zeng, Pengxiang Cai, Zixin Guan et al.

Designing a single molecule that modulates two targets is a promising strategy for polypharmacology, but it remains substantially harder than standard single-target generation because one candidate must satisfy two binding requirements while preserving drug-likeness and synthesizability. Existing dual-target generative methods typically introduce dual-target capability by either retraining the generator or intervening in the diffusion process during sampling. The former can be costly and difficult to stabilize when dual-target supervision is sparse, while the latter may be sensitive to denoising-time target balancing and competing update directions. These limitations motivate a generator-preserving alternative that keeps the pretrained prior intact: can dual-target candidates instead be recovered from the input space of a frozen single-target diffusion model, without modifying its parameters or denoising dynamics? We formulate this task as a constrained multi-objective optimization problem and propose REUSE, a hierarchical evolutionary input-space search framework that combines pair-conditioned exploration with structured multi-stage selection to enforce dual-target affinity, chemical quality, and diversity. Experiments show that, compared with methods that modify the diffusion process, REUSE consistently improves dual-target affinity and balance, achieving a 20.9-percentage-point gain in Dual High Affinity over the strongest prior baseline while maintaining competitive molecular quality.

Zhangli Zhou, Shaochen Wang, Ziyang Chen et al.

High-resolution representations are important for vision-based robotic grasping problems. Existing works generally encode the input images into low-resolution representations via sub-networks and then recover high-resolution representations. This will lose spatial information, and errors introduced by the decoder will be more serious when multiple types of objects are considered or objects are far away from the camera. To address these issues, we revisit the design paradigm of CNN for robotic perception tasks. We demonstrate that using parallel branches as opposed to serial stacked convolutional layers will be a more powerful design for robotic visual grasping tasks. In particular, guidelines of neural network design are provided for robotic perception tasks, e.g., high-resolution representation and lightweight design, which respond to the challenges in different manipulation scenarios. We then develop a novel grasping visual architecture referred to as HRG-Net, a parallel-branch structure that always maintains a high-resolution representation and repeatedly exchanges information across resolutions. Extensive experiments validate that these two designs can effectively enhance the accuracy of visual-based grasping and accelerate network training. We show a series of comparative experiments in real physical environments at Youtube: https://youtu.be/Jhlsp-xzHFY.

Qiaoling Yang, Linkai Luo, Haoyu Zhang et al.

Support vector machine (SVM) and neural networks (NN) have strong complementarity. SVM focuses on the inner operation among samples while NN focuses on the operation among the features within samples. Thus, it is promising and attractive to combine SVM and NN, as it may provide a more powerful function than SVM or NN alone. However, current work on combining them lacks true integration. To address this, we propose a sample attention memory network (SAMN) that effectively combines SVM and NN by incorporating sample attention module, class prototypes, and memory block to NN. SVM can be viewed as a sample attention machine. It allows us to add a sample attention module to NN to implement the main function of SVM. Class prototypes are representatives of all classes, which can be viewed as alternatives to support vectors. The memory block is used for the storage and update of class prototypes. Class prototypes and memory block effectively reduce the computational cost of sample attention and make SAMN suitable for multi-classification tasks. Extensive experiments show that SAMN achieves better classification performance than single SVM or single NN with similar parameter sizes, as well as the previous best model for combining SVM and NN. The sample attention mechanism is a flexible module that can be easily deepened and incorporated into neural networks that require it.

Anhao Zhao, Ziyang Chen, Junlong Tong et al.

Large reasoning models (LRMs) are commonly trained with reinforcement learning (RL) to explore long chain-of-thought reasoning, achieving strong performance at high computational cost. Recent methods add multi-reward objectives to jointly optimize correctness and brevity, but these complex extensions often destabilize training and yield suboptimal trade-offs. We revisit this objective and challenge the necessity of such complexity. Through principled analysis, we identify fundamental misalignments in this paradigm: KL regularization loses its intended role when correctness and length are directly verifiable, and group-wise normalization becomes ambiguous under multiple reward signals. By removing these two items and simplifying the reward to a truncation-based length penalty, we show that the optimization problem reduces to supervised fine-tuning on self-generated data filtered for both correctness and conciseness. We term this simplified training strategy on-policy SFT. Despite its simplicity, on-policy SFT consistently defines the accuracy-efficiency Pareto frontier. It reduces CoT length by up to 80 while maintaining original accuracy, surpassing more complex RL-based methods across five benchmarks. Furthermore, it significantly enhances training efficiency, reducing GPU memory usage by 50% and accelerating convergence by 70%. Our code is available at https://github.com/EIT-NLP/On-Policy-SFT.

Longwei Ding, Anhao Zhao, Fanghua Ye et al.

Large language models (LLMs) are increasingly costly to deploy, motivating extensive research on model pruning. However, most existing studies focus on instruction-following LLMs, leaving it unclear whether established pruning strategies transfer to reasoning-augmented models that explicitly generate long intermediate reasoning traces. In this work, we conduct a controlled study of pruning for both instruction-following ($\textbf{LLM-instruct}$) and reasoning-augmented ($\textbf{LLM-think}$) models. To isolate the effects of pruning, we align pruning calibration and post-pruning recovery data with each model's original training distribution, which we show yields more stable and reliable pruning behavior. We evaluate static depth pruning, static width pruning, and dynamic pruning across 17 tasks spanning classification, generation, and reasoning. Our results reveal clear paradigm-dependent differences: depth pruning outperforms width pruning on classification tasks, while width pruning is more robust for generation and reasoning. Moreover, static pruning better preserves reasoning performance, whereas dynamic pruning excels on classification and generation but remains challenging for long-chain reasoning. These findings underscore the need for pruning strategies that explicitly account for the distinct characteristics of reasoning-augmented LLMs. Our code is publicly available at https://github.com/EIT-NLP/LRM-Pruning.

Ziyang Chen, Wei Long, He Yao et al.

Learning-based stereo matching techniques have made significant progress. However, existing methods inevitably lose geometrical structure information during the feature channel generation process, resulting in edge detail mismatches. In this paper, the Motif Cha}nnel Attention Stereo Matching Network (MoCha-Stereo) is designed to address this problem. We provide the Motif Channel Correlation Volume (MCCV) to determine more accurate edge matching costs. MCCV is achieved by projecting motif channels, which capture common geometric structures in feature channels, onto feature maps and cost volumes. In addition, edge variations in %potential feature channels of the reconstruction error map also affect details matching, we propose the Reconstruction Error Motif Penalty (REMP) module to further refine the full-resolution disparity estimation. REMP integrates the frequency information of typical channel features from the reconstruction error. MoCha-Stereo ranks 1st on the KITTI-2015 and KITTI-2012 Reflective leaderboards. Our structure also shows excellent performance in Multi-View Stereo. Code is avaliable at https://github.com/ZYangChen/MoCha-Stereo.

Linkai Luo, Qiaoling Yang, Hong Peng et al.

The selection of Gaussian kernel parameters plays an important role in the applications of support vector classification (SVC). A commonly used method is the k-fold cross validation with grid search (CV), which is extremely time-consuming because it needs to train a large number of SVC models. In this paper, a new approach is proposed to train SVC and optimize the selection of Gaussian kernel parameters. We first formulate the training and the parameter selection of SVC as a minimax optimization problem named as MaxMin-L2-SVC-NCH, in which the minimization problem is an optimization problem of finding the closest points between two normal convex hulls (L2-SVC-NCH) while the maximization problem is an optimization problem of finding the optimal Gaussian kernel parameters. A lower time complexity can be expected in MaxMin-L2-SVC-NCH because CV is not needed. We then propose a projected gradient algorithm (PGA) for the training of L2-SVC-NCH. It is revealed that the famous sequential minimal optimization (SMO) algorithm is a special case of the PGA. Thus, the PGA can provide more flexibility than the SMO. Furthermore, the solution of the maximization problem is done by a gradient ascent algorithm with dynamic learning rate. The comparative experiments between MaxMin-L2-SVC-NCH and the previous best approaches on public datasets show that MaxMin-L2-SVC-NCH greatly reduces the number of models to be trained while maintaining competitive test accuracy. These findings indicate that MaxMin-L2-SVC-NCH is a better choice for SVC tasks.

Pedro R. A. S. Bassi, Wenxuan Li, Yucheng Tang et al.

How can we test AI performance? This question seems trivial, but it isn't. Standard benchmarks often have problems such as in-distribution and small-size test sets, oversimplified metrics, unfair comparisons, and short-term outcome pressure. As a consequence, good performance on standard benchmarks does not guarantee success in real-world scenarios. To address these problems, we present Touchstone, a large-scale collaborative segmentation benchmark of 9 types of abdominal organs. This benchmark is based on 5,195 training CT scans from 76 hospitals around the world and 5,903 testing CT scans from 11 additional hospitals. This diverse test set enhances the statistical significance of benchmark results and rigorously evaluates AI algorithms across various out-of-distribution scenarios. We invited 14 inventors of 19 AI algorithms to train their algorithms, while our team, as a third party, independently evaluated these algorithms on three test sets. In addition, we also evaluated pre-existing AI frameworks--which, differing from algorithms, are more flexible and can support different algorithms--including MONAI from NVIDIA, nnU-Net from DKFZ, and numerous other open-source frameworks. We are committed to expanding this benchmark to encourage more innovation of AI algorithms for the medical domain.

Zhisong Wang, Yiwen Ye, Ziyang Chen et al.

Scribble-based weakly supervised segmentation methods have shown promising results in medical image segmentation, significantly reducing annotation costs. However, existing approaches often rely on auxiliary tasks to enforce semantic consistency and use hard pseudo labels for supervision, overlooking the unique challenges faced by models trained with sparse annotations. These models must predict pixel-wise segmentation maps from limited data, making it crucial to handle varying levels of annotation richness effectively. In this paper, we propose MaCo, a weakly supervised model designed for medical image segmentation, based on the principle of "from few to more." MaCo leverages Masked Context Modeling (MCM) and Continuous Pseudo Labels (CPL). MCM employs an attention-based masking strategy to perturb the input image, ensuring that the model's predictions align with those of the original image. CPL converts scribble annotations into continuous pixel-wise labels by applying an exponential decay function to distance maps, producing confidence maps that represent the likelihood of each pixel belonging to a specific category, rather than relying on hard pseudo labels. We evaluate MaCo on three public datasets, comparing it with other weakly supervised methods. Our results show that MaCo outperforms competing methods across all datasets, establishing a new record in weakly supervised medical image segmentation.

Ziyang Chen, Yansong Qu, You Shen et al.

Driven by the advancement of 3D devices, stereo vision tasks including stereo matching and stereo conversion have emerged as a critical research frontier. Contemporary stereo vision backbones typically rely on either monocular depth estimation (MDE) models or visual foundation models (VFMs). Crucially, these models are predominantly pretrained without explicit supervision of camera poses. Given that such geometric knowledge is indispensable for stereo vision, the absence of explicit spatial constraints constitutes a significant performance bottleneck for existing architectures. Recognizing that the Visual Geometry Grounded Transformer (VGGT) operates as a foundation model pretrained on extensive 3D priors, including camera poses, we investigate its potential as a robust backbone for stereo vision tasks. Nevertheless, empirical results indicate that its direct application to stereo vision yields suboptimal performance. We observe that VGGT suffers from a more significant degradation of geometric details during feature extraction. Such characteristics conflict with the requirements of binocular stereo vision, thereby constraining its efficacy for relative tasks. To bridge this gap, we propose StereoVGGT, a feature backbone specifically tailored for stereo vision. By leveraging the frozen VGGT and introducing a training-free feature adjustment pipeline, we mitigate geometric degradation and harness the latent camera calibration knowledge embedded within the model. StereoVGGT-based stereo matching network achieved the $1^{st}$ rank among all published methods on the KITTI benchmark, validating that StereoVGGT serves as a highly effective backbone for stereo vision.

Xingliang Lei, Yiwen Ye, Zhisong Wang et al.

Parameter-efficient fine-tuning (PEFT) techniques have emerged to address overfitting and high computational costs associated with fully fine-tuning in self-supervised learning. Mainstream PEFT methods add a few trainable parameters while keeping the pre-trained backbone parameters fixed. These methods achieve comparative, and often superior, performance to fully fine-tuning, demonstrating the powerful representation ability of the pre-trained backbone. Despite this success, these methods typically ignore the initialization of the new parameters, often relying solely on random initialization. We argue that if pre-training is significantly beneficial, it should be applied to all parameters requiring representational capacity. Motivated by this, we propose Target Parameter Pre-training (TPP), a simple yet effective fine-tuning framework. TPP pre-trains target parameters, i.e., the new parameters introduced during fine-tuning, in an additional stage before PEFT. During this stage, the pre-trained backbone parameters are frozen, and only the new parameters are trainable. A defined pretext task encourages the new parameters to learn specific representations of downstream data. Subsequently, when PEFT is employed, the pre-trained new parameters are loaded to enhance fine-tuning efficiency. The proposed TPP framework is versatile, allowing integration with various pre-trained backbones, pretext tasks, and PEFT methods. We evaluated the fine-tuning performance of our method on seven public datasets, covering four modalities and two task types. The results demonstrate that TPP can be easily integrated into existing PEFT methods, significantly improving performance.

Shaoguang Wang, Weiyu Guo, Ziyang Chen et al.

Long video understanding remains a formidable challenge for Multimodal Large Language Models (MLLMs) due to the prohibitive computational cost of processing dense frame sequences. Prevailing solutions, which select a keyframe subset, typically rely on either a single visual-centric metric (e.g., CLIP similarity) or a static fusion of heuristic scores. This ``one-size-fits-all'' paradigm frequently fails: visual-only metrics are ineffective for plot-driven narrative queries, while indiscriminately incorporating textual scores introduces severe ``modal noise'' for purely visual tasks. To break this bottleneck, we propose Q-Gate, a plug-and-play and training-free framework that treats keyframe selection as a dynamic modality routing problem. We decouple the retrieval process into three lightweight expert streams: Visual Grounding for local details, Global Matching for scene semantics, and Contextual Alignment for subtitle-driven narratives. Crucially, Q-Gate introduces a Query-Modulated Gating Mechanism that leverages the in-context reasoning of an LLM to assess the query's intent and dynamically allocate attention weights across the experts. This mechanism intelligently activates necessary modalities while ``muting'' irrelevant ones, thereby maximizing the signal-to-noise ratio. Extensive experiments on LongVideoBench and Video-MME across multiple MLLM backbones demonstrate that Q-Gate substantially outperforms state-of-the-art baselines. By effectively suppressing modality-specific noise, it provides a robust, highly interpretable solution for scalable video reasoning.

Ziyang Chen, Yongjun Zhang, Wenting Li et al.

Real-world applications of stereo matching, such as autonomous driving, place stringent demands on both safety and accuracy. However, learning-based stereo matching methods inherently suffer from the loss of geometric structures in certain feature channels, creating a bottleneck in achieving precise detail matching. Additionally, these methods lack interpretability due to the black-box nature of deep learning. In this paper, we propose MoCha-V2, a novel learning-based paradigm for stereo matching. MoCha-V2 introduces the Motif Correlation Graph (MCG) to capture recurring textures, which are referred to as ``motifs" within feature channels. These motifs reconstruct geometric structures and are learned in a more interpretable way. Subsequently, we integrate features from multiple frequency domains through wavelet inverse transformation. The resulting motif features are utilized to restore geometric structures in the stereo matching process. Experimental results demonstrate the effectiveness of MoCha-V2. MoCha-V2 achieved 1st place on the Middlebury benchmark at the time of its release. Code is available at https://github.com/ZYangChen/MoCha-Stereo.

Ziyang Chen, Renbing Chen, Daowei Li et al.

Simulating group-level user behavior enables scalable counterfactual evaluation of merchant strategies without costly online experiments. However, building a trustworthy simulator faces two structural challenges. First, information incompleteness causes reasoning-based simulators to over-rationalize when unobserved factors such as offline context and implicit habits are missing. Second, mechanism duality requires capturing both interpretable preferences and implicit statistical regularities, which no single paradigm achieves alone. We propose Policy-Guided Hybrid Simulation (PGHS), a dual-process framework that mines transferable decision policies from behavioral trajectories and uses them as a shared alignment layer. This layer anchors an LLM-based reasoning branch that prevents over-rationalization and an ML-based fitting branch that absorbs implicit regularities. Group-level predictions from both branches are fused for complementary correction. We deploy PGHS on Meituan with 101 merchants and over 26,000 trajectories. PGHS achieves a group simulation error of 8.80%, improving over the best reasoning-based and fitting-based baselines by 45.8% and 40.9% respectively.

Zhanpeng Chen, Mingxiao Li, Ziyang Chen et al.

Vision-language Models (VLMs) have shown remarkable capabilities in advancing general artificial intelligence, yet the irrational encoding of visual positions persists in inhibiting the models' comprehensive perception performance across different levels of granularity. In this work, we propose Pyramid-descent Visual Position Encoding (PyPE), a novel approach designed to enhance the perception of visual tokens within VLMs. By assigning visual position indexes from the periphery to the center and expanding the central receptive field incrementally, PyPE addresses the limitations of traditional raster-scan methods and mitigates the long-term decay effects induced by Rotary Position Embedding (RoPE). Our method reduces the relative distance between interrelated visual elements and instruction tokens, promoting a more rational allocation of attention weights and allowing for a multi-granularity perception of visual elements and countering the over-reliance on anchor tokens. Extensive experimental evaluations demonstrate that PyPE consistently improves the general capabilities of VLMs across various sizes. Code is available at https://github.com/SakuraTroyChen/PyPE.

Ziyang Chen, Xiaobin Wang, Yong Jiang et al.

Question Answering (QA) systems face challenges in handling complex questions that require multi-domain knowledge synthesis. The naive RAG models, although effective in information retrieval, struggle with complex questions that require comprehensive and in-depth answers. The pioneering task is defined as explanatory answer generation, which entails handling identified challenges such as the requirement for comprehensive information and logical coherence within the generated context. To address these issues, we refer to systematic thinking theory and propose SynthRAG, an innovative framework designed to enhance QA performance. SynthRAG improves on conventional models by employing adaptive outlines for dynamic content structuring, generating systematic information to ensure detailed coverage, and producing customized answers tailored to specific user inquiries. This structured approach guarantees logical coherence and thorough integration of information, yielding responses that are both insightful and methodically organized. Empirical evaluations underscore SynthRAG's effectiveness, demonstrating its superiority in handling complex questions, overcoming the limitations of naive RAG models, and significantly improving answer quality and depth. Furthermore, an online deployment on the Zhihu platform revealed that SynthRAG's answers achieved notable user engagement, with each response averaging 5.73 upvotes and surpassing the performance of 79.8% of human contributors, highlighting the practical relevance and impact of the proposed framework. Our code is available at https://github.com/czy1999/SynthRAG .

Zhisong Wang, Ziyang Chen, Zanting Ye et al.

Medical visual question answering (Med-VQA) aims to answer clinically relevant questions grounded in medical images. However, existing multimodal large language models (MLLMs) often exhibit shortcut answering, producing plausible responses by exploiting language priors or dataset biases while insufficiently attending to visual evidence. This behavior undermines clinical reliability, especially when subtle imaging findings are decisive. We propose a lightweight plug-in framework, termed Intent-aware Visual Cues (InViC), to explicitly enhance image-based answer generation in medical VQA. InViC introduces a Cue Tokens Extraction (CTE) module that distills dense visual tokens into a compact set of K question-conditioned cue tokens, which serve as structured visual intermediaries injected into the LLM decoder to promote intent-aligned visual evidence. To discourage bypassing of visual information, we further design a two-stage fine-tuning strategy with a cue-bottleneck attention mask. In Stage I, we employ an attention mask to block the LLM's direct view of raw visual features, thereby funneling all visual evidence through the cue pathway. In Stage II, standard causal attention is restored to train the LLM to jointly exploit the visual and cue tokens. We evaluate InViC on three public Med-VQA benchmarks (VQA-RAD, SLAKE, and ImageCLEF VQA-Med 2019) across multiple representative MLLMs. InViC consistently improves over zero-shot inference and standard LoRA fine-tuning, demonstrating that intent-aware visual cues with bottlenecked training is a practical and effective strategy for improving trustworthy Med-VQA.

Weiyu Guo, He Zhang, Pengteng Li et al.

Recent advances in embodied intelligence have leveraged massive scaling of data and model parameters to master natural-language command following and multi-task control. In contrast, biological systems demonstrate an innate ability to acquire skills rapidly from sparse experience. Crucially, current robotic policies struggle to replicate the dynamic stability, reflexive responsiveness, and temporal memory inherent in biological motion. Here we present Neuromorphic Vision-Language-Action (NeuroVLA), a framework that mimics the structural organization of the bio-nervous system between the cortex, cerebellum, and spinal cord. We adopt a system-level bio-inspired design: a high-level model plans goals, an adaptive cerebellum module stabilizes motion using high-frequency sensors feedback, and a bio-inspired spinal layer executes lightning-fast actions generation. NeuroVLA represents the first deployment of a neuromorphic VLA on physical robotics, achieving state-of-the-art performance. We observe the emergence of biological motor characteristics without additional data or special guidance: it stops the shaking in robotic arms, saves significant energy(only 0.4w on Neuromorphic Processor), shows temporal memory ability and triggers safety reflexes in less than 20 milliseconds.

Siqi Miao, Ziyang Chen, Yuhong Luo et al.

While Large Language Model (LLM) multi-agent systems (MAS) offer a transformative approach to simulating human behavior in complex systems, it remains largely unexplored whether these simulations can replicate realistic structural and temporal dynamics from a dynamic network perspective. Our evaluation indicates that existing frameworks excel at generating plausible micro-level interactions but fail to capture the emergent, macroscopic topologies necessary for domains that rely on realistic network dynamics, such as modeling information propagation and cybersecurity threats. To bridge this gap, we introduce two easily integrable extensions to simulation frameworks to ensure they preserve macroscopic network fidelity: 1) augmenting LLM agents with data-driven event triggers to organically sustain long-horizon interactions, and 2) integrating Hawkes processes to accurately model temporal activation dynamics. Our approach allows LLM MAS to capture both plausible micro-level patterns and macroscopic topologies. We further demonstrate the utility of this framework in synthesizing realistic phishing campaigns within evolving communication networks. The study reveals how threats exploit structural vulnerabilities, highlighting the potential of our framework for developing next-generation defenses. Our code is available at https://github.com/Graph-COM/NSL.

Ziyang Chen, Wenting Li, Yongjun Zhang et al.

Constrained by the low-rank bottleneck inherent in attention mechanisms, current stereo matching transformers suffer from limited nonlinear expressivity, which renders their feature representations sensitive to challenging conditions such as reflections. To overcome this difficulty, we present the Hadamard Attention Recurrent Stereo Transformer (HART). HART includes a novel attention mechanism that incorporates the following components: 1) The Dense Attention Kernel (DAK) maps the attention weight distribution into a high-dimensional space over (0, +$\infty$). By removing the upper bound constraint on attention weights, DAK enables more flexible modeling of complex feature interactions. This reduces feature collinearity. 2) The Multi Kernel & Order Interaction (MKOI) module extends the attention mechanism by unifying semantic and spatial knowledge learning. This integration improves the ability of HART to learn features in binocular images. Experimental results demonstrate the effectiveness of our HART. In reflective area, HART ranked 1st on the KITTI 2012 benchmark among all published methods at the time of submission. Code is available at https://github.com/ZYangChen/HART.

Ziyang Chen, Yongsheng Pan, Yiwen Ye et al.

Although recent years have witnessed the great success of convolutional neural networks (CNNs) in medical image segmentation, the domain shift issue caused by the highly variable image quality of medical images hinders the deployment of CNNs in real-world clinical applications. Domain generalization (DG) methods aim to address this issue by training a robust model on the source domain, which has a strong generalization ability. Previously, many DG methods based on feature-space domain randomization have been proposed, which, however, suffer from the limited and unordered search space of feature styles. In this paper, we propose a multi-source DG method called Treasure in Distribution (TriD), which constructs an unprecedented search space to obtain the model with strong robustness by randomly sampling from a uniform distribution. To learn the domain-invariant representations explicitly, we further devise a style-mixing strategy in our TriD, which mixes the feature styles by randomly mixing the augmented and original statistics along the channel wise and can be extended to other DG methods. Extensive experiments on two medical segmentation tasks with different modalities demonstrate that our TriD achieves superior generalization performance on unseen target-domain data. Code is available at https://github.com/Chen-Ziyang/TriD.

Fengyu Yang, Chao Feng, Ziyang Chen et al.

The ability to associate touch with other modalities has huge implications for humans and computational systems. However, multimodal learning with touch remains challenging due to the expensive data collection process and non-standardized sensor outputs. We introduce UniTouch, a unified tactile model for vision-based touch sensors connected to multiple modalities, including vision, language, and sound. We achieve this by aligning our UniTouch embeddings to pretrained image embeddings already associated with a variety of other modalities. We further propose learnable sensor-specific tokens, allowing the model to learn from a set of heterogeneous tactile sensors, all at the same time. UniTouch is capable of conducting various touch sensing tasks in the zero-shot setting, from robot grasping prediction to touch image question answering. To the best of our knowledge, UniTouch is the first to demonstrate such capabilities. Project page: https://cfeng16.github.io/UniTouch/

Wei Luo, Yiting Lu, Xin Li et al.

This paper reports on the LoViF 2026 PhyScore challenge, a competition on holistic quality assessment of world-model-generated videos across both 2D and 4D generation settings. The challenge is motivated by a central gap in current evaluation practice: perceptual quality alone is insufficient to judge whether generated dynamics are physically plausible, temporally coherent, and consistent with input conditions. Participants are required to build a metric that jointly predicts four dimensions, i.e., Video Quality, Physical Realism, Condition-Video Alignment, and Temporal Consistency. Depart from that, participants also need to localize physical anomaly timestamps for fine-grained diagnosis. The benchmark dataset contains 1,554 videos generated by seven representative world generative models, organized into three tracks (text-2D, image-to-4D, and video-to-4D) and spanning 26 categories. These categories explicitly cover physics-relevant scenarios, including dynamics, optics, and thermodynamics, together with diverse real-world and creative content. To ensure label reliability, scores and anomaly timestamps are produced through trained human annotation with an additional automated quality-control pass. Evaluation is based on both score prediction and anomaly localization, with a composite protocol that combines TimeStamp_IOU and SRCC/PLCC. This report summarizes the challenge design and provides method-level insights from submitted solutions.

Xiangde Luo, Jia Fu, Yunxin Zhong et al.

Radiation therapy is a primary and effective NasoPharyngeal Carcinoma (NPC) treatment strategy. The precise delineation of Gross Tumor Volumes (GTVs) and Organs-At-Risk (OARs) is crucial in radiation treatment, directly impacting patient prognosis. Previously, the delineation of GTVs and OARs was performed by experienced radiation oncologists. Recently, deep learning has achieved promising results in many medical image segmentation tasks. However, for NPC OARs and GTVs segmentation, few public datasets are available for model development and evaluation. To alleviate this problem, the SegRap2023 challenge was organized in conjunction with MICCAI2023 and presented a large-scale benchmark for OAR and GTV segmentation with 400 Computed Tomography (CT) scans from 200 NPC patients, each with a pair of pre-aligned non-contrast and contrast-enhanced CT scans. The challenge's goal was to segment 45 OARs and 2 GTVs from the paired CT scans. In this paper, we detail the challenge and analyze the solutions of all participants. The average Dice similarity coefficient scores for all submissions ranged from 76.68\% to 86.70\%, and 70.42\% to 73.44\% for OARs and GTVs, respectively. We conclude that the segmentation of large-size OARs is well-addressed, and more efforts are needed for GTVs and small-size or thin-structure OARs. The benchmark will remain publicly available here: https://segrap2023.grand-challenge.org

Ziyang Chen, Israel D. Gebru, Christian Richardt et al.

We present a new dataset called Real Acoustic Fields (RAF) that captures real acoustic room data from multiple modalities. The dataset includes high-quality and densely captured room impulse response data paired with multi-view images, and precise 6DoF pose tracking data for sound emitters and listeners in the rooms. We used this dataset to evaluate existing methods for novel-view acoustic synthesis and impulse response generation which previously relied on synthetic data. In our evaluation, we thoroughly assessed existing audio and audio-visual models against multiple criteria and proposed settings to enhance their performance on real-world data. We also conducted experiments to investigate the impact of incorporating visual data (i.e., images and depth) into neural acoustic field models. Additionally, we demonstrated the effectiveness of a simple sim2real approach, where a model is pre-trained with simulated data and fine-tuned with sparse real-world data, resulting in significant improvements in the few-shot learning approach. RAF is the first dataset to provide densely captured room acoustic data, making it an ideal resource for researchers working on audio and audio-visual neural acoustic field modeling techniques. Demos and datasets are available on our project page: https://facebookresearch.github.io/real-acoustic-fields/

Ziyang Chen, Prem Seetharaman, Bryan Russell et al.

Generating sound effects for videos often requires creating artistic sound effects that diverge significantly from real-life sources and flexible control in the sound design. To address this problem, we introduce MultiFoley, a model designed for video-guided sound generation that supports multimodal conditioning through text, audio, and video. Given a silent video and a text prompt, MultiFoley allows users to create clean sounds (e.g., skateboard wheels spinning without wind noise) or more whimsical sounds (e.g., making a lion's roar sound like a cat's meow). MultiFoley also allows users to choose reference audio from sound effects (SFX) libraries or partial videos for conditioning. A key novelty of our model lies in its joint training on both internet video datasets with low-quality audio and professional SFX recordings, enabling high-quality, full-bandwidth (48kHz) audio generation. Through automated evaluations and human studies, we demonstrate that MultiFoley successfully generates synchronized high-quality sounds across varied conditional inputs and outperforms existing methods. Please see our project page for video results: https://ificl.github.io/MultiFoley/

Ziyang Chen, Daniel Geng, Andrew Owens

Spectrograms are 2D representations of sound that look very different from the images found in our visual world. And natural images, when played as spectrograms, make unnatural sounds. In this paper, we show that it is possible to synthesize spectrograms that simultaneously look like natural images and sound like natural audio. We call these visual spectrograms images that sound. Our approach is simple and zero-shot, and it leverages pre-trained text-to-image and text-to-spectrogram diffusion models that operate in a shared latent space. During the reverse process, we denoise noisy latents with both the audio and image diffusion models in parallel, resulting in a sample that is likely under both models. Through quantitative evaluations and perceptual studies, we find that our method successfully generates spectrograms that align with a desired audio prompt while also taking the visual appearance of a desired image prompt. Please see our project page for video results: https://ificl.github.io/images-that-sound/

Weiyu Guo, Ziyang Chen, Shaoguang Wang et al.

Understanding long video content is a complex endeavor that often relies on densely sampled frame captions or end-to-end feature selectors, yet these techniques commonly overlook the logical relationships between textual queries and visual elements. In practice, computational constraints necessitate coarse frame subsampling, a challenge analogous to "finding a needle in a haystack." To address this issue, we introduce a semantics-driven search framework that reformulates keyframe selection under the paradigm of Visual Semantic-Logical Search. Specifically, we systematically define four fundamental logical dependencies: 1) spatial co-occurrence, 2) temporal proximity, 3) attribute dependency, and 4) causal order. These relations dynamically update frame sampling distributions through an iterative refinement process, enabling context-aware identification of semantically critical frames tailored to specific query requirements. Our method establishes new SOTA performance on the manually annotated benchmark in key-frame selection metrics. Furthermore, when applied to downstream video question-answering tasks, the proposed approach demonstrates the best performance gains over existing methods on LongVideoBench and Video-MME, validating its effectiveness in bridging the logical gap between textual queries and visual-temporal reasoning. The code will be publicly available.

Zijing Zhang, Ziyang Chen, Mingxiao Li et al.

The development of autonomous agents for complex, long-horizon tasks is a central goal in AI. However, dominant training paradigms face a critical limitation: reinforcement learning (RL) methods that optimize solely for final task success often reinforce flawed or inefficient reasoning paths, a problem we term inefficient exploration. This leads to agents that are brittle and fail to generalize, as they learn to find solutions without learning how to reason coherently. To address this, we introduce RLVMR, a novel framework that integrates dense, process-level supervision into end-to-end RL by rewarding verifiable, meta-reasoning behaviors. RLVMR equips an agent to explicitly tag its cognitive steps, such as planning, exploration, and reflection, and provides programmatic, rule-based rewards for actions that contribute to effective problem-solving. These process-centric rewards are combined with the final outcome signal and optimized using a critic-free policy gradient method. On the challenging ALFWorld and ScienceWorld benchmarks, RLVMR achieves new state-of-the-art results, with our 7B model reaching an 83.6% success rate on the most difficult unseen task split. Our analysis confirms these gains stem from improved reasoning quality, including significant reductions in redundant actions and enhanced error recovery, leading to more robust, efficient, and interpretable agents.

Ming Hu, Chenglong Ma, Wei Li et al. · pku

Scientific Large Language Models (Sci-LLMs) are transforming how knowledge is represented, integrated, and applied in scientific research, yet their progress is shaped by the complex nature of scientific data. This survey presents a comprehensive, data-centric synthesis that reframes the development of Sci-LLMs as a co-evolution between models and their underlying data substrate. We formulate a unified taxonomy of scientific data and a hierarchical model of scientific knowledge, emphasizing the multimodal, cross-scale, and domain-specific challenges that differentiate scientific corpora from general natural language processing datasets. We systematically review recent Sci-LLMs, from general-purpose foundations to specialized models across diverse scientific disciplines, alongside an extensive analysis of over 270 pre-/post-training datasets, showing why Sci-LLMs pose distinct demands -- heterogeneous, multi-scale, uncertainty-laden corpora that require representations preserving domain invariance and enabling cross-modal reasoning. On evaluation, we examine over 190 benchmark datasets and trace a shift from static exams toward process- and discovery-oriented assessments with advanced evaluation protocols. These data-centric analyses highlight persistent issues in scientific data development and discuss emerging solutions involving semi-automated annotation pipelines and expert validation. Finally, we outline a paradigm shift toward closed-loop systems where autonomous agents based on Sci-LLMs actively experiment, validate, and contribute to a living, evolving knowledge base. Collectively, this work provides a roadmap for building trustworthy, continually evolving artificial intelligence (AI) systems that function as a true partner in accelerating scientific discovery.

Ziyang Chen, Zhenxuan Huang, Yile Wang et al.

Traditional sentence embedding methods employ token-level contrastive learning on non-generative pre-trained models. Recently, there have emerged embedding methods based on generative large language models (LLMs). These methods either rely on fixed prompt templates or involve modifications to the model architecture. The former lacks further optimization of the model and results in limited performance, while the latter alters the internal computational mechanisms of the model, thereby compromising its generative capabilities. We propose SemPA, a novel approach that boosts the sentence representations while preserving the generative ability of LLMs via semantic preference alignment. We leverage sentence-level Direct Preference Optimization (DPO) to efficiently optimize LLMs on a paraphrase generation task, where the model learns to discriminate semantically equivalent sentences while preserving inherent generative capacity. Theoretically, we establish a formal connection between DPO and contrastive learning under the Plackett-Luce model framework. Empirically, experimental results on both semantic textual similarity tasks and various benchmarks for LLMs show that SemPA achieves better semantic representations without sacrificing the inherent generation capability of LLMs.

Shaoguang Wang, Weiyu Guo, Ziyang Chen et al.

The practical application of Multimodal Large Language Models (MLLMs) to Video Question Answering (Video-QA) is severely hindered by the high token cost of processing numerous video frames. While keyframe selection is the dominant strategy for mitigating this, we identify a critical flaw: even state-of-the-art selectors produce prompts suffering from significant temporal redundancy, a challenge unique to video that we term 'visual echoes'. This issue leads to context dilution and can paradoxically degrade performance. To address this dual challenge, we propose a novel refinement framework that synergistically combines Adaptive Frame-Pruning(AFP) with a lightweight text-based semantic graph. AFP intelligently prunes 'visual echoes' by adaptively clustering frames, while the semantic graph provides crucial, low-cost semantic compensation. Conducting extensive experiments on the LongVideoBench and Video-MME benchmarks against multiple state-of-the-art selectors, our approach demonstrates a drastic reduction in total input tokens by up to 82.2%. Crucially, by creating a concise, high-quality prompt, our framework not only enhances efficiency but also demonstrates a remarkable ability to robustify and improve the accuracy of upstream selectors, achieving results that are highly competitive with, and often superior to, baselines that use vastly more frames.