Bo Zhao

Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo et al. · tsinghua

While next-token prediction is considered a promising path towards artificial general intelligence, it has struggled to excel in multimodal tasks, which are still dominated by diffusion models (e.g., Stable Diffusion) and compositional approaches (e.g., CLIP combined with LLMs). In this paper, we introduce Emu3, a new suite of state-of-the-art multimodal models trained solely with next-token prediction. By tokenizing images, text, and videos into a discrete space, we train a single transformer from scratch on a mixture of multimodal sequences. Emu3 outperforms several well-established task-specific models in both generation and perception tasks, surpassing flagship models such as SDXL and LLaVA-1.6, while eliminating the need for diffusion or compositional architectures. Emu3 is also capable of generating high-fidelity video via predicting the next token in a video sequence. We simplify complex multimodal model designs by converging on a singular focus: tokens, unlocking great potential for scaling both during training and inference. Our results demonstrate that next-token prediction is a promising path towards building general multimodal intelligence beyond language. We open-source key techniques and models to support further research in this direction.

Lei Zhang, Jie Zhang, Bowen Lei et al. · microsoft-research

Dataset Distillation (DD), a newly emerging field, aims at generating much smaller but efficient synthetic training datasets from large ones. Existing DD methods based on gradient matching achieve leading performance; however, they are extremely computationally intensive as they require continuously optimizing a dataset among thousands of randomly initialized models. In this paper, we assume that training the synthetic data with diverse models leads to better generalization performance. Thus we propose two model augmentation techniques, i.e. using early-stage models and parameter perturbation to learn an informative synthetic set with significantly reduced training cost. Extensive experiments demonstrate that our method achieves up to 20x speedup and comparable performance on par with state-of-the-art methods.

Bo Zhao, Boya Wu, Muyang He et al.

Thanks to the emerging of foundation models, the large language and vision models are integrated to acquire the multimodal ability of visual captioning, question answering, etc. Although existing multimodal models present impressive performance of visual understanding and reasoning, their limits are still largely under-explored due to the scarcity of high-quality instruction tuning data. To push the limits of multimodal capability, we Scale up Visual Instruction Tuning (SVIT) by constructing a dataset of 4.2 million visual instruction tuning data including 1.6M conversation question-answer (QA) pairs, 1.6M complex reasoning QA pairs, 1.0M referring QA pairs and 106K detailed image descriptions. Besides the volume, the proposed dataset is also featured by the high quality and rich diversity, which is generated by prompting GPT-4 with the abundant manual annotations of images. We also propose a new data recipe to select subset with better diversity and balance, which evokes model's superior capabilities. Extensive experiments verify that SVIT-v1.5, trained on the proposed dataset, outperforms state-of-the-art Multimodal Large Language Models on popular benchmarks. The data and code are publicly available at https://github.com/BAAI-DCAI/Visual-Instruction-Tuning.

Peter Eckmann, Kunyang Sun, Bo Zhao et al.

Generation of drug-like molecules with high binding affinity to target proteins remains a difficult and resource-intensive task in drug discovery. Existing approaches primarily employ reinforcement learning, Markov sampling, or deep generative models guided by Gaussian processes, which can be prohibitively slow when generating molecules with high binding affinity calculated by computationally-expensive physics-based methods. We present Latent Inceptionism on Molecules (LIMO), which significantly accelerates molecule generation with an inceptionism-like technique. LIMO employs a variational autoencoder-generated latent space and property prediction by two neural networks in sequence to enable faster gradient-based reverse-optimization of molecular properties. Comprehensive experiments show that LIMO performs competitively on benchmark tasks and markedly outperforms state-of-the-art techniques on the novel task of generating drug-like compounds with high binding affinity, reaching nanomolar range against two protein targets. We corroborate these docking-based results with more accurate molecular dynamics-based calculations of absolute binding free energy and show that one of our generated drug-like compounds has a predicted $K_D$ (a measure of binding affinity) of $6 \cdot 10^{-14}$ M against the human estrogen receptor, well beyond the affinities of typical early-stage drug candidates and most FDA-approved drugs to their respective targets. Code is available at https://github.com/Rose-STL-Lab/LIMO.

Muyang He, Shuo Yang, Tiejun Huang et al.

The state of the art of many learning tasks, e.g., image classification, is advanced by collecting larger datasets and then training larger models on them. As the outcome, the increasing computational cost is becoming unaffordable. In this paper, we investigate how to prune the large-scale datasets, and thus produce an informative subset for training sophisticated deep models with negligible performance drop. We propose a simple yet effective dataset pruning method by exploring both the prediction uncertainty and training dynamics. We study dataset pruning by measuring the variation of predictions during the whole training process on large-scale datasets, i.e., ImageNet-1K and ImageNet-21K, and advanced models, i.e., Swin Transformer and ConvNeXt. Extensive experimental results indicate that our method outperforms the state of the art and achieves 25% lossless pruning ratio on both ImageNet-1K and ImageNet-21K. The code and pruned datasets are available at https://github.com/BAAI-DCAI/Dataset-Pruning.

Xiang Li, Yiqun Yao, Xin Jiang et al.

Large Language Models (LLMs) represent a significant stride toward Artificial General Intelligence. As scaling laws underscore the potential of increasing model sizes, the academic community has intensified its investigations into LLMs with capacities exceeding 50 billion parameters. This technical report builds on our prior work with Tele-FLM (also known as FLM-2), a publicly available 52-billion-parameter model. We delve into two primary areas: we first discuss our observation of Supervised Fine-tuning (SFT) on Tele-FLM-52B, which supports the "less is more" approach for SFT data construction; second, we demonstrate our experiments and analyses on the best practices for progressively growing a model from 52 billion to 102 billion, and subsequently to 1 trillion parameters. We will open-source a 1T model checkpoint, namely Tele-FLM-1T, to advance further training and research.

Sidan Yin, Bo Zhao

Synthetic and manipulated speech can reduce the reliability of automatic speaker verification systems, so anti-spoofing methods need to be both accurate and efficient in training and inference. This paper focuses on the ASVspoof 5 Track 1 closed condition, where standard cross-entropy training may not give enough attention to hard trials and is not directly aligned with ranking- and threshold-based evaluation metrics. We propose TFPARN, a Transformer-based focal-pairwise attentive ranking network. The system extracts log-Mel features from speech, uses a Transformer encoder to model frame-level information, applies attention pooling to obtain utterance-level representations, and is trained with a combination of focal classification loss and pairwise ranking loss. RawBoost augmentation is used during training, and test-time augmentation is applied during evaluation to improve robustness. Compared with re-implemented AASIST and RawNet2 baselines under the same protocol, TFPARN achieves the best results, with a minDCF of 0.2430 and an EER of 12.52%. Ablation experiments further show that the pairwise loss, focal loss, and attention pooling all improve performance. TFPARN also uses the lowest inference memory among the compared systems, at 1.4 GB, runs at about 0.79 ms per utterance, and reaches its best checkpoint in less training time than AASIST. These results show that TFPARN provides a good balance between detection accuracy and computational cost for logical access anti-spoofing.

Yuxin Du, Fan Bai, Tiejun Huang et al.

Precise image segmentation provides clinical study with instructive information. Despite the remarkable progress achieved in medical image segmentation, there is still an absence of a 3D foundation segmentation model that can segment a wide range of anatomical categories with easy user interaction. In this paper, we propose a 3D foundation segmentation model, named SegVol, supporting universal and interactive volumetric medical image segmentation. By scaling up training data to 90K unlabeled Computed Tomography (CT) volumes and 6K labeled CT volumes, this foundation model supports the segmentation of over 200 anatomical categories using semantic and spatial prompts. To facilitate efficient and precise inference on volumetric images, we design a zoom-out-zoom-in mechanism. Extensive experiments on 22 anatomical segmentation tasks verify that SegVol outperforms the competitors in 19 tasks, with improvements up to 37.24% compared to the runner-up methods. We demonstrate the effectiveness and importance of specific designs by ablation study. We expect this foundation model can promote the development of volumetric medical image analysis. The model and code are publicly available at: https://github.com/BAAI-DCAI/SegVol.

Xuan Rao, Mingming Ha, Bo Zhao et al.

Class-incremental learning (CIL) with Vision Transformers (ViTs) faces a major computational bottleneck during the classifier reconstruction phase, where most existing methods rely on costly iterative stochastic gradient descent (SGD). We observe that analytic Regularized Gaussian Discriminant Analysis (RGDA) provides a Bayes-optimal alternative with accuracy comparable to SGD-based classifiers; however, its quadratic inference complexity limits its use in large-scale CIL scenarios. To overcome this, we propose Low-Rank Factorized RGDA (LR-RGDA), a scalable classifier that combines RGDA's expressivity with the efficiency of linear classifiers. By exploiting the low-rank structure of the covariance via the Woodbury matrix identity, LR-RGDA decomposes the discriminant function into a global affine term refined by a low-rank quadratic perturbation, reducing the inference complexity from $\mathcal{O}(Cd^2)$ to $\mathcal{O}(d^2 + Crd^2)$, where $C$ is the class number, $d$ the feature dimension, and $r \ll d$ the subspace rank. To mitigate representation drift caused by backbone updates, we further introduce Hopfield-based Distribution Compensator (HopDC), a training-free mechanism that uses modern continuous Hopfield Networks to recalibrate historical class statistics through associative memory dynamics on unlabeled anchors, accompanied by a theoretical bound on the estimation error. Extensive experiments on diverse CIL benchmarks demonstrate that our framework achieves state-of-the-art performance, providing a scalable solution for large-scale class-incremental learning with ViTs. Code: https://github.com/raoxuan98-hash/lr_rgda_hopdc.

Xuan Rao, Simian Xu, Zheng Li et al.

Recent advances have shown that sequential fine-tuning (SeqFT) of pre-trained vision transformers (ViTs), followed by classifier refinement using approximate distributions of class features, can be an effective strategy for class-incremental learning (CIL). However, this approach is susceptible to distribution drift, caused by the sequential optimization of shared backbone parameters. This results in a mismatch between the distributions of the previously learned classes and that of the updater model, ultimately degrading the effectiveness of classifier performance over time. To address this issue, we introduce a latent space transition operator and propose Sequential Learning with Drift Compensation (SLDC). SLDC aims to align feature distributions across tasks to mitigate the impact of drift. First, we present a linear variant of SLDC, which learns a linear operator by solving a regularized least-squares problem that maps features before and after fine-tuning. Next, we extend this with a weakly nonlinear SLDC variant, which assumes that the ideal transition operator lies between purely linear and fully nonlinear transformations. This is implemented using learnable, weakly nonlinear mappings that balance flexibility and generalization. To further reduce representation drift, we apply knowledge distillation (KD) in both algorithmic variants. Extensive experiments on standard CIL benchmarks demonstrate that SLDC significantly improves the performance of SeqFT. Notably, by combining KD to address representation drift with SLDC to compensate distribution drift, SeqFT achieves performance comparable to joint training across all evaluated datasets. Code: https://github.com/raoxuan98-hash/sldc.git.

Jie Zhang, Xiaohua Qi, Bo Zhao

Existing approaches in Federated Learning (FL) mainly focus on sending model parameters or gradients from clients to a server. However, these methods are plagued by significant inefficiency, privacy, and security concerns. Thanks to the emerging foundation generative models, we propose a novel federated learning framework, namely Federated Generative Learning. In this framework, each client can create text embeddings that are tailored to their local data, and send embeddings to the server. Then the informative training data can be synthesized remotely on the server using foundation generative models with these embeddings, which can benefit FL tasks. Our proposed framework offers several advantages, including increased communication efficiency, robustness to data heterogeneity, substantial performance improvements, and enhanced privacy protection. We validate these benefits through extensive experiments conducted on 12 datasets. For example, on the ImageNet100 dataset with a highly skewed data distribution, our method outperforms FedAvg by 12% in a single communication round, compared to FedAvg's performance over 200 communication rounds. We have released the code for all experiments conducted in this study.

Salva Rühling Cachay, Bo Zhao, Hailey Joren et al.

While diffusion models can successfully generate data and make predictions, they are predominantly designed for static images. We propose an approach for efficiently training diffusion models for probabilistic spatiotemporal forecasting, where generating stable and accurate rollout forecasts remains challenging, Our method, DYffusion, leverages the temporal dynamics in the data, directly coupling it with the diffusion steps in the model. We train a stochastic, time-conditioned interpolator and a forecaster network that mimic the forward and reverse processes of standard diffusion models, respectively. DYffusion naturally facilitates multi-step and long-range forecasting, allowing for highly flexible, continuous-time sampling trajectories and the ability to trade-off performance with accelerated sampling at inference time. In addition, the dynamics-informed diffusion process in DYffusion imposes a strong inductive bias and significantly improves computational efficiency compared to traditional Gaussian noise-based diffusion models. Our approach performs competitively on probabilistic forecasting of complex dynamics in sea surface temperatures, Navier-Stokes flows, and spring mesh systems.

Tian Dong, Bo Zhao, Lingjuan Lyu

To prevent unintentional data leakage, research community has resorted to data generators that can produce differentially private data for model training. However, for the sake of the data privacy, existing solutions suffer from either expensive training cost or poor generalization performance. Therefore, we raise the question whether training efficiency and privacy can be achieved simultaneously. In this work, we for the first time identify that dataset condensation (DC) which is originally designed for improving training efficiency is also a better solution to replace the traditional data generators for private data generation, thus providing privacy for free. To demonstrate the privacy benefit of DC, we build a connection between DC and differential privacy, and theoretically prove on linear feature extractors (and then extended to non-linear feature extractors) that the existence of one sample has limited impact ($O(m/n)$) on the parameter distribution of networks trained on $m$ samples synthesized from $n (n \gg m)$ raw samples by DC. We also empirically validate the visual privacy and membership privacy of DC-synthesized data by launching both the loss-based and the state-of-the-art likelihood-based membership inference attacks. We envision this work as a milestone for data-efficient and privacy-preserving machine learning.

Bo Zhao, Hakan Bilen

Remarkable progress has been achieved in synthesizing photo-realistic images with generative adversarial networks (GANs). Recently, GANs are utilized as the training sample generator when obtaining or storing real training data is expensive even infeasible. However, traditional GANs generated images are not as informative as the real training samples when being used to train deep neural networks. In this paper, we propose a novel method to synthesize Informative Training samples with GAN (IT-GAN). Specifically, we freeze a pre-trained GAN model and learn the informative latent vectors that correspond to informative training samples. The synthesized images are required to preserve information for training deep neural networks rather than visual reality or fidelity. Experiments verify that the deep neural networks can learn faster and achieve better performance when being trained with our IT-GAN generated images. We also show that our method is a promising solution to dataset condensation problem.

Yan Shu, Zheng Liu, Peitian Zhang et al.

Long video understanding poses a significant challenge for current Multi-modal Large Language Models (MLLMs). Notably, the MLLMs are constrained by their limited context lengths and the substantial costs while processing long videos. Although several existing methods attempt to reduce visual tokens, their strategies encounter severe bottleneck, restricting MLLMs' ability to perceive fine-grained visual details. In this work, we propose Video-XL, a novel approach that leverages MLLMs' inherent key-value (KV) sparsification capacity to condense the visual input. Specifically, we introduce a new special token, the Visual Summarization Token (VST), for each interval of the video, which summarizes the visual information within the interval as its associated KV. The VST module is trained by instruction fine-tuning, where two optimizing strategies are offered. 1.Curriculum learning, where VST learns to make small (easy) and large compression (hard) progressively. 2. Composite data curation, which integrates single-image, multi-image, and synthetic data to overcome the scarcity of long-video instruction data. The compression quality is further improved by dynamic compression, which customizes compression granularity based on the information density of different video intervals. Video-XL's effectiveness is verified from three aspects. First, it achieves a superior long-video understanding capability, outperforming state-of-the-art models of comparable sizes across multiple popular benchmarks. Second, it effectively preserves video information, with minimal compression loss even at 16x compression ratio. Third, it realizes outstanding cost-effectiveness, enabling high-quality processing of thousands of frames on a single A100 GPU.

Kai Wang, Bo Zhao, Xiangyu Peng et al.

Face recognition, as one of the most successful applications in artificial intelligence, has been widely used in security, administration, advertising, and healthcare. However, the privacy issues of public face datasets have attracted increasing attention in recent years. Previous works simply mask most areas of faces or synthesize samples using generative models to construct privacy-preserving face datasets, which overlooks the trade-off between privacy protection and data utility. In this paper, we propose a novel framework FaceMAE, where the face privacy and recognition performance are considered simultaneously. Firstly, randomly masked face images are used to train the reconstruction module in FaceMAE. We tailor the instance relation matching (IRM) module to minimize the distribution gap between real faces and FaceMAE reconstructed ones. During the deployment phase, we use trained FaceMAE to reconstruct images from masked faces of unseen identities without extra training. The risk of privacy leakage is measured based on face retrieval between reconstructed and original datasets. Experiments prove that the identities of reconstructed images are difficult to be retrieved. We also perform sufficient privacy-preserving face recognition on several public face datasets (i.e. CASIA-WebFace and WebFace260M). Compared to previous state of the arts, FaceMAE consistently \textbf{reduces at least 50\% error rate} on LFW, CFP-FP and AgeDB.

Hui Liu, Bo Zhao, Kehuan Zhang et al.

Although deep neural networks (DNNs) have shown impressive performance on many perceptual tasks, they are vulnerable to adversarial examples that are generated by adding slight but maliciously crafted perturbations to benign images. Adversarial detection is an important technique for identifying adversarial examples before they are entered into target DNNs. Previous studies to detect adversarial examples either targeted specific attacks or required expensive computation. How design a lightweight unsupervised detector is still a challenging problem. In this paper, we propose an AutoEncoder-based Adversarial Examples (AEAE) detector, that can guard DNN models by detecting adversarial examples with low computation in an unsupervised manner. The AEAE includes only a shallow autoencoder but plays two roles. First, a well-trained autoencoder has learned the manifold of benign examples. This autoencoder can produce a large reconstruction error for adversarial images with large perturbations, so we can detect significantly perturbed adversarial examples based on the reconstruction error. Second, the autoencoder can filter out the small noise and change the DNN's prediction on adversarial examples with small perturbations. It helps to detect slightly perturbed adversarial examples based on the prediction distance. To cover these two cases, we utilize the reconstruction error and prediction distance from benign images to construct a two-tuple feature set and train an adversarial detector using the isolation forest algorithm. We show empirically that the AEAE is unsupervised and inexpensive against the most state-of-the-art attacks. Through the detection in these two cases, there is nowhere to hide adversarial examples.

Yu Wang, Xuelin Qian, Jingyang Huo et al.

We present a significant breakthrough in 3D shape generation by scaling it to unprecedented dimensions. Through the adaptation of the Auto-Regressive model and the utilization of large language models, we have developed a remarkable model with an astounding 3.6 billion trainable parameters, establishing it as the largest 3D shape generation model to date, named Argus-3D. Our approach addresses the limitations of existing methods by enhancing the quality and diversity of generated 3D shapes. To tackle the challenges of high-resolution 3D shape generation, our model incorporates tri-plane features as latent representations, effectively reducing computational complexity. Additionally, we introduce a discrete codebook for efficient quantization of these representations. Leveraging the power of transformers, we enable multi-modal conditional generation, facilitating the production of diverse and visually impressive 3D shapes. To train our expansive model, we leverage an ensemble of publicly-available 3D datasets, consisting of a comprehensive collection of approximately 900,000 objects from renowned repositories such as ModelNet40, ShapeNet, Pix3D, 3D-Future, and Objaverse. This diverse dataset empowers our model to learn from a wide range of object variations, bolstering its ability to generate high-quality and diverse 3D shapes. Extensive experimentation demonstrate the remarkable efficacy of our approach in significantly improving the visual quality of generated 3D shapes. By pushing the boundaries of 3D generation, introducing novel methods for latent representation learning, and harnessing the power of transformers for multi-modal conditional generation, our contributions pave the way for substantial advancements in the field. Our work unlocks new possibilities for applications in gaming, virtual reality, product design, and other domains that demand high-quality and diverse 3D objects.

Xiaolong Han, Zehong Wang, Bo Zhao et al.

Neural network weights are typically viewed as the end product of training, while most deep learning research focuses on data, features, and architectures. However, recent advances show that the set of all possible weight values (weight space) itself contains rich structure: pretrained models form organized distributions, exhibit symmetries, and can be embedded, compared, or even generated. Understanding such structures has tremendous impact on how neural networks are analyzed and compared, and on how knowledge is transferred across models, beyond individual training instances. This emerging research direction, which we refer to as Weight Space Learning (WSL), treats neural weights as a meaningful domain for analysis and modeling. This survey provides the first unified taxonomy of WSL. We categorize existing methods into three core dimensions: Weight Space Understanding (WSU), which studies the geometry and symmetries of weights; Weight Space Representation (WSR), which learns embeddings over model weights; and Weight Space Generation (WSG), which synthesizes new weights through hypernetworks or generative models. We further show how these developments enable practical applications, including model retrieval, continual and federated learning, neural architecture search, and data-free reconstruction. By consolidating fragmented progress under a coherent framework, this survey highlights weight space as a learnable, structured domain with growing impact across model analysis, transferring, and weight generation. We release an accompanying resource at https://github.com/Zehong-Wang/Awesome-Weight-Space-Learning.

Bo Zhao, Iordan Ganev, Robin Walters et al.

Empirical studies of the loss landscape of deep networks have revealed that many local minima are connected through low-loss valleys. Yet, little is known about the theoretical origin of such valleys. We present a general framework for finding continuous symmetries in the parameter space, which carve out low-loss valleys. Our framework uses equivariances of the activation functions and can be applied to different layer architectures. To generalize this framework to nonlinear neural networks, we introduce a novel set of nonlinear, data-dependent symmetries. These symmetries can transform a trained model such that it performs similarly on new samples, which allows ensemble building that improves robustness under certain adversarial attacks. We then show that conserved quantities associated with linear symmetries can be used to define coordinates along low-loss valleys. The conserved quantities help reveal that using common initialization methods, gradient flow only explores a small part of the global minimum. By relating conserved quantities to convergence rate and sharpness of the minimum, we provide insights on how initialization impacts convergence and generalizability.

Yanwen Zou, Zhaoye Zhou, Chenyang Shi et al.

We propose U-Arm, a low-cost and rapidly adaptable leader-follower teleoperation framework designed to interface with most of commercially available robotic arms. Our system supports teleoperation through three structurally distinct 3D-printed leader arms that share consistent control logic, enabling seamless compatibility with diverse commercial robot configurations. Compared with previous open-source leader-follower interfaces, we further optimized both the mechanical design and servo selection, achieving a bill of materials (BOM) cost of only \$50.5 for the 6-DoF leader arm and \$56.8 for the 7-DoF version. To enhance usability, we mitigate the common challenge in controlling redundant degrees of freedom by %engineering methods mechanical and control optimizations. Experimental results demonstrate that U-Arm achieves 39\% higher data collection efficiency and comparable task success rates across multiple manipulation scenarios compared with Joycon, another low-cost teleoperation interface. We have open-sourced all CAD models of three configs and also provided simulation support for validating teleoperation workflows. We also open-sourced real-world manipulation data collected with U-Arm. The project website is https://github.com/MINT-SJTU/LeRobot-Anything-U-Arm.

Bo Zhao, Nima Dehmamy, Robin Walters et al.

Existing gradient-based optimization methods update parameters locally, in a direction that minimizes the loss function. We study a different approach, symmetry teleportation, that allows parameters to travel a large distance on the loss level set, in order to improve the convergence speed in subsequent steps. Teleportation exploits symmetries in the loss landscape of optimization problems. We derive loss-invariant group actions for test functions in optimization and multi-layer neural networks, and prove a necessary condition for teleportation to improve convergence rate. We also show that our algorithm is closely related to second order methods. Experimentally, we show that teleportation improves the convergence speed of gradient descent and AdaGrad for several optimization problems including test functions, multi-layer regressions, and MNIST classification.

Chengcheng Guo, Bo Zhao, Yanbing Bai

Coreset selection, which aims to select a subset of the most informative training samples, is a long-standing learning problem that can benefit many downstream tasks such as data-efficient learning, continual learning, neural architecture search, active learning, etc. However, many existing coreset selection methods are not designed for deep learning, which may have high complexity and poor generalization performance. In addition, the recently proposed methods are evaluated on models, datasets, and settings of different complexities. To advance the research of coreset selection in deep learning, we contribute a comprehensive code library, namely DeepCore, and provide an empirical study on popular coreset selection methods on CIFAR10 and ImageNet datasets. Extensive experiments on CIFAR10 and ImageNet datasets verify that, although various methods have advantages in certain experiment settings, random selection is still a strong baseline.

Huanzhou Zhu, Bo Zhao, Gang Chen et al.

Reinforcement learning (RL) trains many agents, which is resource-intensive and must scale to large GPU clusters. Different RL training algorithms offer different opportunities for distributing and parallelising the computation. Yet, current distributed RL systems tie the definition of RL algorithms to their distributed execution: they hard-code particular distribution strategies and only accelerate specific parts of the computation (e.g. policy network updates) on GPU workers. Fundamentally, current systems lack abstractions that decouple RL algorithms from their execution. We describe MindSpore Reinforcement Learning (MSRL), a distributed RL training system that supports distribution policies that govern how RL training computation is parallelised and distributed on cluster resources, without requiring changes to the algorithm implementation. MSRL introduces the new abstraction of a fragmented dataflow graph, which maps Python functions from an RL algorithm's training loop to parallel computational fragments. Fragments are executed on different devices by translating them to low-level dataflow representations, e.g. computational graphs as supported by deep learning engines, CUDA implementations or multi-threaded CPU processes. We show that MSRL subsumes the distribution strategies of existing systems, while scaling RL training to 64 GPUs.

Xiao Ma, Guang Zheng, Chi Xu et al.

Characterizing urban environments with broad coverages and high precision is more important than ever for achieving the UN's Sustainable Development Goals (SDGs) as half of the world's populations are living in cities. Urban building height as a fundamental 3D urban structural feature has far-reaching applications. However, so far, producing readily available datasets of recent urban building heights with fine spatial resolutions and global coverages remains a challenging task. Here, we provide an up-to-date global product of urban building heights based on a fine grid size of 150 m around 2020 by combining the spaceborne lidar instrument of GEDI and multi-sourced data including remotely sensed images (i.e., Landsat-8, Sentinel-2, and Sentinel-1) and topographic data. Our results revealed that the estimated method of building height samples based on the GEDI data was effective with 0.78 of Pearson's r and 3.67 m of RMSE in comparison to the reference data. The mapping product also demonstrated good performance as indicated by its strong correlation with the reference data (i.e., Pearson's r = 0.71, RMSE = 4.60 m). Compared with the currently existing products, our global urban building height map holds the ability to provide a higher spatial resolution (i.e., 150 m) with a great level of inherent details about the spatial heterogeneity and flexibility of updating using the GEDI samples as inputs. This work will boost future urban studies across many fields including climate, environmental, ecological, and social sciences.

Jianhao Yuan, Jie Zhang, Shuyang Sun et al.

Synthetic training data has gained prominence in numerous learning tasks and scenarios, offering advantages such as dataset augmentation, generalization evaluation, and privacy preservation. Despite these benefits, the efficiency of synthetic data generated by current methodologies remains inferior when training advanced deep models exclusively, limiting its practical utility. To address this challenge, we analyze the principles underlying training data synthesis for supervised learning and elucidate a principled theoretical framework from the distribution-matching perspective that explicates the mechanisms governing synthesis efficacy. Through extensive experiments, we demonstrate the effectiveness of our synthetic data across diverse image classification tasks, both as a replacement for and augmentation to real datasets, while also benefits such as out-of-distribution generalization, privacy preservation, and scalability. Specifically, we achieve 70.9% top1 classification accuracy on ImageNet1K when training solely with synthetic data equivalent to 1 X the original real data size, which increases to 76.0% when scaling up to 10 X synthetic data.

Hui Shi, Yupeng Gu, Yitong Zhou et al.

User embeddings (vectorized representations of a user) are essential in recommendation systems. Numerous approaches have been proposed to construct a representation for the user in order to find similar items for retrieval tasks, and they have been proven effective in industrial recommendation systems as well. Recently people have discovered the power of using multiple embeddings to represent a user, with the hope that each embedding represents the user's interest in a certain topic. With multi-interest representation, it's important to model the user's preference over the different topics and how the preference change with time. However, existing approaches either fail to estimate the user's affinity to each interest or unreasonably assume every interest of every user fades with an equal rate with time, thus hurting the recall of candidate retrieval. In this paper, we propose the Multi-Interest Preference (MIP) model, an approach that not only produces multi-interest for users by using the user's sequential engagement more effectively but also automatically learns a set of weights to represent the preference over each embedding so that the candidates can be retrieved from each interest proportionally. Extensive experiments have been done on various industrial-scale datasets to demonstrate the effectiveness of our approach.

Zeheng Wang, Zitong Yu, Yijie Zhu et al.

LLM-based multimodal emotion recognition relies on static parametric memory and often hallucinates when interpreting nuanced affective states. In this paper, given that single-round retrieval-augmented generation is highly susceptible to modal ambiguity and therefore struggles to capture complex affective dependencies across modalities, we introduce AffectAgent, an affect-oriented multi-agent retrieval-augmented generation framework that leverages collaborative decision-making among agents for fine-grained affective understanding. Specifically, AffectAgent comprises three jointly optimized specialized agents, namely a query planner, an evidence filter, and an emotion generator, which collaboratively perform analytical reasoning to retrieve cross-modal samples, assess evidence, and generate predictions. These agents are optimized end-to-end using Multi-Agent Proximal Policy Optimization (MAPPO) with a shared affective reward to ensure consistent emotion understanding. Furthermore, we introduce Modality-Balancing Mixture of Experts (MB-MoE) and Retrieval-Augmented Adaptive Fusion (RAAF), where MB-MoE dynamically regulates the contributions of different modalities to mitigate representation mismatch caused by cross-modal heterogeneity, while RAAF enhances semantic completion under missing-modality conditions by incorporating retrieved audiovisual embeddings. Extensive experiments on MER-UniBench demonstrate that AffectAgent achieves superior performance across complex scenarios. Our code will be released at: https://github.com/Wz1h1NG/AffectAgent.

Xuan Rao, Bo Zhao, Derong Liu

In neural architecture search (NAS) methods based on latent space optimization (LSO), a deep generative model is trained to embed discrete neural architectures into a continuous latent space. In this case, different optimization algorithms that operate in the continuous space can be implemented to search neural architectures. However, the optimization of latent variables is challenging for gradient-based LSO since the mapping from the latent space to the architecture performance is generally non-convex. To tackle this problem, this paper develops a convexity regularized latent space optimization (CR-LSO) method, which aims to regularize the learning process of latent space in order to obtain a convex architecture performance mapping. Specifically, CR-LSO trains a graph variational autoencoder (G-VAE) to learn the continuous representations of discrete architectures. Simultaneously, the learning process of latent space is regularized by the guaranteed convexity of input convex neural networks (ICNNs). In this way, the G-VAE is forced to learn a convex mapping from the architecture representation to the architecture performance. Hereafter, the CR-LSO approximates the performance mapping using the ICNN and leverages the estimated gradient to optimize neural architecture representations. Experimental results on three popular NAS benchmarks show that CR-LSO achieves competitive evaluation results in terms of both computational complexity and architecture performance.

Yikai Gu, Lele Cao, Bo Zhao et al.

Counterfactual explanations (CE) explain model decisions by identifying input modifications that lead to different predictions. Most existing methods operate at the instance level. Distributional Counterfactual Explanations (DCE) extend this setting by optimizing an optimal transport objective that balances proximity to a factual input distribution and alignment to a target output distribution, with statistical certification via chance constrained bounds. However, DCE relies on gradient based optimization, while many real-world tabular pipelines are dominated by non-differentiable models. We propose DISCOVER, a model-agnostic solver for distributional counterfactual explanations. DISCOVER preserves the original DCE objective and certification while replacing gradient descent with a sparse propose-and-select search paradigm. It exploits a sample-wise decomposition of the transport objective to compute per-row impact scores and enforce a top-$k$ intervention budget, focusing edits on the most influential samples. To guide candidate generation without predictor gradients, DISCOVER introduces an OT-guided cone sampling primitive driven by input-side transport geometry. Experiments on multiple tabular datasets demonstrate strong joint alignment of input and output distributions, extending distributional counterfactual reasoning to modern black box learning pipelines. A code repository is available at https://github.com/understanding-ml/DCE.

Bo Zhao, Berkcan Kapusuzoglu, Kartik Balasubramaniam et al.

Reasoning language models perform well on complex tasks but are costly to deploy due to their size and long reasoning traces. We propose a routing approach that assigns each problem to the smallest model likely to solve it, reducing compute without sacrificing accuracy. Using intermediate representations from s1.1-32B, we train lightweight predictors of problem difficulty or model correctness to guide routing across a pool of reasoning models. On diverse math benchmarks, routing improves efficiency over random assignment and matches s1.1-32B's performance while using significantly less compute. Our results demonstrate that difficulty-aware routing is effective for cost-efficient deployment of reasoning models.

Muyang He, Yexin Liu, Boya Wu et al.

Multimodal Large Language Models (MLLMs) have demonstrated notable capabilities in general visual understanding and reasoning tasks. However, their deployment is hindered by substantial computational costs in both training and inference, limiting accessibility to the broader research and user communities. A straightforward solution is to leverage smaller pre-trained vision and language models, which inevitably cause significant performance drops. In this paper, we demonstrate the possibility of training a smaller but better MLLM with high-quality training data. Specifically, we introduce Bunny, a family of lightweight MLLMs with flexible vision and language backbones for efficient multimodal learning from selected training data. Experiments show that our Bunny-4B/8B outperforms the state-of-the-art large MLLMs on multiple benchmarks. We expect that this work can provide the community with a clean and flexible open-source tool for further research and development. The code, models, and data can be found in https://github.com/BAAI-DCAI/Bunny.

Fan Bai, Yuxin Du, Tiejun Huang et al.

Medical image analysis is essential to clinical diagnosis and treatment, which is increasingly supported by multi-modal large language models (MLLMs). However, previous research has primarily focused on 2D medical images, leaving 3D images under-explored, despite their richer spatial information. This paper aims to advance 3D medical image analysis with MLLMs. To this end, we present a large-scale 3D multi-modal medical dataset, M3D-Data, comprising 120K image-text pairs and 662K instruction-response pairs specifically tailored for various 3D medical tasks, such as image-text retrieval, report generation, visual question answering, positioning, and segmentation. Additionally, we propose M3D-LaMed, a versatile multi-modal large language model for 3D medical image analysis. Furthermore, we introduce a new 3D multi-modal medical benchmark, M3D-Bench, which facilitates automatic evaluation across eight tasks. Through comprehensive evaluation, our method proves to be a robust model for 3D medical image analysis, outperforming existing solutions. All code, data, and models are publicly available at: https://github.com/BAAI-DCAI/M3D.

Runze Wang, Yuqian Fu, Yu Li et al.

Vision-language-action (VLA) models have shown strong potential for generalist robot manipulation, yet they remain limited by insufficient spatial reasoning, particularly in determining where to interact in complex visual scenes. While recent efforts introduce various forms of visual planning to address this issue, existing approaches either rely on global geometric cues, symbolic intermediate representations, or externally generated visual signals, which are often weakly coupled with downstream action prediction. In this work, we revisit visual planning in VLA systems and argue that effective planning should be local, visually grounded, internally generated, and directly aligned with action. Based on this insight, we propose Afford-VLA, a unified framework that internalizes task-conditioned affordance as an explicit visual planning interface within VLA models. Concretely, we introduce learnable <AFF> tokens to query task-relevant interaction regions, decode affordance masks from multimodal features, and convert them into compact embeddings that directly condition action generation. This design enables affordance to be both generated and utilized within the VLA, forming a tightly coupled perception-action pathway. To further support this integration, we adopt a training strategy that allows the affordance pathway to be jointly optimized with action prediction, improving its effectiveness for downstream control. We evaluate our method on multiple simulation benchmarks, including LIBERO, LIBERO-Plus, and SimplerEnv, achieving consistent state-of-the-art performance, along with strong real-world results. These findings demonstrate that internalizing affordance as action-aligned visual planning provides a powerful paradigm for improving VLA systems.

Yizhang Jin, Jian Li, Yexin Liu et al.

In the past year, Multimodal Large Language Models (MLLMs) have demonstrated remarkable performance in tasks such as visual question answering, visual understanding and reasoning. However, the extensive model size and high training and inference costs have hindered the widespread application of MLLMs in academia and industry. Thus, studying efficient and lightweight MLLMs has enormous potential, especially in edge computing scenarios. In this survey, we provide a comprehensive and systematic review of the current state of efficient MLLMs. Specifically, we summarize the timeline of representative efficient MLLMs, research state of efficient structures and strategies, and the applications. Finally, we discuss the limitations of current efficient MLLM research and promising future directions. Please refer to our GitHub repository for more details: https://github.com/lijiannuist/Efficient-Multimodal-LLMs-Survey.

Bo Zhao, Yihang Liu, Chenfeng Zhang et al.

Text-guided texture editing aims to modify object appearance while preserving the underlying geometric structure. However, our empirical analysis reveals that even SOTA editing models frequently struggle to maintain structural consistency during texture editing, despite the intended changes being purely appearance-related. Motivated by this observation, we jointly enhance structure preservation from both data and training perspectives, and build TexEditor, a dedicated texture editing model based on Qwen-Image-Edit-2509. Firstly, we construct TexBlender, a high-quality SFT dataset generated with Blender, which provides strong structural priors for a cold start. Sec- ondly, we introduce StructureNFT, a RL-based approach that integrates structure-preserving losses to transfer the structural priors learned during SFT to real-world scenes. Moreover, due to the limited realism and evaluation coverage of existing benchmarks, we introduce TexBench, a general-purpose real-world benchmark for text-guided texture editing. Extensive experiments on existing Blender-based texture benchmarks and our TexBench show that TexEditor consistently outperforms strong baselines such as Nano Banana Pro. In addition, we assess TexEditor on the general purpose benchmark ImgEdit to validate its generalization. Our code and data are available at https://github.com/KlingAIResearch/TexEditor.

Qi Cai, Jingwen Chen, Yang Chen et al.

Recent advancements in image generative foundation models have prioritized quality improvements but often at the cost of increased computational complexity and inference latency. To address this critical trade-off, we introduce HiDream-I1, a new open-source image generative foundation model with 17B parameters that achieves state-of-the-art image generation quality within seconds. HiDream-I1 is constructed with a new sparse Diffusion Transformer (DiT) structure. Specifically, it starts with a dual-stream decoupled design of sparse DiT with dynamic Mixture-of-Experts (MoE) architecture, in which two separate encoders are first involved to independently process image and text tokens. Then, a single-stream sparse DiT structure with dynamic MoE architecture is adopted to trigger multi-model interaction for image generation in a cost-efficient manner. To support flexiable accessibility with varied model capabilities, we provide HiDream-I1 in three variants: HiDream-I1-Full, HiDream-I1-Dev, and HiDream-I1-Fast. Furthermore, we go beyond the typical text-to-image generation and remould HiDream-I1 with additional image conditions to perform precise, instruction-based editing on given images, yielding a new instruction-based image editing model namely HiDream-E1. Ultimately, by integrating text-to-image generation and instruction-based image editing, HiDream-I1 evolves to form a comprehensive image agent (HiDream-A1) capable of fully interactive image creation and refinement. To accelerate multi-modal AIGC research, we have open-sourced all the codes and model weights of HiDream-I1-Full, HiDream-I1-Dev, HiDream-I1-Fast, HiDream-E1 through our project websites: https://github.com/HiDream-ai/HiDream-I1 and https://github.com/HiDream-ai/HiDream-E1. All features can be directly experienced via https://vivago.ai/studio.

Junjie Zhou, Zheng Liu, Ze Liu et al.

Despite the rapidly growing demand for multimodal retrieval, progress in this field remains severely constrained by a lack of training data. In this paper, we introduce MegaPairs, a novel data synthesis method that leverages vision language models (VLMs) and open-domain images, together with a massive synthetic dataset generated from this method. Our empirical analysis shows that MegaPairs generates high-quality data, enabling the multimodal retriever to significantly outperform the baseline model trained on 70$\times$ more data from existing datasets. Moreover, since MegaPairs solely relies on general image corpora and open-source VLMs, it can be easily scaled up, enabling continuous improvements in retrieval performance. In this stage, we produced more than 26 million training instances and trained several models of varying sizes using this data. These new models achieve state-of-the-art zero-shot performance across 4 popular composed image retrieval (CIR) benchmarks and the highest overall performance on the 36 datasets provided by MMEB. They also demonstrate notable performance improvements with additional downstream fine-tuning. Our produced dataset, well-trained models, and data synthesis pipeline will be made publicly available to facilitate the future development of this field.

Dingxin Cheng, Mingda Li, Jingyu Liu et al.

Recently, integrating visual foundation models into large language models (LLMs) to form video understanding systems has attracted widespread attention. Most of the existing models compress diverse semantic information within the whole video and feed it into LLMs for content comprehension. While this method excels in short video understanding, it may result in a blend of multiple event information in long videos due to coarse compression, which causes information redundancy. Consequently, the semantics of key events might be obscured within the vast information that hinders the model's understanding capabilities. To address this issue, we propose a Hierarchical Event-based Memory-enhanced LLM (HEM-LLM) for better understanding of long videos. Firstly, we design a novel adaptive sequence segmentation scheme to divide multiple events within long videos. In this way, we can perform individual memory modeling for each event to establish intra-event contextual connections, thereby reducing information redundancy. Secondly, while modeling current event, we compress and inject the information of the previous event to enhance the long-term inter-event dependencies in videos. Finally, we perform extensive experiments on various video understanding tasks and the results show that our model achieves state-of-the-art performances.

Pedro R. A. S. Bassi, Wenxuan Li, Yucheng Tang et al.

How can we test AI performance? This question seems trivial, but it isn't. Standard benchmarks often have problems such as in-distribution and small-size test sets, oversimplified metrics, unfair comparisons, and short-term outcome pressure. As a consequence, good performance on standard benchmarks does not guarantee success in real-world scenarios. To address these problems, we present Touchstone, a large-scale collaborative segmentation benchmark of 9 types of abdominal organs. This benchmark is based on 5,195 training CT scans from 76 hospitals around the world and 5,903 testing CT scans from 11 additional hospitals. This diverse test set enhances the statistical significance of benchmark results and rigorously evaluates AI algorithms across various out-of-distribution scenarios. We invited 14 inventors of 19 AI algorithms to train their algorithms, while our team, as a third party, independently evaluated these algorithms on three test sets. In addition, we also evaluated pre-existing AI frameworks--which, differing from algorithms, are more flexible and can support different algorithms--including MONAI from NVIDIA, nnU-Net from DKFZ, and numerous other open-source frameworks. We are committed to expanding this benchmark to encourage more innovation of AI algorithms for the medical domain.

Bo Zhao, Fanghua Ye, Yixin Ji et al.

Humans infer emotions by integrating observed multimodal cues with expectations about how affective states may unfold. Existing multimodal large language models (MLLMs), however, often treat emotion recognition as static fusion over complete audiovisual-text inputs, leaving affective dynamics implicit. We propose AffectVerse, a Qwen2.5-Omni-based model equipped with an Emotion World Module (EWM), an action-free representation-level module for short-horizon latent affective prediction. \rev{EWM contains three modules: 1) Cross-Modal Temporal Imagination predicts future video/audio representations from past tokens with multi-step rollout. 2) MAMA(Modality-Aware Multi-step Attention) Belief Aggregation compresses imagined tokens into modality-aware belief tokens. 3) Belief Injection inserts these belief tokens into the LLM for affective reasoning.} AffectVerse uses future prediction as a past-conditioned self-supervised signal: it does not replace modeling observed history or require unseen signals at inference, but forces the current belief state to encode transition cues that are predictive of subsequent affective change. Across nine benchmarks, AffectVerse improves at least 2.57\% over other models, while controlled ablations show additive gains from temporal imagination, cross-modal rollout, and belief aggregation. These results suggest predictive belief-state modeling is a practical alternative for affective computing.

Luca Zhou, Bo Zhao, Rose Yu et al.

Model merging combines knowledge from separately fine-tuned models, yet success factors remain poorly understood. While recent work treats mergeability as an intrinsic property, we show with an architecture-agnostic framework that it fundamentally depends on both the merging method and the partner tasks. Using linear optimization over a set of interpretable pairwise metrics (e.g., gradient L2 distance), we uncover properties correlating with post-merge performance across four merging methods. We find substantial variation in success drivers (46.7% metric overlap; 55.3% sign agreement), revealing method-specific "fingerprints". Crucially, however, subspace overlap and gradient alignment metrics consistently emerge as foundational, method-agnostic prerequisites for compatibility. These findings provide a diagnostic foundation for understanding mergeability and motivate future fine-tuning strategies that explicitly encourage these properties.

Tao Lin, Yilei Zhong, Yuxin Du et al.

Vision-Language-Action (VLA) models have emerged as a powerful framework that unifies perception, language, and control, enabling robots to perform diverse tasks through multimodal understanding. However, current VLA models typically contain massive parameters and rely heavily on large-scale robot data pretraining, leading to high computational costs during training, as well as limited deployability for real-time inference. Moreover, most training paradigms often degrade the perceptual representations of the vision-language backbone, resulting in overfitting and poor generalization to downstream tasks. In this work, we present Evo-1, a lightweight VLA model that reduces computation and improves deployment efficiency, while maintaining strong performance without pretraining on robot data. Evo-1 builds on a native multimodal Vision-Language model (VLM), incorporating a novel cross-modulated diffusion transformer along with an optimized integration module, together forming an effective architecture. We further introduce a two-stage training paradigm that progressively aligns action with perception, preserving the representations of the VLM. Notably, with only 0.77 billion parameters, Evo-1 achieves state-of-the-art results on the Meta-World and RoboTwin suite, surpassing the previous best models by 12.4% and 6.9%, respectively, and also attains a competitive result of 94.8% on LIBERO. In real-world evaluations, Evo-1 attains a 78% success rate with high inference frequency and low memory overhead, outperforming all baseline methods. We release code, data, and model weights to facilitate future research on lightweight and efficient VLA models.

Mingze Gao, Jingyu Liu, Mingda Li et al.

Multimodal Large Language Models (MLLMs) have significantly improved performance across various image-language applications. Recently, there has been a growing interest in adapting image pre-trained MLLMs for video-related tasks. However, most efforts concentrate on enhancing the vision encoder and projector components, while the core part, Large Language Models (LLMs), remains comparatively under-explored. In this paper, we propose two strategies to enhance the model's capability in video understanding tasks by improving inter-layer attention computation in LLMs. Specifically, the first approach focuses on the enhancement of Rotary Position Embedding (RoPE) with Temporal-Aware Dual RoPE, which introduces temporal position information to strengthen the MLLM's temporal modeling capabilities while preserving the relative position relationships of both visual and text tokens. The second approach involves enhancing the Attention Mask with the Frame-wise Block Causal Attention Mask, a simple yet effective method that broadens visual token interactions within and across video frames while maintaining the causal inference mechanism. Based on these proposed methods, we adapt LLaVA for video understanding tasks, naming it Temporal-Considered LLaVA (TC-LLaVA). Our TC-LLaVA achieves new state-of-the-art performance across various video understanding benchmarks with only supervised fine-tuning (SFT) on video-related datasets.

Bo Zhao, Maosheng Pang, Chen Zhang et al.

Existing works typically focus on presentation generation under isolated input settings, whereas real-world use cases span diverse scenarios, including vague user prompts, long documents, multimodal materials, and multiple heterogeneous sources. Moreover, current evaluations are often insufficiently scenario-specific. They mainly rely on generic presentation-quality criteria, such as visual appeal, layout quality, and overall coherence, but fail to assess the core capabilities required by different input settings, including grounded compression, visual-text alignment, and cross-source synthesis. Consequently, the field lacks a unified benchmark and a scenario-aware evaluation framework for faithfully diagnosing presentation-generation systems across diverse real-world settings. We present UniPPTBench, a unified benchmark for presentation generation across four representative input settings: vague-prompt, long-document, multimodal-document, and multi-source generation. We further introduce UniPPTEval, a scenario-aware evaluation protocol that combines shared metrics for cross-setting comparison with scenario-specific metrics tailored to the core requirements of each setting. We also provide transparent reference baselines to support reproducible comparison. Experiments on UniPPTBench reveal substantial performance variation across settings and recurring failure modes in content grounding, multimodal integration, and cross-source synthesis. In particular, strong performance on generic presentation-quality metrics does not necessarily imply strong task fulfillment in grounded scenarios. Together, UniPPTBench and UniPPTEval provide a faithful and diagnostic foundation for evaluating presentation generation across diverse real-world scenarios. Code and data will be publicly available.

Minghao Qin, Xiangrui Liu, Zhengyang Liang et al.

Multi-modal large language models (MLLMs) models have made significant progress in video understanding over the past few years. However, processing long video inputs remains a major challenge due to high memory and computational costs. This makes it difficult for current models to achieve both strong performance and high efficiency in long video understanding. To address this challenge, we propose Video-XL-2, a novel MLLM that delivers superior cost-effectiveness for long-video understanding based on task-aware KV sparsification. The proposed framework operates with two key steps: chunk-based pre-filling and bi-level key-value decoding. Chunk-based pre-filling divides the visual token sequence into chunks, applying full attention within each chunk and sparse attention across chunks. This significantly reduces computational and memory overhead. During decoding, bi-level key-value decoding selectively reloads either dense or sparse key-values for each chunk based on its relevance to the task. This approach further improves memory efficiency and enhances the model's ability to capture fine-grained information. Video-XL-2 achieves state-of-the-art performance on various long video understanding benchmarks, outperforming existing open-source lightweight models. It also demonstrates exceptional efficiency, capable of processing over 10,000 frames on a single NVIDIA A100 (80GB) GPU and thousands of frames in just a few seconds.

Xiang Li, Yiqun Yao, Xin Jiang et al.

Large language models (LLMs) have showcased profound capabilities in language understanding and generation, facilitating a wide array of applications. However, there is a notable paucity of detailed, open-sourced methodologies on efficiently scaling LLMs beyond 50 billion parameters with minimum trial-and-error cost and computational resources. In this report, we introduce Tele-FLM (aka FLM-2), a 52B open-sourced multilingual large language model that features a stable, efficient pre-training paradigm and enhanced factual judgment capabilities. Tele-FLM demonstrates superior multilingual language modeling abilities, measured by BPB on textual corpus. Besides, in both English and Chinese foundation model evaluation, it is comparable to strong open-sourced models that involve larger pre-training FLOPs, such as Llama2-70B and DeepSeek-67B. In addition to the model weights, we share the core designs, engineering practices, and training details, which we expect to benefit both the academic and industrial communities.

Bo Zhao, Kairui Guo, Runnan Du et al.

Instruction guided image editing has advanced substantially with recent generative models, yet it still fails to produce reliable results across many seemingly simple cases. We observe that a large portion of these failures stem not from insufficient model capacity, but from poorly formulated editing tasks, such as those involving small targets, implicit spatial relations, or under-specified instructions. In this work, we frame image editing failures as a task formulation problem and propose an adaptive task reformulation framework that improves editing performance without modifying the underlying model. Our key idea is to transform the original image-instruction pair into a sequence of operations that are dynamically determined and executed by a MLLM agent through analysis, routing, reformulation, and feedback-driven refinement. Experiments on multiple benchmarks, including ImgEdit, PICA, and RePlan, across diverse editing backbones such as Qwen Image Edit and Nano Banana, show consistent improvements, with especially large gains on challenging cases. These results suggest that task reformulation is a critical but underexplored factor, and that substantial gains can be achieved by better matching editing tasks to the effective operating regime of existing models.

Bach Do, Sina Jafari Ghalekohneh, Taiwo Adebiyi et al.

Nonreciprocal thermal emitters that break Kirchhoff's law of thermal radiation promise exciting applications for thermal and energy applications. The design of the bandwidth and angular range of the nonreciprocal effect, which directly affects the performance of nonreciprocal emitters, typically relies on physical intuition. In this study, we present a general numerical approach to maximize the nonreciprocal effect. We choose doped magneto-optic materials and magnetic Weyl semimetal materials as model materials and focus on pattern-free multilayer structures. The optimization randomly starts from a less effective structure and incrementally improves the broadband nonreciprocity through the combination of Bayesian optimization and reparameterization. Optimization results show that the proposed approach can discover structures that can achieve broadband nonreciprocal emission at wavelengths from 5 to 40 micrometers using only a fewer layers, significantly outperforming current state-of-the-art designs based on intuition in terms of both performance and simplicity.

Kai Wang, Bo Zhao, Xiangyu Peng et al.

Dataset condensation aims at reducing the network training effort through condensing a cumbersome training set into a compact synthetic one. State-of-the-art approaches largely rely on learning the synthetic data by matching the gradients between the real and synthetic data batches. Despite the intuitive motivation and promising results, such gradient-based methods, by nature, easily overfit to a biased set of samples that produce dominant gradients, and thus lack global supervision of data distribution. In this paper, we propose a novel scheme to Condense dataset by Aligning FEatures (CAFE), which explicitly attempts to preserve the real-feature distribution as well as the discriminant power of the resulting synthetic set, lending itself to strong generalization capability to various architectures. At the heart of our approach is an effective strategy to align features from the real and synthetic data across various scales, while accounting for the classification of real samples. Our scheme is further backed up by a novel dynamic bi-level optimization, which adaptively adjusts parameter updates to prevent over-/under-fitting. We validate the proposed CAFE across various datasets, and demonstrate that it generally outperforms the state of the art: on the SVHN dataset, for example, the performance gain is up to 11%. Extensive experiments and analyses verify the effectiveness and necessity of proposed designs.