Chenglin Yang

Chenglin Yang, Siyuan Qiao, Qihang Yu et al. · deepmind

This paper presents MOAT, a family of neural networks that build on top of MObile convolution (i.e., inverted residual blocks) and ATtention. Unlike the current works that stack separate mobile convolution and transformer blocks, we effectively merge them into a MOAT block. Starting with a standard Transformer block, we replace its multi-layer perceptron with a mobile convolution block, and further reorder it before the self-attention operation. The mobile convolution block not only enhances the network representation capacity, but also produces better downsampled features. Our conceptually simple MOAT networks are surprisingly effective, achieving 89.1% / 81.5% top-1 accuracy on ImageNet-1K / ImageNet-1K-V2 with ImageNet22K pretraining. Additionally, MOAT can be seamlessly applied to downstream tasks that require large resolution inputs by simply converting the global attention to window attention. Thanks to the mobile convolution that effectively exchanges local information between pixels (and thus cross-windows), MOAT does not need the extra window-shifting mechanism. As a result, on COCO object detection, MOAT achieves 59.2% box AP with 227M model parameters (single-scale inference, and hard NMS), and on ADE20K semantic segmentation, MOAT attains 57.6% mIoU with 496M model parameters (single-scale inference). Finally, the tiny-MOAT family, obtained by simply reducing the channel sizes, also surprisingly outperforms several mobile-specific transformer-based models on ImageNet. The tiny-MOAT family is also benchmarked on downstream tasks, serving as a baseline for the community. We hope our simple yet effective MOAT will inspire more seamless integration of convolution and self-attention. Code is publicly available.

Chenglin Yang, Siyuan Qiao, Yuan Cao et al. · deepmind

Generative training has been demonstrated to be powerful for building visual-language models. However, on zero-shot discriminative benchmarks, there is still a performance gap between models trained with generative and discriminative objectives. In this paper, we aim to narrow this gap by improving the efficacy of generative training on classification tasks, without any finetuning processes or additional modules. Specifically, we focus on narrowing the gap between the generative captioner and the CLIP classifier. We begin by analysing the predictions made by the captioner and classifier and observe that the caption generation inherits the distribution bias from the language model trained with pure text modality, making it less grounded on the visual signal. To tackle this problem, we redesign the scoring objective for the captioner to alleviate the distributional bias and focus on measuring the gain of information brought by the visual inputs. We further design a generative training objective to match the evaluation objective. We name our model trained and evaluated from the novel procedures as Information Gain (IG) captioner. We pretrain the models on the public Laion-5B dataset and perform a series of discriminative evaluations. For the zero-shot classification on ImageNet, IG captioner achieves $> 18\%$ improvements over the standard captioner, achieving comparable performances with the CLIP classifier. IG captioner also demonstrated strong performance on zero-shot image-text retrieval tasks on MSCOCO and Flickr30K. We hope this paper inspires further research towards unifying generative and discriminative training procedures for visual-language models.

Yihan Xie, Sijing Li, Tianwei Lin et al.

Although electrocardiograms (ECG) play a dominant role in cardiovascular diagnosis and treatment, their intrinsic data forms and representational patterns pose significant challenges for medical multimodal large language models (Med-MLLMs) in achieving cross-modal semantic alignment. To address this gap, we propose Heartcare Suite, a unified ECG suite designed for dual signal-image modeling and understanding: (i) Heartcare-400K. A fine-grained ECG instruction dataset on top of our data pipeline engine--HeartAgent--by integrating high quality clinical ECG reports from top hospitals with open-source data. (ii) Heartcare-Bench. A systematic benchmark assessing performance of models in multi-perspective ECG understanding and cross-modal generalization, providing guidance for optimizing ECG comprehension models. (iii) HeartcareGPT. Built upon a structure-aware discrete tokenizer Beat, we propose Dual Stream Projection Alignment (DSPA) paradigm--a dual encoder projection alignment mechanism enabling joint optimizing and modeling native ECG signal-image within a shared feature space. HeartcareGPT achieves consistent improvements across diverse ECG understanding tasks, validating both the effectiveness of the unified modeling paradigm and the necessity of a high-quality data pipeline, and establishing a methodological foundation for extending Med-MLLMs towards physiological signal domains. Our project is available at https://github.com/ZJU4HealthCare/HeartcareGPT .

Sijing Li, Zhongwei Qiu, Jiang Liu et al.

Accurate tumor analysis is central to clinical radiology and precision oncology, where early detection, reliable lesion characterization, and pathology-level risk assessment guide diagnosis and treatment planning. Chain-of-Thought (CoT) reasoning is particularly important in this setting because it enables step-by-step interpretation from imaging findings to clinical impressions and pathology conclusions, improving traceability and reducing diagnostic errors. Here, we target the clinical tumor analysis task and build a large-scale benchmark that operationalizes a multimodal reasoning pipeline, spanning findings, impressions, and pathology predictions. We curate TumorCoT, a large-scale dataset of 1.5M CoT-labeled VQA instructions paired with 3D CT scans, with step-aligned rationales and cross-modal alignments along the trajectory from findings to impression to pathology, enabling evaluation of both answer accuracy and reasoning consistency. We further propose TumorChain, a multimodal interleaved reasoning framework that tightly couples 3D imaging encoders, clinical text understanding, and organ-level vision-language alignment. Through cross-modal alignment and iterative interleaved causal reasoning, TumorChain grounds visual evidence, aggregates conclusions, and issues pathology predictions after multiple rounds of self-refinement, improving traceability and reducing hallucination risk. Experiments show consistent improvements over strong baselines in lesion detection, impression generation, and pathology classification, and demonstrate strong generalization on the DeepTumorVQA benchmark. These results highlight the potential of multimodal reasoning for reliable and interpretable tumor analysis in clinical practice. Detailed information about our project can be found on our project homepage at https://github.com/ZJU4HealthCare/TumorChain.

Chenglin Yang

Modern AI agents execute real-world side effects through tool calls such as file operations, shell commands, HTTP requests, and database queries. A single unsafe action, including accidental deletion, credential exposure, or data exfiltration, can cause irreversible harm. Existing defenses are incomplete: post-hoc benchmarks measure behavior after execution, static guardrails miss obfuscation and multi-step context, and infrastructure sandboxes constrain where code runs without understanding what an action means. We present AgentTrust, a runtime safety layer that intercepts agent tool calls before execution and returns a structured verdict: allow, warn, block, or review. AgentTrust combines a shell deobfuscation normalizer, SafeFix suggestions for safer alternatives, RiskChain detection for multi-step attack chains, and a cache-aware LLM-as-Judge for ambiguous inputs. We release a 300-scenario benchmark across six risk categories and an additional 630 independently constructed real-world adversarial scenarios. On the internal benchmark, the production-only ruleset achieves 95.0% verdict accuracy and 73.7% risk-level accuracy at low-millisecond end-to-end latency. On the 630-scenario benchmark, evaluated under a patched ruleset and not claimed as zero-shot, AgentTrust achieves 96.7% verdict accuracy, including about 93% on shell-obfuscated payloads. AgentTrust is released under the AGPL-3.0 license and provides a Model Context Protocol server for MCP-compatible agents.

Dongwon Kim, Ju He, Qihang Yu et al.

Image tokenizers form the foundation of modern text-to-image generative models but are notoriously difficult to train. Furthermore, most existing text-to-image models rely on large-scale, high-quality private datasets, making them challenging to replicate. In this work, we introduce Text-Aware Transformer-based 1-Dimensional Tokenizer (TA-TiTok), an efficient and powerful image tokenizer that can utilize either discrete or continuous 1-dimensional tokens. TA-TiTok uniquely integrates textual information during the tokenizer decoding stage (i.e., de-tokenization), accelerating convergence and enhancing performance. TA-TiTok also benefits from a simplified, yet effective, one-stage training process, eliminating the need for the complex two-stage distillation used in previous 1-dimensional tokenizers. This design allows for seamless scalability to large datasets. Building on this, we introduce a family of text-to-image Masked Generative Models (MaskGen), trained exclusively on open data while achieving comparable performance to models trained on private data. We aim to release both the efficient, strong TA-TiTok tokenizers and the open-data, open-weight MaskGen models to promote broader access and democratize the field of text-to-image masked generative models.

Chenglin Yang, Celong Liu, Xueqing Deng et al.

We present 1.58-bit FLUX, the first successful approach to quantizing the state-of-the-art text-to-image generation model, FLUX.1-dev, using 1.58-bit weights (i.e., values in {-1, 0, +1}) while maintaining comparable performance for generating 1024 x 1024 images. Notably, our quantization method operates without access to image data, relying solely on self-supervision from the FLUX.1-dev model. Additionally, we develop a custom kernel optimized for 1.58-bit operations, achieving a 7.7x reduction in model storage, a 5.1x reduction in inference memory, and improved inference latency. Extensive evaluations on the GenEval and T2I Compbench benchmarks demonstrate the effectiveness of 1.58-bit FLUX in maintaining generation quality while significantly enhancing computational efficiency.

Xueqing Deng, Qihang Yu, Ali Athar et al.

This paper introduces the COCONut-PanCap dataset, created to enhance panoptic segmentation and grounded image captioning. Building upon the COCO dataset with advanced COCONut panoptic masks, this dataset aims to overcome limitations in existing image-text datasets that often lack detailed, scene-comprehensive descriptions. The COCONut-PanCap dataset incorporates fine-grained, region-level captions grounded in panoptic segmentation masks, ensuring consistency and improving the detail of generated captions. Through human-edited, densely annotated descriptions, COCONut-PanCap supports improved training of vision-language models (VLMs) for image understanding and generative models for text-to-image tasks. Experimental results demonstrate that COCONut-PanCap significantly boosts performance across understanding and generation tasks, offering complementary benefits to large-scale datasets. This dataset sets a new benchmark for evaluating models on joint panoptic segmentation and grounded captioning tasks, addressing the need for high-quality, detailed image-text annotations in multi-modal learning.

Inkyu Shin, Chenglin Yang, Liang-Chieh Chen

Flow based generative models have charted an impressive path across multiple visual generation tasks by adhering to a simple principle: learning velocity representations of a linear interpolant. However, we observe that training velocity solely from the final layer output underutilizes the rich inter layer representations, potentially impeding model convergence. To address this limitation, we introduce DeepFlow, a novel framework that enhances velocity representation through inter layer communication. DeepFlow partitions transformer layers into balanced branches with deep supervision and inserts a lightweight Velocity Refiner with Acceleration (VeRA) block between adjacent branches, which aligns the intermediate velocity features within transformer blocks. Powered by the improved deep supervision via the internal velocity alignment, DeepFlow converges 8 times faster on ImageNet with equivalent performance and further reduces FID by 2.6 while halving training time compared to previous flow based models without a classifier free guidance. DeepFlow also outperforms baselines in text to image generation tasks, as evidenced by evaluations on MSCOCO and zero shot GenEval.

Chenglin Yang, Yilin Wang, Jianming Zhang et al.

Despite the impressive representation capacity of vision transformer models, current light-weight vision transformer models still suffer from inconsistent and incorrect dense predictions at local regions. We suspect that the power of their self-attention mechanism is limited in shallower and thinner networks. We propose Lite Vision Transformer (LVT), a novel light-weight transformer network with two enhanced self-attention mechanisms to improve the model performances for mobile deployment. For the low-level features, we introduce Convolutional Self-Attention (CSA). Unlike previous approaches of merging convolution and self-attention, CSA introduces local self-attention into the convolution within a kernel of size 3x3 to enrich low-level features in the first stage of LVT. For the high-level features, we propose Recursive Atrous Self-Attention (RASA), which utilizes the multi-scale context when calculating the similarity map and a recursive mechanism to increase the representation capability with marginal extra parameter cost. The superiority of LVT is demonstrated on ImageNet recognition, ADE20K semantic segmentation, and COCO panoptic segmentation. The code is made publicly available.

Chenglin Yang, Siyuan Qiao, Adam Kortylewski et al.

Self-Attention has become prevalent in computer vision models. Inspired by fully connected Conditional Random Fields (CRFs), we decompose self-attention into local and context terms. They correspond to the unary and binary terms in CRF and are implemented by attention mechanisms with projection matrices. We observe that the unary terms only make small contributions to the outputs, and meanwhile standard CNNs that rely solely on the unary terms achieve great performances on a variety of tasks. Therefore, we propose Locally Enhanced Self-Attention (LESA), which enhances the unary term by incorporating it with convolutions, and utilizes a fusion module to dynamically couple the unary and binary operations. In our experiments, we replace the self-attention modules with LESA. The results on ImageNet and COCO show the superiority of LESA over convolution and self-attention baselines for the tasks of image recognition, object detection, and instance segmentation. The code is made publicly available.

Chenglin Yang, Yilin Wang, Jianming Zhang et al.

Compared with common image segmentation tasks targeted at low-resolution images, higher resolution detailed image segmentation receives much less attention. In this paper, we propose and study a task named Meticulous Object Segmentation (MOS), which is focused on segmenting well-defined foreground objects with elaborate shapes in high resolution images (e.g. 2k - 4k). To this end, we propose the MeticulousNet which leverages a dedicated decoder to capture the object boundary details. Specifically, we design a Hierarchical Point-wise Refining (HierPR) block to better delineate object boundaries, and reformulate the decoding process as a recursive coarse to fine refinement of the object mask. To evaluate segmentation quality near object boundaries, we propose the Meticulosity Quality (MQ) score considering both the mask coverage and boundary precision. In addition, we collect a MOS benchmark dataset including 600 high quality images with complex objects. We provide comprehensive empirical evidence showing that MeticulousNet can reveal pixel-accurate segmentation boundaries and is superior to state-of-the-art methods for high resolution object segmentation tasks.

Christian Cosgrove, Adam Kortylewski, Chenglin Yang et al.

Patch-based adversarial attacks introduce a perceptible but localized change to the input that induces misclassification. While progress has been made in defending against imperceptible attacks, it remains unclear how patch-based attacks can be resisted. In this work, we study two different approaches for defending against black-box patch attacks. First, we show that adversarial training, which is successful against imperceptible attacks, has limited effectiveness against state-of-the-art location-optimized patch attacks. Second, we find that compositional deep networks, which have part-based representations that lead to innate robustness to natural occlusion, are robust to patch attacks on PASCAL3D+ and the German Traffic Sign Recognition Benchmark, without adversarial training. Moreover, the robustness of compositional models outperforms that of adversarially trained standard models by a large margin. However, on GTSRB, we observe that they have problems discriminating between similar traffic signs with fine-grained differences. We overcome this limitation by introducing part-based finetuning, which improves fine-grained recognition. By leveraging compositional representations, this is the first work that defends against black-box patch attacks without expensive adversarial training. This defense is more robust than adversarial training and more interpretable because it can locate and ignore adversarial patches.

Chenglin Yang, Adam Kortylewski, Cihang Xie et al.

Patch-based attacks introduce a perceptible but localized change to the input that induces misclassification. A limitation of current patch-based black-box attacks is that they perform poorly for targeted attacks, and even for the less challenging non-targeted scenarios, they require a large number of queries. Our proposed PatchAttack is query efficient and can break models for both targeted and non-targeted attacks. PatchAttack induces misclassifications by superimposing small textured patches on the input image. We parametrize the appearance of these patches by a dictionary of class-specific textures. This texture dictionary is learned by clustering Gram matrices of feature activations from a VGG backbone. PatchAttack optimizes the position and texture parameters of each patch using reinforcement learning. Our experiments show that PatchAttack achieves > 99% success rate on ImageNet for a wide range of architectures, while only manipulating 3% of the image for non-targeted attacks and 10% on average for targeted attacks. Furthermore, we show that PatchAttack circumvents state-of-the-art adversarial defense methods successfully.

Chenglin Yang, Lingxi Xie, Chi Su et al.

Optimizing a deep neural network is a fundamental task in computer vision, yet direct training methods often suffer from over-fitting. Teacher-student optimization aims at providing complementary cues from a model trained previously, but these approaches are often considerably slow due to the pipeline of training a few generations in sequence, i.e., time complexity is increased by several times. This paper presents snapshot distillation (SD), the first framework which enables teacher-student optimization in one generation. The idea of SD is very simple: instead of borrowing supervision signals from previous generations, we extract such information from earlier epochs in the same generation, meanwhile make sure that the difference between teacher and student is sufficiently large so as to prevent under-fitting. To achieve this goal, we implement SD in a cyclic learning rate policy, in which the last snapshot of each cycle is used as the teacher for all iterations in the next cycle, and the teacher signal is smoothed to provide richer information. In standard image classification benchmarks such as CIFAR100 and ILSVRC2012, SD achieves consistent accuracy gain without heavy computational overheads. We also verify that models pre-trained with SD transfers well to object detection and semantic segmentation in the PascalVOC dataset.

Chenglin Yang, Lingxi Xie, Siyuan Qiao et al.

We focus on the problem of training a deep neural network in generations. The flowchart is that, in order to optimize the target network (student), another network (teacher) with the same architecture is first trained, and used to provide part of supervision signals in the next stage. While this strategy leads to a higher accuracy, many aspects (e.g., why teacher-student optimization helps) still need further explorations. This paper studies this problem from a perspective of controlling the strictness in training the teacher network. Existing approaches mostly used a hard distribution (e.g., one-hot vectors) in training, leading to a strict teacher which itself has a high accuracy, but we argue that the teacher needs to be more tolerant, although this often implies a lower accuracy. The implementation is very easy, with merely an extra loss term added to the teacher network, facilitating a few secondary classes to emerge and complement to the primary class. Consequently, the teacher provides a milder supervision signal (a less peaked distribution), and makes it possible for the student to learn from inter-class similarity and potentially lower the risk of over-fitting. Experiments are performed on standard image classification tasks (CIFAR100 and ILSVRC2012). Although the teacher network behaves less powerful, the students show a persistent ability growth and eventually achieve higher classification accuracies than other competitors. Model ensemble and transfer feature extraction also verify the effectiveness of our approach.