Zhenlin Xu

Alexander H. Liu, Alexis Tacnet, Andy Ehrenberg et al. · deepmind, tsinghua

We introduce Voxtral TTS, an expressive multilingual text-to-speech model that generates natural speech from as little as 3 seconds of reference audio. Voxtral TTS adopts a hybrid architecture that combines auto-regressive generation of semantic speech tokens with flow-matching for acoustic tokens. These tokens are encoded and decoded with Voxtral Codec, a speech tokenizer trained from scratch with a hybrid VQ-FSQ quantization scheme. In human evaluations conducted by native speakers, Voxtral TTS is preferred for multilingual voice cloning due to its naturalness and expressivity, achieving a 68.4\% win rate over ElevenLabs Flash v2.5. We release the model weights under a CC BY-NC license.

Qin Liu, Zhenlin Xu, Gedas Bertasius et al.

Click-based interactive image segmentation aims at extracting objects with a limited user clicking. A hierarchical backbone is the de-facto architecture for current methods. Recently, the plain, non-hierarchical Vision Transformer (ViT) has emerged as a competitive backbone for dense prediction tasks. This design allows the original ViT to be a foundation model that can be finetuned for downstream tasks without redesigning a hierarchical backbone for pretraining. Although this design is simple and has been proven effective, it has not yet been explored for interactive image segmentation. To fill this gap, we propose SimpleClick, the first interactive segmentation method that leverages a plain backbone. Based on the plain backbone, we introduce a symmetric patch embedding layer that encodes clicks into the backbone with minor modifications to the backbone itself. With the plain backbone pretrained as a masked autoencoder (MAE), SimpleClick achieves state-of-the-art performance. Remarkably, our method achieves 4.15 NoC@90 on SBD, improving 21.8% over the previous best result. Extensive evaluation on medical images demonstrates the generalizability of our method. We further develop an extremely tiny ViT backbone for SimpleClick and provide a detailed computational analysis, highlighting its suitability as a practical annotation tool.

Zhenlin Xu, Marc Niethammer, Colin Raffel

Deep learning models struggle with compositional generalization, i.e. the ability to recognize or generate novel combinations of observed elementary concepts. In hopes of enabling compositional generalization, various unsupervised learning algorithms have been proposed with inductive biases that aim to induce compositional structure in learned representations (e.g. disentangled representation and emergent language learning). In this work, we evaluate these unsupervised learning algorithms in terms of how well they enable compositional generalization. Specifically, our evaluation protocol focuses on whether or not it is easy to train a simple model on top of the learned representation that generalizes to new combinations of compositional factors. We systematically study three unsupervised representation learning algorithms - $β$-VAE, $β$-TCVAE, and emergent language (EL) autoencoders - on two datasets that allow directly testing compositional generalization. We find that directly using the bottleneck representation with simple models and few labels may lead to worse generalization than using representations from layers before or after the learned representation itself. In addition, we find that the previously proposed metrics for evaluating the levels of compositionality are not correlated with actual compositional generalization in our framework. Surprisingly, we find that increasing pressure to produce a disentangled representation produces representations with worse generalization, while representations from EL models show strong compositional generalization. Taken together, our results shed new light on the compositional generalization behavior of different unsupervised learning algorithms with a new setting to rigorously test this behavior, and suggest the potential benefits of delevoping EL learning algorithms for more generalizable representations.

Yanbei Chen, Manchen Wang, Abhay Mittal et al.

Multi-dataset training provides a viable solution for exploiting heterogeneous large-scale datasets without extra annotation cost. In this work, we propose a scalable multi-dataset detector (ScaleDet) that can scale up its generalization across datasets when increasing the number of training datasets. Unlike existing multi-dataset learners that mostly rely on manual relabelling efforts or sophisticated optimizations to unify labels across datasets, we introduce a simple yet scalable formulation to derive a unified semantic label space for multi-dataset training. ScaleDet is trained by visual-textual alignment to learn the label assignment with label semantic similarities across datasets. Once trained, ScaleDet can generalize well on any given upstream and downstream datasets with seen and unseen classes. We conduct extensive experiments using LVIS, COCO, Objects365, OpenImages as upstream datasets, and 13 datasets from Object Detection in the Wild (ODinW) as downstream datasets. Our results show that ScaleDet achieves compelling strong model performance with an mAP of 50.7 on LVIS, 58.8 on COCO, 46.8 on Objects365, 76.2 on OpenImages, and 71.8 on ODinW, surpassing state-of-the-art detectors with the same backbone.

Zhenlin Xu, Yi Zhu, Tiffany Deng et al.

This paper presents novel benchmarks for evaluating vision-language models (VLMs) in zero-shot recognition, focusing on granularity and specificity. Although VLMs excel in tasks like image captioning, they face challenges in open-world settings. Our benchmarks test VLMs' consistency in understanding concepts across semantic granularity levels and their response to varying text specificity. Findings show that VLMs favor moderately fine-grained concepts and struggle with specificity, often misjudging texts that differ from their training data. Extensive evaluations reveal limitations in current VLMs, particularly in distinguishing between correct and subtly incorrect descriptions. While fine-tuning offers some improvements, it doesn't fully address these issues, highlighting the need for VLMs with enhanced generalization capabilities for real-world applications. This study provides insights into VLM limitations and suggests directions for developing more robust models.

Berk Iskender, Zhenlin Xu, Simon Kornblith et al.

Many contrastive representation learning methods learn a single global representation of an entire image. However, dense contrastive representation learning methods such as DenseCL (Wang et al., 2021) can learn better representations for tasks requiring stronger spatial localization of features, such as multi-label classification, detection, and segmentation. In this work, we study how to improve the quality of the representations learned by DenseCL by modifying the training scheme and objective function, and propose DenseCL++. We also conduct several ablation studies to better understand the effects of: (i) various techniques to form dense negative pairs among augmentations of different images, (ii) cross-view dense negative and positive pairs, and (iii) an auxiliary reconstruction task. Our results show 3.5% and 4% mAP improvement over SimCLR (Chen et al., 2020a) andDenseCL in COCO multi-label classification. In COCO and VOC segmentation tasks, we achieve 1.8% and 0.7% mIoU improvements over SimCLR, respectively.

Zhenlin Xu, Xiaogang Zhu, Yu Yao et al.

Modern agentic systems allow Large Language Model (LLM) agents to tackle complex tasks through extensive tool usage, forming structured control flows of tool selection and execution. Existing security analyses often treat these control flows as ephemeral, one-off sessions, overlooking the persistent influence of memory. This paper identifies a new threat from Memory Control Flow Attacks (MCFA) that memory retrieval can dominate the control flow, forcing unintended tool usage even against explicit user instructions and inducing persistent behavioral deviations across tasks. To understand the impact of this vulnerability, we further design MEMFLOW, an automated evaluation framework that systematically identifies and quantifies MCFA across heterogeneous tasks and long interaction horizons. To evaluate MEMFLOW, we attack state-of-the-art LLMs, including GPT-5 mini, Claude Sonnet 4.5 and Gemini 2.5 Flash on real-world tools from two major LLM agent development frameworks, LangChain and LlamaIndex. The results show that in general over 90% trials are vulnerable to MCFA even under strict safety constraints, highlighting critical security risks that demand immediate attention.

Haokun Liu, Gyung Hyun Je, Marco Ciccone et al.

The widespread availability of fine-tuned LoRA modules for open pre-trained models has led to an interest in methods that can adaptively merge LoRAs to improve performance. These methods typically include some way of selecting LoRAs from a pool and tune merging coefficients based on a task-specific dataset. While adaptive merging methods have demonstrated improvements in some settings, no past work has attempted to recycle LoRAs found "in the wild" on model repositories like the Hugging Face Hub. To address this gap, we consider recycling from a pool of nearly 1,000 user-contributed LoRAs trained from the Llama 3.1 8B-Instruct language model. Our empirical study includes a range of adaptive and non-adaptive merging methods in addition to a new method designed via a wide search over the methodological design space. We demonstrate that adaptive merging methods can improve performance over the base model but provide limited benefit over training a new LoRA on the same data used to set merging coefficients. We additionally find not only that the specific choice of LoRAs to merge has little importance, but that using LoRAs with randomly initialized parameter values yields similar performance. This raises the possibility that adaptive merging from recycled LoRAs primarily works via some kind of regularization effect, rather than by enabling positive cross-task transfer. To better understand why past work has proven successful, we confirm that positive transfer is indeed possible when there are highly relevant LoRAs in the pool. We release the model checkpoints and code online.

Yue Zhao, Hanwen Jiang, Zhenlin Xu et al.

Non-parametric quantization has received much attention due to its efficiency on parameters and scalability to a large codebook. In this paper, we present a unified formulation of different non-parametric quantization methods through the lens of lattice coding. The geometry of lattice codes explains the necessity of auxiliary loss terms when training auto-encoders with certain existing lookup-free quantization variants such as BSQ. As a step forward, we explore a few possible candidates, including random lattices, generalized Fibonacci lattices, and densest sphere packing lattices. Among all, we find the Leech lattice-based quantization method, which is dubbed as Spherical Leech Quantization ($Λ_{24}$-SQ), leads to both a simplified training recipe and an improved reconstruction-compression tradeoff thanks to its high symmetry and even distribution on the hypersphere. In image tokenization and compression tasks, this quantization approach achieves better reconstruction quality across all metrics than BSQ, the best prior art, while consuming slightly fewer bits. The improvement also extends to state-of-the-art auto-regressive image generation frameworks.

Zhengyang Shen, Zhenlin Xu, Sahin Olut et al.

We introduce a fluid-based image augmentation method for medical image analysis. In contrast to existing methods, our framework generates anatomically meaningful images via interpolation from the geodesic subspace underlying given samples. Our approach consists of three steps: 1) given a source image and a set of target images, we construct a geodesic subspace using the Large Deformation Diffeomorphic Metric Mapping (LDDMM) model; 2) we sample transformations from the resulting geodesic subspace; 3) we obtain deformed images and segmentations via interpolation. Experiments on brain (LPBA) and knee (OAI) data illustrate the performance of our approach on two tasks: 1) data augmentation during training and testing for image segmentation; 2) one-shot learning for single atlas image segmentation. We demonstrate that our approach generates anatomically meaningful data and improves performance on these tasks over competing approaches. Code is available at https://github.com/uncbiag/easyreg.

Jiahui Guang, Zexun Zhan, Zhenlin Xu et al.

Multimodal Large Language Models (MLLMs) may memorize sensitive cross-modal information during pretraining. However, existing MLLM unlearning benchmarks rely on synthetic knowledge injection or complete subject-level deletion, which fail to capture realistic, personalized deletion requests that require fine-grained factual control. In this paper, we introduce PPU-Bench, a real-world and fine-tuning-free benchmark for personalized partial unlearning in MLLMs. PPU-Bench contains 24K multimodal and unimodal samples derived from pre-existing knowledge of 500 public figures under three progressively challenging settings: Complete, Selective, and Personalized unlearning. The benchmark evaluates whether methods can remove target knowledge while preserving non-target facts, model utility, and cross-modal consistency. Extensive experiments show that Complete Unlearning often suppresses visual identity rather than factual knowledge, while Selective and Personalized Unlearning expose significant forget--retain trade-offs and challenges in intra-subject factual boundaries. Robustness analysis under cross-image and prompt-based attacks reveals distinct vulnerabilities across different unlearning settings. Motivated by these findings, we propose Boundary-Aware Optimization (BAO), which explicitly models intra-subject forget-retain boundaries. Experimental results on two representative methods demonstrate that BAO can effectively enforce intra-subject factual boundaries.

Zijia Lu, Bing Shuai, Yanbei Chen et al.

In this paper, we propose a novel concept of path consistency to learn robust object matching without using manual object identity supervision. Our key idea is that, to track a object through frames, we can obtain multiple different association results from a model by varying the frames it can observe, i.e., skipping frames in observation. As the differences in observations do not alter the identities of objects, the obtained association results should be consistent. Based on this rationale, we generate multiple observation paths, each specifying a different set of frames to be skipped, and formulate the Path Consistency Loss that enforces the association results are consistent across different observation paths. We use the proposed loss to train our object matching model with only self-supervision. By extensive experiments on three tracking datasets (MOT17, PersonPath22, KITTI), we demonstrate that our method outperforms existing unsupervised methods with consistent margins on various evaluation metrics, and even achieves performance close to supervised methods.

Pengfei Yu, Dongming Shen, Silin Meng et al.

We present RPGBench, the first benchmark designed to evaluate large language models (LLMs) as text-based role-playing game (RPG) engines. RPGBench comprises two core tasks: Game Creation (GC) and Game Simulation (GS). In GC, an LLM must craft a valid and playable RPG world using a structured event-state representation, ensuring logical coherence and proper termination conditions. In GS, the LLM simulates interactive gameplay across multiple rounds while consistently updating states and enforcing game rules. To comprehensively assess performance, RPGBench integrates objective and subjective evaluation methodologies. Objective measures verify adherence to event mechanics and check variable updates without requiring human intervention. Subjective measures, such as content interestingness, action quality, and role-playing capability, are evaluated via an LLM-as-a-judge framework, where a strong LLM grades each candidate's outputs. Empirical results demonstrate that state-of-the-art LLMs can produce engaging stories but often struggle to implement consistent, verifiable game mechanics, particularly in long or complex scenarios. By combining structured, rule-based assessments with LLM-based judgments, RPGBench provides a new standard for evaluating how well LLMs can balance creativity, coherence, and complexity in text-based RPGs, opening avenues for more immersive and controllable interactive storytelling.

Nikhil Kandpal, Brian Lester, Colin Raffel et al. · allen-ai, cmu

Large language models (LLMs) are typically trained on enormous quantities of unlicensed text, a practice that has led to scrutiny due to possible intellectual property infringement and ethical concerns. Training LLMs on openly licensed text presents a first step towards addressing these issues, but prior data collection efforts have yielded datasets too small or low-quality to produce performant LLMs. To address this gap, we collect, curate, and release the Common Pile v0.1, an eight terabyte collection of openly licensed text designed for LLM pretraining. The Common Pile comprises content from 30 sources that span diverse domains including research papers, code, books, encyclopedias, educational materials, audio transcripts, and more. Crucially, we validate our efforts by training two 7 billion parameter LLMs on text from the Common Pile: Comma v0.1-1T and Comma v0.1-2T, trained on 1 and 2 trillion tokens respectively. Both models attain competitive performance to LLMs trained on unlicensed text with similar computational budgets, such as Llama 1 and 2 7B. In addition to releasing the Common Pile v0.1 itself, we also release the code used in its creation as well as the training mixture and checkpoints for the Comma v0.1 models.

Yicheng Wang, Zhikang Zhang, Jue Wang et al. · amazon-science

In various video-language learning tasks, the challenge of achieving cross-modality alignment with multi-grained data persists. We propose a method to tackle this challenge from two crucial perspectives: data and modeling. Given the absence of a multi-grained video-text pretraining dataset, we introduce a Granularity EXpansion (GEX) method with Integration and Compression operations to expand the granularity of a single-grained dataset. To better model multi-grained data, we introduce an Iterative Approximation Module (IAM), which embeds multi-grained videos and texts into a unified, low-dimensional semantic space while preserving essential information for cross-modal alignment. Furthermore, GEXIA is highly scalable with no restrictions on the number of video-text granularities for alignment. We evaluate our work on three categories of video tasks across seven benchmark datasets, showcasing state-of-the-art or comparable performance. Remarkably, our model excels in tasks involving long-form video understanding, even though the pretraining dataset only contains short video clips.

Alexander H. Liu, Andy Ehrenberg, Andy Lo et al.

We introduce Voxtral Realtime, a natively streaming automatic speech recognition model that matches offline transcription quality at sub-second latency. Unlike approaches that adapt offline models through chunking or sliding windows, Voxtral Realtime is trained end-to-end for streaming, with explicit alignment between audio and text streams. Our architecture builds on the Delayed Streams Modeling framework, introducing a new causal audio encoder and Ada RMS-Norm for improved delay conditioning. We scale pretraining to a large-scale dataset spanning 13 languages. At a delay of 480ms, Voxtral Realtime achieves performance on par with Whisper, the most widely deployed offline transcription system. We release the model weights under the Apache 2.0 license.

Saeed Rahmani, Zhenlin Xu, Simeon C. Calvert et al.

The integration of automated vehicles (AVs) into transportation systems presents an unprecedented opportunity to enhance road safety and efficiency. However, understanding the interactions between AVs and human-driven vehicles (HVs) at intersections remains an open research question. This study aims to bridge this gap by examining behavioral differences and adaptations of AVs and HVs at unsignalized intersections by utilizing two large-scale AV datasets from Waymo and Lyft. By using a systematic methodology, the research identifies and analyzes merging and crossing conflicts by calculating key safety and efficiency metrics, including time to collision (TTC), post-encroachment time (PET), maximum required deceleration (MRD), time advantage (TA), and speed and acceleration profiles. Through this approach, the study assesses the safety and efficiency implications of these behavioral differences and adaptations for mixed-autonomy traffic. The findings reveal a paradox: while AVs maintain larger safety margins, their conservative behavior can lead to unexpected situations for human drivers, potentially causing unsafe conditions. From a performance point of view, human drivers tend to exhibit more consistent behavior when interacting with AVs versus other HVs, suggesting AVs may contribute to harmonizing traffic flow patterns. Moreover, notable differences were observed between Waymo and Lyft vehicles, which highlights the importance of considering manufacturer-specific AV behaviors in traffic modeling and management strategies for the safe integration of AVs. The processed dataset, as well as the developed algorithms and scripts, are openly published to foster research on AV-HV interactions.

Qin Liu, Zhenlin Xu, Yining Jiao et al.

We propose iSegFormer, a memory-efficient transformer that combines a Swin transformer with a lightweight multilayer perceptron (MLP) decoder. With the efficient Swin transformer blocks for hierarchical self-attention and the simple MLP decoder for aggregating both local and global attention, iSegFormer learns powerful representations while achieving high computational efficiencies. Specifically, we apply iSegFormer to interactive 3D medical image segmentation.

Xu Han, Zhengyang Shen, Zhenlin Xu et al.

Registration of images with pathologies is challenging due to tissue appearance changes and missing correspondences caused by the pathologies. Moreover, mass effects as observed for brain tumors may displace tissue, creating larger deformations over time than what is observed in a healthy brain. Deep learning models have successfully been applied to image registration to offer dramatic speed up and to use surrogate information (e.g., segmentations) during training. However, existing approaches focus on learning registration models using images from healthy patients. They are therefore not designed for the registration of images with strong pathologies for example in the context of brain tumors, and traumatic brain injuries. In this work, we explore a deep learning approach to register images with brain tumors to an atlas. Our model learns an appearance mapping from images with tumors to the atlas, while simultaneously predicting the transformation to atlas space. Using separate decoders, the network disentangles the tumor mass effect from the reconstruction of quasi-normal images. Results on both synthetic and real brain tumor scans show that our approach outperforms cost function masking for registration to the atlas and that reconstructed quasi-normal images can be used for better longitudinal registrations.

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.

Zhenlin Xu, Marc Niethammer

Deep convolutional neural networks (CNNs) are state-of-the-art for semantic image segmentation, but typically require many labeled training samples. Obtaining 3D segmentations of medical images for supervised training is difficult and labor intensive. Motivated by classical approaches for joint segmentation and registration we therefore propose a deep learning framework that jointly learns networks for image registration and image segmentation. In contrast to previous work on deep unsupervised image registration, which showed the benefit of weak supervision via image segmentations, our approach can use existing segmentations when available and computes them via the segmentation network otherwise, thereby providing the same registration benefit. Conversely, segmentation network training benefits from the registration, which essentially provides a realistic form of data augmentation. Experiments on knee and brain 3D magnetic resonance (MR) images show that our approach achieves large simultaneous improvements of segmentation and registration accuracy (over independently trained networks) and allows training high-quality models with very limited training data. Specifically, in a one-shot-scenario (with only one manually labeled image) our approach increases Dice scores (%) over an unsupervised registration network by 2.7 and 1.8 on the knee and brain images respectively.

Zhengyang Shen, Xu Han, Zhenlin Xu et al.

We introduce an end-to-end deep-learning framework for 3D medical image registration. In contrast to existing approaches, our framework combines two registration methods: an affine registration and a vector momentum-parameterized stationary velocity field (vSVF) model. Specifically, it consists of three stages. In the first stage, a multi-step affine network predicts affine transform parameters. In the second stage, we use a Unet-like network to generate a momentum, from which a velocity field can be computed via smoothing. Finally, in the third stage, we employ a self-iterable map-based vSVF component to provide a non-parametric refinement based on the current estimate of the transformation map. Once the model is trained, a registration is completed in one forward pass. To evaluate the performance, we conducted longitudinal and cross-subject experiments on 3D magnetic resonance images (MRI) of the knee of the Osteoarthritis Initiative (OAI) dataset. Results show that our framework achieves comparable performance to state-of-the-art medical image registration approaches, but it is much faster, with a better control of transformation regularity including the ability to produce approximately symmetric transformations, and combining affine and non-parametric registration.