Jiawei Mao

Junqi Liu, Salena Song, Yuhan Wang et al.

Autonomous agents are increasingly expected to support end-to-end medical-AI research workflows, moving beyond isolated prediction tasks or short-form clinical question answering. However, existing medical agent benchmarks primarily evaluate final outputs, providing limited visibility into agent behavior within the research process. To address this gap, we present AutoMedBench, a workflow-aware benchmark for autonomous medical-AI research across diverse medical imaging and multimodal inference tasks, organizing agent execution into a unified five-stage workflow (S1-S5): Plan, Setup, Validate, Inference, and Submit. It comprises long-horizon tasks with each run averaging 33 agent turns, spanning five research tracks: segmentation, image enhancement, visual question answering (VQA), report generation, and lesion detection. Each task is evaluated under two difficulty tiers, Lite and Standard, which use the same data and metrics but differ in the amount of task-brief scaffolding, and each run is scored using both final task performance and S1-S5 stage scores, enabling stage-level analysis from the initial task brief to the final submitted artifact. Across thousands of recorded runs, stage-level scoring reveals that Validate is the weakest workflow stage on average, whereas Setup is the strongest, suggesting that current agents are better at making pipelines executable than at verifying their reliability. Post-run error analysis further shows that verification and submission failures dominate tagged errors, accounting for 37.7% and 38.1% of fired codes respectively, whereas task-understanding errors are rare at 0.9%, and runs with one fired error code have a 48% lower overall score than runs with no error code on average.

Jiawei Mao, Honggu Zhou, Xuesong Yin et al.

Vision Transformers (ViTs) outperforms convolutional neural networks (CNNs) in several vision tasks with its global modeling capabilities. However, ViT lacks the inductive bias inherent to convolution making it require a large amount of data for training. This results in ViT not performing as well as CNNs on small datasets like medicine and science. We experimentally found that masked autoencoders (MAE) can make the transformer focus more on the image itself, thus alleviating the data-hungry issue of ViT to some extent. Yet the current MAE model is too complex resulting in over-fitting problems on small datasets. This leads to a gap between MAEs trained on small datasets and advanced CNNs models still. Therefore, we investigated how to reduce the decoder complexity in MAE and found a more suitable architectural configuration for it with small datasets. Besides, we additionally designed a location prediction task and a contrastive learning task to introduce localization and invariance characteristics for MAE. Our contrastive learning task not only enables the model to learn high-level visual information but also allows the training of MAE's class token. This is something that most MAE improvement efforts do not consider. Extensive experiments have shown that our method shows state-of-the-art performance on standard small datasets as well as medical datasets with few samples compared to the current popular masked image modeling (MIM) and vision transformers for small datasets.The code and models are available at https://github.com/Talented-Q/SDMAE.

Jiawei Mao, Guangyi Zhao, Yuanqi Chang et al.

Facial expression recognition (FER) plays a significant role in the ubiquitous application of computer vision. We revisit this problem with a new perspective on whether it can acquire useful representations that improve FER performance in the image generation process, and propose a novel generative method based on the image inversion mechanism for the FER task, termed Inversion FER (IFER). Particularly, we devise a novel Adversarial Style Inversion Transformer (ASIT) towards IFER to comprehensively extract features of generated facial images. In addition, ASIT is equipped with an image inversion discriminator that measures the cosine similarity of semantic features between source and generated images, constrained by a distribution alignment loss. Finally, we introduce a feature modulation module to fuse the structural code and latent codes from ASIT for the subsequent FER work. We extensively evaluate ASIT on facial datasets such as FFHQ and CelebA-HQ, showing that our approach achieves state-of-the-art facial inversion performance. IFER also achieves competitive results in facial expression recognition datasets such as RAF-DB, SFEW and AffectNet. The code and models are available at https://github.com/Talented-Q/IFER-master.

Jiawei Mao, Rui Xu, Xuesong Yin et al.

Facial expression recognition (FER) plays an important role in a variety of real-world applications such as human-computer interaction. POSTER achieves the state-of-the-art (SOTA) performance in FER by effectively combining facial landmark and image features through two-stream pyramid cross-fusion design. However, the architecture of POSTER is undoubtedly complex. It causes expensive computational costs. In order to relieve the computational pressure of POSTER, in this paper, we propose POSTER++. It improves POSTER in three directions: cross-fusion, two-stream, and multi-scale feature extraction. In cross-fusion, we use window-based cross-attention mechanism replacing vanilla cross-attention mechanism. We remove the image-to-landmark branch in the two-stream design. For multi-scale feature extraction, POSTER++ combines images with landmark's multi-scale features to replace POSTER's pyramid design. Extensive experiments on several standard datasets show that our POSTER++ achieves the SOTA FER performance with the minimum computational cost. For example, POSTER++ reached 92.21% on RAF-DB, 67.49% on AffectNet (7 cls) and 63.77% on AffectNet (8 cls), respectively, using only 8.4G floating point operations (FLOPs) and 43.7M parameters (Param). This demonstrates the effectiveness of our improvements.

Jiawei Mao, Hardy Chen, Haoqin Tu et al. · princeton

Large vision-language models (LVLMs) have become increasingly strong but remain prone to hallucinations in multimodal tasks, which significantly narrows their deployment. As training these LVLMs to avoid hallucinations becomes prohibitively expensive for larger models, training-free methods offer a cheap and flexible solution to this problem, yet existing approaches based on decoding or tool use often bring limited gains and/or weak interpretability. We propose Kestrel, a training-free framework for LVLM hallucination mitigation that combines an explicit visual-grounding agent with evidence-verified self-refinement mechanism. In detail, Kestrel first collects explicit visual evidence and converts tool outputs into reusable and structured textual evidence. Second, to take full advantage of these evidence, Kestrel verifies them via an LVLM judge for evidence checking, then iteratively self-refine answers based on verified evidence to reduce the risk of over-correction. Extensive experiments show that Kestrel improves performance over strong baselines across hallucination benchmarks (e.g., average +3.31% on POPE and +28.34 on MME-Hallucination with Qwen3-VL), while providing transparent verification traces for hallucination diagnosis and analysis -- e.g., both the integrated self-refinement module and grounding agent contributing an average +2.0% gain on POPE.

Honggu Zhou, Xiaogang Peng, Jiawei Mao et al.

Some self-supervised cross-modal learning approaches have recently demonstrated the potential of image signals for enhancing point cloud representation. However, it remains a question on how to directly model cross-modal local and global correspondences in a self-supervised fashion. To solve it, we proposed PointCMC, a novel cross-modal method to model multi-scale correspondences across modalities for self-supervised point cloud representation learning. In particular, PointCMC is composed of: (1) a local-to-local (L2L) module that learns local correspondences through optimized cross-modal local geometric features, (2) a local-to-global (L2G) module that aims to learn the correspondences between local and global features across modalities via local-global discrimination, and (3) a global-to-global (G2G) module, which leverages auxiliary global contrastive loss between the point cloud and image to learn high-level semantic correspondences. Extensive experiment results show that our approach outperforms existing state-of-the-art methods in various downstream tasks such as 3D object classification and segmentation. Code will be made publicly available upon acceptance.

Jiawei Mao, Yuanqi Chang, Xuesong Yin et al.

Compared to other severe weather image restoration tasks, single image desnowing is a more challenging task. This is mainly due to the diversity and irregularity of snow shape, which makes it extremely difficult to restore images in snowy scenes. Moreover, snow particles also have a veiling effect similar to haze or mist. Although current works can effectively remove snow particles with various shapes, they also bring distortion to the restored image. To address these issues, we propose a novel single image desnowing network called Star-Net. First, we design a Star type Skip Connection (SSC) to establish information channels for all different scale features, which can deal with the complex shape of snow particles.Second, we present a Multi-Stage Interactive Transformer (MIT) as the base module of Star-Net, which is designed to better understand snow particle shapes and to address image distortion by explicitly modeling a variety of important image recovery features. Finally, we propose a Degenerate Filter Module (DFM) to filter the snow particle and snow fog residual in the SSC on the spatial and channel domains. Extensive experiments show that our Star-Net achieves state-of-the-art snow removal performances on three standard snow removal datasets and retains the original sharpness of the images.

Jiawei Mao, Xuesong Yin, Yuanqi Chang et al.

This paper explores improvements to the masked image modeling (MIM) paradigm. The MIM paradigm enables the model to learn the main object features of the image by masking the input image and predicting the masked part by the unmasked part. We found the following three main directions for MIM to be improved. First, since both encoders and decoders contribute to representation learning, MIM uses only encoders for downstream tasks, which ignores the impact of decoders on representation learning. Although the MIM paradigm already employs small decoders with asymmetric structures, we believe that continued reduction of decoder parameters is beneficial to improve the representational learning capability of the encoder . Second, MIM solves the image prediction task by training the encoder and decoder together , and does not design a separate task for the encoder . To further enhance the performance of the encoder when performing downstream tasks, we designed the encoder for the tasks of comparative learning and token position prediction. Third, since the input image may contain background and other objects, and the proportion of each object in the image varies, reconstructing the tokens related to the background or to other objects is not meaningful for MIM to understand the main object representations. Therefore we use ContrastiveCrop to crop the input image so that the input image contains as much as possible only the main objects. Based on the above three improvements to MIM, we propose a new model, Contrastive Masked AutoEncoders (CMAE). We achieved a Top-1 accuracy of 65.84% on tinyimagenet using the ViT-B backbone, which is +2.89 outperforming the MAE of competing methods when all conditions are equal. Code will be made available.

Jiawei Mao, Yuanqi Chang, Xuesong Yin

Compared with the vanilla transformer, the window-based transformer offers a better trade-off between accuracy and efficiency. Although the window-based transformer has made great progress, its long-range modeling capabilities are limited due to the size of the local window and the window connection scheme. To address this problem, we propose a novel Token Transformer (TT). The core mechanism of TT is the addition of a Class (CLS) token for summarizing window information in each local window. We refer to this type of token interaction as CLS Attention. These CLS tokens will interact spatially with the tokens in each window to enable long-range modeling. In order to preserve the hierarchical design of the window-based transformer, we designed Feature Inheritance Module (FIM) in each phase of TT to deliver the local window information from the previous phase to the CLS token in the next phase. In addition, we have designed a Spatial-Channel Feedforward Network (SCFFN) in TT, which can mix CLS tokens and embedded tokens on the spatial domain and channel domain without additional parameters. Extensive experiments have shown that our TT achieves competitive results with low parameters in image classification and downstream tasks.

Jiawei Mao, Yuhan Wang, Yucheng Tang et al.

This paper presents MedSegFactory, a versatile medical synthesis framework that generates high-quality paired medical images and segmentation masks across modalities and tasks. It aims to serve as an unlimited data repository, supplying image-mask pairs to enhance existing segmentation tools. The core of MedSegFactory is a dual-stream diffusion model, where one stream synthesizes medical images and the other generates corresponding segmentation masks. To ensure precise alignment between image-mask pairs, we introduce Joint Cross-Attention (JCA), enabling a collaborative denoising paradigm by dynamic cross-conditioning between streams. This bidirectional interaction allows both representations to guide each other's generation, enhancing consistency between generated pairs. MedSegFactory unlocks on-demand generation of paired medical images and segmentation masks through user-defined prompts that specify the target labels, imaging modalities, anatomical regions, and pathological conditions, facilitating scalable and high-quality data generation. This new paradigm of medical image synthesis enables seamless integration into diverse medical imaging workflows, enhancing both efficiency and accuracy. Extensive experiments show that MedSegFactory generates data of superior quality and usability, achieving competitive or state-of-the-art performance in 2D and 3D segmentation tasks while addressing data scarcity and regulatory constraints.

Jiawei Mao, Yuhan Wang, Lifeng Chen et al.

Recent advances in generative medical models are constrained by modality-specific scenarios that hinder the integration of complementary evidence from imaging, pathology, and clinical notes. This fragmentation limits their evolution into foundation models that can learn and reason across the full spectrum of biomedical data. We propose MeDiM, the first medical discrete diffusion model that learns shared distributions across modalities without modality-specific components. MeDiM unifies multiple generative tasks: translating between images and text, and jointly producing image-report pairs across domains in response to prompts. Built on a discrete diffusion framework, MeDiM bridges vision and language representations through a shared probabilistic space. To enable unified and flexible medical generation, we employ a multimodal large language model (MLLM) as the diffusion backbone, leveraging its prior knowledge and cross-modal reasoning. Two key designs are introduced: (1) removing the causal attention mask for bidirectional context, and (2) injecting continuous timestep embeddings for diffusion awareness. Experiments demonstrate high-fidelity medical generation (FID 16.60 on MIMIC-CXR and FID 24.19 on PathGen) and accurate report generation (METEOR 0.2650 and 0.2580). Jointly generated image-report pairs further enhance downstream performance (plus6.43 percent BLEU-1, plus18.57 percent BLEU-2, plus31.58 percent BLEU-3, plus4.80 percent METEOR), showing that MeDiM supports coherent and clinically grounded multimodal outputs.

Jiawei Mao, Yu Yang, Xuesong Yin et al.

Image restoration models often face the simultaneous interaction of multiple degradations in real-world scenarios. Existing approaches typically handle single or composite degradations based on scene descriptors derived from text or image embeddings. However, due to the varying proportions of different degradations within an image, these scene descriptors may not accurately differentiate between degradations, leading to suboptimal restoration in practical applications. To address this issue, we propose a novel Transformer-based restoration framework, AllRestorer. In AllRestorer, we enable the model to adaptively consider all image impairments, thereby avoiding errors from scene descriptor misdirection. Specifically, we introduce an All-in-One Transformer Block (AiOTB), which adaptively removes all degradations present in a given image by modeling the relationships between all degradations and the image embedding in latent space. To accurately address different variations potentially present within the same type of degradation and minimize ambiguity, AiOTB utilizes a composite scene descriptor consisting of both image and text embeddings to define the degradation. Furthermore, AiOTB includes an adaptive weight for each degradation, allowing for precise control of the restoration intensity. By leveraging AiOTB, AllRestorer avoids misdirection caused by inaccurate scene descriptors, achieving a 5.00 dB increase in PSNR compared to the baseline on the CDD-11 dataset.

Jiawei Mao, Guangyi Zhao, Xuesong Yin et al.

This paper proposes the first pure Transformer structure inversion network called SwinStyleformer, which can compensate for the shortcomings of the CNNs inversion framework by handling long-range dependencies and learning the global structure of objects. Experiments found that the inversion network with the Transformer backbone could not successfully invert the image. The above phenomena arise from the differences between CNNs and Transformers, such as the self-attention weights favoring image structure ignoring image details compared to convolution, the lack of multi-scale properties of Transformer, and the distribution differences between the latent code extracted by the Transformer and the StyleGAN style vector. To address these differences, we employ the Swin Transformer with a smaller window size as the backbone of the SwinStyleformer to enhance the local detail of the inversion image. Meanwhile, we design a Transformer block based on learnable queries. Compared to the self-attention transformer block, the Transformer block based on learnable queries provides greater adaptability and flexibility, enabling the model to update the attention weights according to specific tasks. Thus, the inversion focus is not limited to the image structure. To further introduce multi-scale properties, we design multi-scale connections in the extraction of feature maps. Multi-scale connections allow the model to gain a comprehensive understanding of the image to avoid loss of detail due to global modeling. Moreover, we propose an inversion discriminator and distribution alignment loss to minimize the distribution differences. Based on the above designs, our SwinStyleformer successfully solves the Transformer's inversion failure issue and demonstrates SOTA performance in image inversion and several related vision tasks.

Jiawei Mao, Juncheng Wu, Yuyin Zhou et al.

There are many excellent solutions in image restoration.However, most methods require on training separate models to restore images with different types of degradation.Although existing all-in-one models effectively address multiple types of degradation simultaneously, their performance in real-world scenarios is still constrained by the task confusion problem.In this work, we attempt to address this issue by introducing \textbf{Restorer}, a novel Transformer-based all-in-one image restoration model.To effectively address the complex degradation present in real-world images, we propose All-Axis Attention (AAA), a mechanism that simultaneously models long-range dependencies across both spatial and channel dimensions, capturing potential correlations along all axes.Additionally, we introduce textual prompts in Restorer to incorporate explicit task priors, enabling the removal of specific degradation types based on user instructions. By iterating over these prompts, Restorer can handle composite degradation in real-world scenarios without requiring additional training.Based on these designs, Restorer with one set of parameters demonstrates state-of-the-art performance in multiple image restoration tasks compared to existing all-in-one and even single-task models.Additionally, Restorer is efficient during inference, suggesting the potential in real-world applications.

Jiawei Mao, Shujian Guo, Yuanqi Chang et al.

Masked autoencoders (MAEs) have displayed significant potential in the classification and semantic segmentation of medical images in the last year. Due to the high similarity of human tissues, even slight changes in medical images may represent diseased tissues, necessitating fine-grained inspection to pinpoint diseased tissues. The random masking strategy of MAEs is likely to result in areas of lesions being overlooked by the model. At the same time, inconsistencies between the pre-training and fine-tuning phases impede the performance and efficiency of MAE in medical image classification. To address these issues, we propose a medical supervised masked autoencoder (MSMAE) in this paper. In the pre-training phase, MSMAE precisely masks medical images via the attention maps obtained from supervised training, contributing to the representation learning of human tissue in the lesion area. During the fine-tuning phase, MSMAE is also driven by attention to the accurate masking of medical images. This improves the computational efficiency of the MSMAE while increasing the difficulty of fine-tuning, which indirectly improves the quality of MSMAE medical diagnosis. Extensive experiments demonstrate that MSMAE achieves state-of-the-art performance in case with three official medical datasets for various diseases. Meanwhile, transfer learning for MSMAE also demonstrates the great potential of our approach for medical semantic segmentation tasks. Moreover, the MSMAE accelerates the inference time in the fine-tuning phase by 11.2% and reduces the number of floating-point operations (FLOPs) by 74.08% compared to a traditional MAE.

Jiawei Mao, Xuesong Yin, Yuanqi Chang et al.

Weakly-supervised learning has become a popular technology in recent years. In this paper, we propose a novel medical image classification algorithm, called Weakly-Supervised Generative Adversarial Networks (WSGAN), which only uses a small number of real images without labels to generate fake images or mask images to enlarge the sample size of the training set. First, we combine with MixMatch to generate pseudo labels for the fake images and unlabeled images to do the classification. Second, contrastive learning and self-attention mechanism are introduced into the proposed problem to enhance the classification accuracy. Third, the problem of mode collapse is well addressed by cyclic consistency loss. Finally, we design global and local classifiers to complement each other with the key information needed for classification. The experimental results on four medical image datasets show that WSGAN can obtain relatively high learning performance by using few labeled and unlabeled data. For example, the classification accuracy of WSGAN is 11% higher than that of the second-ranked MIXMATCH with 100 labeled images and 1000 unlabeled images on the OCT dataset. In addition, we also conduct ablation experiments to verify the effectiveness of our algorithm.