Junlin Yang

Bohan Lyu, Yucheng Yang, Siqiao Huang et al.

Modern AI progress has been driven by ML methods that are generalizable across settings and scalable to larger regimes. As large language models demonstrate advanced capabilities in reasoning, coding, and engineering tasks, it is increasingly important to understand whether they can discover such methods rather than only apply existing ones. We introduce MLS-Bench, a benchmark for evaluating whether AI systems can invent generalizable and scalable ML methods. MLS-Bench contains 140 tasks across 12 domains, each requiring an agent to improve one targeted component of an ML system or algorithm and demonstrate that the improvement generalizes across controlled settings and scales. We find that current agents remain far from reliably surpassing human-designed methods, and that engineering-style tuning is easier for them than genuine method invention. We further study the effects of test-time scaling, adaptive compute allocation, and context provision on agents' discovery performance, together with case studies of their behavior. Our analyses suggest that the bottleneck is not only in proposing new methods, but also in the scientific insight needed to plan, validate, and scale claims about them. More search, compute, or context alone does not remove this bottleneck. We build and maintain a community platform for cumulative and comparable iteration, and release the data and code at https://mls-bench.com.

Junlin Yang, Dylan Zhang, Xiangchen Song et al.

We introduce CausaLab, a scalable environment for evaluating interactive causal discovery by LLM agents. Unlike prior evaluations, CausaLab evaluates both whether an agent can solve a problem using causal evidence and whether its answer is supported by a correct hypothesis about the underlying causal mechanism. Each episode places an agent in a synthetic laboratory: it receives prior measurement records, intervenes on a manipulator crystal, and predicts the resonance frequency of a held-out reactor crystal governed by the same mechanism. The hidden data-generating process is a randomly sampled structural causal model (SCM), so success requires recovering both a causal graph and structural equations rather than recalling prior knowledge. CausaLab also includes a domain-specific language that records the agent's evolving SCM hypothesis, making trajectories inspectable and comparable with ground truth. Experiments show a persistent gap between prediction and mechanism recovery: in the purely observational 6-node setting, GPT-5.2-high reaches 92% task accuracy but only 0.471 all-edge $F_1$. This observation further motivates our exploration of different interaction strategies: Mixed observation--intervention strategies improve structural fidelity: in the mixed 6-node setting, GPT-5.2-high achieves 80% on both task accuracy and all-edge $F_1$. Yet even strong agents struggle to design informative interventions, as pure intervention strategies perform poorly on both task accuracy and all-edge $F_1$. We identify premature stopping as a major weakness of agents, and show that asking the model to verify the consistency between its hypothesis and past data can help mitigate this issue. CausaLab therefore separates predictive success from causal understanding and exposes current LLM agents' limits as experimental causal reasoners.

Juntang Zhuang, Nicha C. Dvornek, Xiaoxiao Li et al.

Deep neural networks are vulnerable to adversarial attacks and hard to interpret because of their black-box nature. The recently proposed invertible network is able to accurately reconstruct the inputs to a layer from its outputs, thus has the potential to unravel the black-box model. An invertible network classifier can be viewed as a two-stage model: (1) invertible transformation from input space to the feature space; (2) a linear classifier in the feature space. We can determine the decision boundary of a linear classifier in the feature space; since the transform is invertible, we can invert the decision boundary from the feature space to the input space. Furthermore, we propose to determine the projection of a data point onto the decision boundary, and define explanation as the difference between data and its projection. Finally, we propose to locally approximate a neural network with its first-order Taylor expansion, and define feature importance using a local linear model. We provide the implementation of our method: \url{https://github.com/juntang-zhuang/explain_invertible}.

Juntang Zhuang, Junlin Yang, Lin Gu et al.

In this paper, we present ShelfNet, a novel architecture for accurate fast semantic segmentation. Different from the single encoder-decoder structure, ShelfNet has multiple encoder-decoder branch pairs with skip connections at each spatial level, which looks like a shelf with multiple columns. The shelf-shaped structure can be viewed as an ensemble of multiple deep and shallow paths, thus improving accuracy. We significantly reduce computation burden by reducing channel number, at the same time achieving high accuracy with this unique structure. In addition, we propose a shared-weight strategy in the residual block which reduces parameter number without sacrificing performance. Compared with popular non real-time methods such as PSPNet, our ShelfNet achieves 4$\times$ faster inference speed with similar accuracy on PASCAL VOC dataset. Compared with real-time segmentation models such as BiSeNet, our model achieves higher accuracy at comparable speed on the Cityscapes Dataset, enabling the application in speed-demanding tasks such as street-scene understanding for autonomous driving. Furthermore, our ShelfNet achieves 79.0\% mIoU on Cityscapes Dataset with ResNet34 backbone, outperforming PSPNet and BiSeNet with large backbones such as ResNet101. Through extensive experiments, we validated the superior performance of ShelfNet. We provide link to the implementation \url{https://github.com/juntang-zhuang/ShelfNet-lw-cityscapes}.

Dunjie Lu, Yiheng Xu, Junli Wang et al.

Training computer-use agents requires massive amounts of GUI interaction data, but manually annotating action trajectories at scale is prohibitively expensive. We present VideoAgentTrek, a scalable pipeline that automatically mines training data from publicly available screen-recorded videos at web scale, eliminating the need for manual annotation. Our approach addresses a key challenge: raw videos contain implicit demonstrations but lack explicit action labels. To solve this, we develop Video2Action, an inverse dynamics module (IDM) with two components: (1) a video grounding model that detects and localizes GUI actions with precise temporal boundaries and context, and (2) an action-content recognizer that extracts structured parameters like click coordinates and typed text with high fidelity. Applied to 39,000 YouTube tutorial videos, our pipeline generates 1.52 million interaction steps automatically. We leverage this data through continued pretraining followed by supervised fine-tuning. On OSWorld-Verified, our approach improves task success rates from 9.3% (SFT-only baseline) to 15.8%, a 70% relative improvement. On AgentNetBench, step accuracy increases from 64.1% to 69.3%. Our results demonstrate that passive internet videos can be transformed into high-quality supervision for computer-use agents, providing a scalable alternative to expensive manual annotation.

Daiqing Li, Junlin Yang, Karsten Kreis et al.

Training deep networks with limited labeled data while achieving a strong generalization ability is key in the quest to reduce human annotation efforts. This is the goal of semi-supervised learning, which exploits more widely available unlabeled data to complement small labeled data sets. In this paper, we propose a novel framework for discriminative pixel-level tasks using a generative model of both images and labels. Concretely, we learn a generative adversarial network that captures the joint image-label distribution and is trained efficiently using a large set of unlabeled images supplemented with only few labeled ones. We build our architecture on top of StyleGAN2, augmented with a label synthesis branch. Image labeling at test time is achieved by first embedding the target image into the joint latent space via an encoder network and test-time optimization, and then generating the label from the inferred embedding. We evaluate our approach in two important domains: medical image segmentation and part-based face segmentation. We demonstrate strong in-domain performance compared to several baselines, and are the first to showcase extreme out-of-domain generalization, such as transferring from CT to MRI in medical imaging, and photographs of real faces to paintings, sculptures, and even cartoons and animal faces. Project Page: \url{https://nv-tlabs.github.io/semanticGAN/}

Chenyu You, Junlin Yang, Julius Chapiro et al.

Deep neural networks have shown exceptional learning capability and generalizability in the source domain when massive labeled data is provided. However, the well-trained models often fail in the target domain due to the domain shift. Unsupervised domain adaptation aims to improve network performance when applying robust models trained on medical images from source domains to a new target domain. In this work, we present an approach based on the Wasserstein distance guided disentangled representation to achieve 3D multi-domain liver segmentation. Concretely, we embed images onto a shared content space capturing shared feature-level information across domains and domain-specific appearance spaces. The existing mutual information-based representation learning approaches often fail to capture complete representations in multi-domain medical imaging tasks. To mitigate these issues, we utilize Wasserstein distance to learn more complete representation, and introduces a content discriminator to further facilitate the representation disentanglement. Experiments demonstrate that our method outperforms the state-of-the-art on the multi-modality liver segmentation task.

Zhenlin Xu, Deyi Liu, Junlin Yang et al.

While successful for various computer vision tasks, deep neural networks have shown to be vulnerable to texture style shifts and small perturbations to which humans are robust. In this work, we show that the robustness of neural networks can be greatly improved through the use of random convolutions as data augmentation. Random convolutions are approximately shape-preserving and may distort local textures. Intuitively, randomized convolutions create an infinite number of new domains with similar global shapes but random local textures. Therefore, we explore using outputs of multi-scale random convolutions as new images or mixing them with the original images during training. When applying a network trained with our approach to unseen domains, our method consistently improves the performance on domain generalization benchmarks and is scalable to ImageNet. In particular, in the challenging scenario of generalizing to the sketch domain in PACS and to ImageNet-Sketch, our method outperforms state-of-art methods by a large margin. More interestingly, our method can benefit downstream tasks by providing a more robust pretrained visual representation.

Junlin Yang, Nicha C. Dvornek, Fan Zhang et al.

This work proposes a pipeline to predict treatment response to intra-arterial therapy of patients with Hepatocellular Carcinoma (HCC) for improved therapeutic decision-making. Our graph neural network model seamlessly combines heterogeneous inputs of baseline MR scans, pre-treatment clinical information, and planned treatment characteristics and has been validated on patients with HCC treated by transarterial chemoembolization (TACE). It achieves Accuracy of $0.713 \pm 0.075$, F1 of $0.702 \pm 0.082$ and AUC of $0.710 \pm 0.108$. In addition, the pipeline incorporates uncertainty estimation to select hard cases and most align with the misclassified cases. The proposed pipeline arrives at more informed intra-arterial therapeutic decisions for patients with HCC via improving model accuracy and incorporating uncertainty estimation.

Junlin Yang, Nicha C. Dvornek, Fan Zhang et al.

Domain Adaptation (DA) has the potential to greatly help the generalization of deep learning models. However, the current literature usually assumes to transfer the knowledge from the source domain to a specific known target domain. Domain Agnostic Learning (DAL) proposes a new task of transferring knowledge from the source domain to data from multiple heterogeneous target domains. In this work, we propose the Domain-Agnostic Learning framework with Anatomy-Consistent Embedding (DALACE) that works on both domain-transfer and task-transfer to learn a disentangled representation, aiming to not only be invariant to different modalities but also preserve anatomical structures for the DA and DAL tasks in cross-modality liver segmentation. We validated and compared our model with state-of-the-art methods, including CycleGAN, Task Driven Generative Adversarial Network (TD-GAN), and Domain Adaptation via Disentangled Representations (DADR). For the DA task, our DALACE model outperformed CycleGAN, TD-GAN ,and DADR with DSC of 0.847 compared to 0.721, 0.793 and 0.806. For the DAL task, our model improved the performance with DSC of 0.794 from 0.522, 0.719 and 0.742 by CycleGAN, TD-GAN, and DADR. Further, we visualized the success of disentanglement, which added human interpretability of the learned meaningful representations. Through ablation analysis, we specifically showed the concrete benefits of disentanglement for downstream tasks and the role of supervision for better disentangled representation with segmentation consistency to be invariant to domains with the proposed Domain-Agnostic Module (DAM) and to preserve anatomical information with the proposed Anatomy-Preserving Module (APM).

Junlin Yang, Nicha C. Dvornek, Fan Zhang et al.

A deep learning model trained on some labeled data from a certain source domain generally performs poorly on data from different target domains due to domain shifts. Unsupervised domain adaptation methods address this problem by alleviating the domain shift between the labeled source data and the unlabeled target data. In this work, we achieve cross-modality domain adaptation, i.e. between CT and MRI images, via disentangled representations. Compared to learning a one-to-one mapping as the state-of-art CycleGAN, our model recovers a many-to-many mapping between domains to capture the complex cross-domain relations. It preserves semantic feature-level information by finding a shared content space instead of a direct pixelwise style transfer. Domain adaptation is achieved in two steps. First, images from each domain are embedded into two spaces, a shared domain-invariant content space and a domain-specific style space. Next, the representation in the content space is extracted to perform a task. We validated our method on a cross-modality liver segmentation task, to train a liver segmentation model on CT images that also performs well on MRI. Our method achieved Dice Similarity Coefficient (DSC) of 0.81, outperforming a CycleGAN-based method of 0.72. Moreover, our model achieved good generalization to joint-domain learning, in which unpaired data from different modalities are jointly learned to improve the segmentation performance on each individual modality. Lastly, under a multi-modal target domain with significant diversity, our approach exhibited the potential for diverse image generation and remained effective with DSC of 0.74 on multi-phasic MRI while the CycleGAN-based method performed poorly with a DSC of only 0.52.

Yinchuan Xu, Junlin Yang

Gender bias has been found in existing coreference resolvers. In order to eliminate gender bias, a gender-balanced dataset Gendered Ambiguous Pronouns (GAP) has been released and the best baseline model achieves only 66.9% F1. Bidirectional Encoder Representations from Transformers (BERT) has broken several NLP task records and can be used on GAP dataset. However, fine-tune BERT on a specific task is computationally expensive. In this paper, we propose an end-to-end resolver by combining pre-trained BERT with Relational Graph Convolutional Network (R-GCN). R-GCN is used for digesting structural syntactic information and learning better task-specific embeddings. Empirical results demonstrate that, under explicit syntactic supervision and without the need to fine tune BERT, R-GCN's embeddings outperform the original BERT embeddings on the coreference task. Our work significantly improves the snippet-context baseline F1 score on GAP dataset from 66.9% to 80.3%. We participated in the 2019 GAP Coreference Shared Task, and our codes are available online.