Zhuowei Li

Zhuowei Li, Long Zhao, Zizhao Zhang et al. · deepmind

In the context of continual learning, prototypes-as representative class embeddings-offer advantages in memory conservation and the mitigation of catastrophic forgetting. However, challenges related to semantic drift and prototype interference persist. In this study, we introduce the Contrastive Prototypical Prompt (CPP) approach. Through task-specific prompt-tuning, underpinned by a contrastive learning objective, we effectively address both aforementioned challenges. Our evaluations on four challenging class-incremental benchmarks reveal that CPP achieves a significant 4% to 6% improvement over state-of-the-art methods. Importantly, CPP operates without a rehearsal buffer and narrows the performance divergence between continual and offline joint-learning, suggesting an innovative scheme for Transformer-based continual learning systems.

Zhuowei Li, Yibo Gao, Zhenzhou Zha et al.

Neural architecture search (NAS) algorithms save tremendous labor from human experts. Recent advancements further reduce the computational overhead to an affordable level. However, it is still cumbersome to deploy the NAS techniques in real-world applications due to the fussy procedures and the supervised learning paradigm. In this work, we propose the self-supervised and weight-preserving neural architecture search (SSWP-NAS) as an extension of the current NAS framework by allowing the self-supervision and retaining the concomitant weights discovered during the search stage. As such, we simplify the workflow of NAS to a one-stage and proxy-free procedure. Experiments show that the architectures searched by the proposed framework achieve state-of-the-art accuracy on CIFAR-10, CIFAR-100, and ImageNet datasets without using manual labels. Moreover, we show that employing the concomitant weights as initialization consistently outperforms the random initialization and the two-stage weight pre-training method by a clear margin under semi-supervised learning scenarios. Codes are publicly available at https://github.com/LzVv123456/SSWP-NAS.

Yunhe Gao, Zhuowei Li, Di Liu et al.

A major focus of clinical imaging workflow is disease diagnosis and management, leading to medical imaging datasets strongly tied to specific clinical objectives. This scenario has led to the prevailing practice of developing task-specific segmentation models, without gaining insights from widespread imaging cohorts. Inspired by the training program of medical radiology residents, we propose a shift towards universal medical image segmentation, a paradigm aiming to build medical image understanding foundation models by leveraging the diversity and commonality across clinical targets, body regions, and imaging modalities. Towards this goal, we develop Hermes, a novel context-prior learning approach to address the challenges of data heterogeneity and annotation differences in medical image segmentation. In a large collection of eleven diverse datasets (2,438 3D images) across five modalities (CT, PET, T1, T2 and cine MRI) and multiple body regions, we demonstrate the merit of the universal paradigm over the traditional paradigm on addressing multiple tasks within a single model. By exploiting the synergy across tasks, Hermes achieves state-of-the-art performance on all testing datasets and shows superior model scalability. Results on two additional datasets reveals Hermes' strong performance for transfer learning, incremental learning, and generalization to downstream tasks. Hermes's learned priors demonstrate an appealing trait to reflect the intricate relations among tasks and modalities, which aligns with the established anatomical and imaging principles in radiology. The code is available: https://github.com/yhygao/universal-medical-image-segmentation.

Di Liu, Xiang Yu, Meng Ye et al.

Accurate 3D shape abstraction from a single 2D image is a long-standing problem in computer vision and graphics. By leveraging a set of primitives to represent the target shape, recent methods have achieved promising results. However, these methods either use a relatively large number of primitives or lack geometric flexibility due to the limited expressibility of the primitives. In this paper, we propose a novel bi-channel Transformer architecture, integrated with parameterized deformable models, termed DeFormer, to simultaneously estimate the global and local deformations of primitives. In this way, DeFormer can abstract complex object shapes while using a small number of primitives which offer a broader geometry coverage and finer details. Then, we introduce a force-driven dynamic fitting and a cycle-consistent re-projection loss to optimize the primitive parameters. Extensive experiments on ShapeNet across various settings show that DeFormer achieves better reconstruction accuracy over the state-of-the-art, and visualizes with consistent semantic correspondences for improved interpretability.

Tunyu Zhang, Xinxi Zhang, Ligong Han et al.

Diffusion large language models (DLLMs) have the potential to enable fast text generation by decoding multiple tokens in parallel. However, in practice, their inference efficiency is constrained by the need for many refinement steps, while aggressively reducing the number of steps leads to a substantial degradation in generation quality. To alleviate this, we propose a trajectory self-distillation framework that improves few-step decoding by distilling the model's own generative trajectories. We incorporate Direct Discriminative Optimization (DDO), a reverse-KL objective that promotes mode-seeking distillation and encourages the student to concentrate on high-probability teacher modes. Across benchmarks, our approach consistently outperforms strong few-step baselines and standard training under tight step budgets. Although full-step decoding remains superior, we substantially narrow the gap, establishing a strong foundation towards practical few-step DLLMs. The source code is available at https://github.com/Tyrion58/T3D.

Tunyu Zhang, Haizhou Shi, Yibin Wang et al.

While Large Language Models (LLMs) have demonstrated impressive capabilities, their output quality remains inconsistent across various application scenarios, making it difficult to identify trustworthy responses, especially in complex tasks requiring multi-step reasoning. In this paper, we propose a Token-level Uncertainty estimation framework for Reasoning (TokUR) that enables LLMs to self-assess and self-improve their responses in mathematical reasoning. Specifically, we introduce low-rank random weight perturbation during LLM decoding to generate predictive distributions for token-level uncertainty estimation, and we aggregate these uncertainty quantities to capture the semantic uncertainty of generated responses. Experiments on mathematical reasoning datasets of varying difficulty demonstrate that TokUR exhibits a strong correlation with answer correctness and model robustness, and the uncertainty signals produced by TokUR can be leveraged to enhance the model's reasoning performance at test time. These results highlight the effectiveness of TokUR as a principled and scalable approach for improving the reliability and interpretability of LLMs in challenging reasoning tasks.

Zhuowei Li, Haizhou Shi, Yunhe Gao et al.

Large Vision-Language Models (LVLMs) can reason effectively over both textual and visual inputs, but they tend to hallucinate syntactically coherent yet visually ungrounded contents. In this paper, we investigate the internal dynamics of hallucination by examining the tokens logits ranking throughout the generation process, revealing three key patterns in how LVLMs process information: (1) gradual visual information loss - visually grounded tokens gradually become less favored throughout generation, and (2) early excitation - semantically meaningful tokens achieve peak activation in the layers earlier than the final layer. (3) hidden genuine information - visually grounded tokens though not being eventually decoded still retain relatively high rankings at inference. Based on these insights, we propose VISTA (Visual Information Steering with Token-logit Augmentation), a training-free inference-time intervention framework that reduces hallucination while promoting genuine information. VISTA works by combining two complementary approaches: reinforcing visual information in activation space and leveraging early layer activations to promote semantically meaningful decoding. Compared to existing methods, VISTA requires no external supervision and is applicable to various decoding strategies. Extensive experiments show that VISTA on average reduces hallucination by about 40% on evaluated open-ended generation task, and it consistently outperforms existing methods on four benchmarks across four architectures under three decoding strategies. Code is available at https://github.com/LzVv123456/VISTA.

Gerasimos Chatzoudis, Konstantinos D. Polyzos, Zhuowei Li et al.

Understanding the internal activations of Vision Transformers (ViTs) is critical for building interpretable and trustworthy models. While Sparse Autoencoders (SAEs) have been used to extract human-interpretable features, they operate on individual layers and fail to capture the cross-layer computational structure of Transformers, as well as the relative significance of each layer in forming the last-layer representation. Alternatively, we introduce the adoption of Cross-Layer Transcoders (CLTs) as reliable, sparse, and depth-aware proxy models for MLP blocks in ViTs. CLTs use an encoder-decoder scheme to reconstruct each post-MLP activation from learned sparse embeddings of preceding layers, yielding a linear decomposition that transforms the final representation of ViTs from an opaque embedding into an additive, layer-resolved construction that enables faithful attribution and process-level interpretability. We train CLTs on CLIP ViT-B/32 and ViT-B/16 across CIFAR-100, COCO, and ImageNet-100. We show that CLTs achieve high reconstruction fidelity with post-MLP activations while preserving and even improving, in some cases, CLIP zero-shot classification accuracy. In terms of interpretability, we show that the cross-layer contribution scores provide faithful attribution, revealing that the final representation is concentrated in a smaller set of dominant layer-wise terms whose removal degrades performance and whose retention largely preserves it. These results showcase the significance of adopting CLTs as an alternative interpretable proxy of ViTs in the vision domain.

Zhuowei Li, Zihao Xu, Ligong Han et al.

In-context Learning (ICL) empowers large language models (LLMs) to swiftly adapt to unseen tasks at inference-time by prefixing a few demonstration examples before queries. Despite its versatility, ICL incurs substantial computational and memory overheads compared to zero-shot learning and is sensitive to the selection and order of demonstration examples. In this work, we introduce Implicit In-context Learning (I2CL), an innovative paradigm that reduces the inference cost of ICL to that of zero-shot learning with minimal information loss. I2CL operates by first generating a condensed vector representation, namely a context vector, extracted from the demonstration examples. It then conducts an inference-time intervention through injecting a linear combination of the context vector and query activations back into the model's residual streams. Empirical evaluation on nine real-world tasks across three model architectures demonstrates that I2CL achieves few-shot level performance at zero-shot inference cost, and it exhibits robustness against variations in demonstration examples. Furthermore, I2CL facilitates a novel representation of task-ids, enhancing task similarity detection and fostering effective transfer learning. We also perform a comprehensive analysis and ablation study on I2CL, offering deeper insights into its internal mechanisms. Code is available at https://github.com/LzVv123456/I2CL.

Jason Wu, Tianchen Zhao, Chang Liu et al.

Large Vision-Language Models (LVLMs) use their vision encoders to translate images into representations for downstream reasoning, but the encoders often underperform in domain-specific visual tasks such as medical image diagnosis or fine-grained classification, where representation errors can cascade through the language model, leading to incorrect responses. Existing adaptation methods modify the continuous feature interface between encoder and language model through projector tuning or other parameter-efficient updates, which still couples the two components and requires re-alignment whenever the encoder changes. We introduce CRAFT (Codebook RegulAted Fine-Tuning), a lightweight method that fine-tunes the encoder using a discrete codebook that anchors visual representations to a stable token space, achieving domain adaptation without modifying other parts of the model. This decoupled design allows the adapted encoder to seamlessly boost the performance of LVLMs with different language architectures, as long as they share the same codebook. Empirically, CRAFT achieves an average gain of 13.51% across 10 domain-specific benchmarks such as VQARAD and PlantVillage, while preserving the LLM's linguistic capabilities and outperforming peer methods that operate on continuous tokens.

Zhenting Wang, Shuming Hu, Shiyu Zhao et al.

Image content safety has become a significant challenge with the rise of visual media on online platforms. Meanwhile, in the age of AI-generated content (AIGC), many image generation models are capable of producing harmful content, such as images containing sexual or violent material. Thus, it becomes crucial to identify such unsafe images based on established safety rules. Pre-trained Multimodal Large Language Models (MLLMs) offer potential in this regard, given their strong pattern recognition abilities. Existing approaches typically fine-tune MLLMs with human-labeled datasets, which however brings a series of drawbacks. First, relying on human annotators to label data following intricate and detailed guidelines is both expensive and labor-intensive. Furthermore, users of safety judgment systems may need to frequently update safety rules, making fine-tuning on human-based annotation more challenging. This raises the research question: Can we detect unsafe images by querying MLLMs in a zero-shot setting using a predefined safety constitution (a set of safety rules)? Our research showed that simply querying pre-trained MLLMs does not yield satisfactory results. This lack of effectiveness stems from factors such as the subjectivity of safety rules, the complexity of lengthy constitutions, and the inherent biases in the models. To address these challenges, we propose a MLLM-based method includes objectifying safety rules, assessing the relevance between rules and images, making quick judgments based on debiased token probabilities with logically complete yet simplified precondition chains for safety rules, and conducting more in-depth reasoning with cascaded chain-of-thought processes if necessary. Experiment results demonstrate that our method is highly effective for zero-shot image safety judgment tasks.

Zhuowei Li, Miao Zhang, Xiaotian Lin et al.

This paper introduces GAgent: an Gripping Agent designed for open-world environments that provides advanced cognitive abilities via VLM agents and flexible grasping abilities with variable stiffness soft grippers. GAgent comprises three primary components - Prompt Engineer module, Visual-Language Model (VLM) core and Workflow module. These three modules enhance gripper success rates by recognizing objects and materials and accurately estimating grasp area even under challenging lighting conditions. As part of creativity, researchers also created a bionic hybrid soft gripper with variable stiffness capable of gripping heavy loads while still gently engaging objects. This intelligent agent, featuring VLM-based cognitive processing with bionic design, shows promise as it could potentially benefit UAVs in various scenarios.

Yunhe Gao, Di Liu, Zhuowei Li et al.

Medical image segmentation remains challenging due to the vast diversity of anatomical structures, imaging modalities, and segmentation tasks. While deep learning has made significant advances, current approaches struggle to generalize as they require task-specific training or fine-tuning on unseen classes. We present Iris, a novel In-context Reference Image guided Segmentation framework that enables flexible adaptation to novel tasks through the use of reference examples without fine-tuning. At its core, Iris features a lightweight context task encoding module that distills task-specific information from reference context image-label pairs. This rich context embedding information is used to guide the segmentation of target objects. By decoupling task encoding from inference, Iris supports diverse strategies from one-shot inference and context example ensemble to object-level context example retrieval and in-context tuning. Through comprehensive evaluation across twelve datasets, we demonstrate that Iris performs strongly compared to task-specific models on in-distribution tasks. On seven held-out datasets, Iris shows superior generalization to out-of-distribution data and unseen classes. Further, Iris's task encoding module can automatically discover anatomical relationships across datasets and modalities, offering insights into medical objects without explicit anatomical supervision.

Zhuowei Li, Tianchen Zhao, Xiang Xu et al.

Developing a face anti-spoofing model that meets the security requirements of clients worldwide is challenging due to the domain gap between training datasets and diverse end-user test data. Moreover, for security and privacy reasons, it is undesirable for clients to share a large amount of their face data with service providers. In this work, we introduce a novel method in which the face anti-spoofing model can be adapted by the client itself to a target domain at test time using only a small sample of data while keeping model parameters and training data inaccessible to the client. Specifically, we develop a prototype-based base model and an optimal transport-guided adaptor that enables adaptation in either a lightweight training or training-free fashion, without updating base model's parameters. Furthermore, we propose geodesic mixup, an optimal transport-based synthesis method that generates augmented training data along the geodesic path between source prototypes and target data distribution. This allows training a lightweight classifier to effectively adapt to target-specific characteristics while retaining essential knowledge learned from the source domain. In cross-domain and cross-attack settings, compared with recent methods, our method achieves average relative improvements of 19.17% in HTER and 8.58% in AUC, respectively.

Gerasimos Chatzoudis, Zhuowei Li, Gemma E. Moran et al.

Steering vision foundation models at inference time without retraining or access to large labeled datasets is a desirable yet challenging objective, particularly in dynamic or resource-constrained settings. In this paper, we introduce Visual Sparse Steering (VS2), a lightweight, test-time method that guides vision models using steering vectors derived from sparse features learned by top-$k$ Sparse Autoencoders without requiring contrastive data. Specifically, VS2 surpasses zero-shot CLIP by 4.12% on CIFAR-100, 1.08% on CUB-200, and 1.84% on Tiny-ImageNet. We further propose VS2++, a retrieval-augmented variant that selectively amplifies relevant sparse features using pseudo-labeled neighbors at inference time. With oracle positive/negative sets, VS2++ achieves absolute top-1 gains over CLIP zero-shot of up to 21.44% on CIFAR-100, 7.08% on CUB-200, and 20.47% on Tiny-ImageNet. Interestingly, VS2 and VS2++ raise per-class accuracy by up to 25% and 38%, respectively, showing that sparse steering benefits specific classes by disambiguating visually or taxonomically proximate categories rather than providing a uniform boost. Finally, to better align the sparse features learned through the SAE reconstruction task with those relevant for downstream performance, we propose Prototype-Aligned Sparse Steering (PASS). By incorporating a prototype-alignment loss during SAE training, using labels only during training while remaining fully test-time unsupervised, PASS consistently, though modestly, outperforms VS2, achieving a 6.12% gain over VS2 only on CIFAR-100 with ViT-B/32.

Zhuowei Li, Zihao Liu, Zhiqiang Hu et al.

Medical image segmentation has been widely recognized as a pivot procedure for clinical diagnosis, analysis, and treatment planning. However, the laborious and expensive annotation process lags down the speed of further advances. Contrastive learning-based weight pre-training provides an alternative by leveraging unlabeled data to learn a good representation. In this paper, we investigate how contrastive learning benefits the general supervised medical segmentation tasks. To this end, patch-dragsaw contrastive regularization (PDCR) is proposed to perform patch-level tugging and repulsing with the extent controlled by a continuous affinity score. And a new structure dubbed uncertainty-aware feature selection block (UAFS) is designed to perform the feature selection process, which can handle the learning target shift caused by minority features with high uncertainty. By plugging the proposed 2 modules into the existing segmentation architecture, we achieve state-of-the-art results across 8 public datasets from 6 domains. Newly designed modules further decrease the amount of training data to a quarter while achieving comparable, if not better, performances. From this perspective, we take the opposite direction of the original self/un-supervised contrastive learning by further excavating information contained within the label.

Jinxi Xiang, Zhuowei Li, Wenji Wang et al.

Deep learning has demonstrated significant improvements in medical image segmentation using a sufficiently large amount of training data with manual labels. Acquiring well-representative labels requires expert knowledge and exhaustive labors. In this paper, we aim to boost the performance of semi-supervised learning for medical image segmentation with limited labels using a self-ensembling contrastive learning technique. To this end, we propose to train an encoder-decoder network at image-level with small amounts of labeled images, and more importantly, we learn latent representations directly at feature-level by imposing contrastive loss on unlabeled images. This method strengthens intra-class compactness and inter-class separability, so as to get a better pixel classifier. Moreover, we devise a student encoder for online learning and an exponential moving average version of it, called teacher encoder, to improve the performance iteratively in a self-ensembling manner. To construct contrastive samples with unlabeled images, two sampling strategies that exploit structure similarity across medical images and utilize pseudo-labels for construction, termed region-aware and anatomical-aware contrastive sampling, are investigated. We conduct extensive experiments on an MRI and a CT segmentation dataset and demonstrate that in a limited label setting, the proposed method achieves state-of-the-art performance. Moreover, the anatomical-aware strategy that prepares contrastive samples on-the-fly using pseudo-labels realizes better contrastive regularization on feature representations.