Siyuan Qiao

Qihang Yu, Huiyu Wang, Siyuan Qiao et al. · deepmind

The rise of transformers in vision tasks not only advances network backbone designs, but also starts a brand-new page to achieve end-to-end image recognition (e.g., object detection and panoptic segmentation). Originated from Natural Language Processing (NLP), transformer architectures, consisting of self-attention and cross-attention, effectively learn long-range interactions between elements in a sequence. However, we observe that most existing transformer-based vision models simply borrow the idea from NLP, neglecting the crucial difference between languages and images, particularly the extremely large sequence length of spatially flattened pixel features. This subsequently impedes the learning in cross-attention between pixel features and object queries. In this paper, we rethink the relationship between pixels and object queries and propose to reformulate the cross-attention learning as a clustering process. Inspired by the traditional k-means clustering algorithm, we develop a k-means Mask Xformer (kMaX-DeepLab) for segmentation tasks, which not only improves the state-of-the-art, but also enjoys a simple and elegant design. As a result, our kMaX-DeepLab achieves a new state-of-the-art performance on COCO val set with 58.0% PQ, Cityscapes val set with 68.4% PQ, 44.0% AP, and 83.5% mIoU, and ADE20K val set with 50.9% PQ and 55.2% mIoU without test-time augmentation or external dataset. We hope our work can shed some light on designing transformers tailored for vision tasks. TensorFlow code and models are available at https://github.com/google-research/deeplab2 A PyTorch re-implementation is also available at https://github.com/bytedance/kmax-deeplab

Qihang Yu, Huiyu Wang, Dahun Kim et al. · deepmind

We propose Clustering Mask Transformer (CMT-DeepLab), a transformer-based framework for panoptic segmentation designed around clustering. It rethinks the existing transformer architectures used in segmentation and detection; CMT-DeepLab considers the object queries as cluster centers, which fill the role of grouping the pixels when applied to segmentation. The clustering is computed with an alternating procedure, by first assigning pixels to the clusters by their feature affinity, and then updating the cluster centers and pixel features. Together, these operations comprise the Clustering Mask Transformer (CMT) layer, which produces cross-attention that is denser and more consistent with the final segmentation task. CMT-DeepLab improves the performance over prior art significantly by 4.4% PQ, achieving a new state-of-the-art of 55.7% PQ on the COCO test-dev set.

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

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

Dahun Kim, Jun Xie, Huiyu Wang et al. · deepmind

We present TubeFormer-DeepLab, the first attempt to tackle multiple core video segmentation tasks in a unified manner. Different video segmentation tasks (e.g., video semantic/instance/panoptic segmentation) are usually considered as distinct problems. State-of-the-art models adopted in the separate communities have diverged, and radically different approaches dominate in each task. By contrast, we make a crucial observation that video segmentation tasks could be generally formulated as the problem of assigning different predicted labels to video tubes (where a tube is obtained by linking segmentation masks along the time axis) and the labels may encode different values depending on the target task. The observation motivates us to develop TubeFormer-DeepLab, a simple and effective video mask transformer model that is widely applicable to multiple video segmentation tasks. TubeFormer-DeepLab directly predicts video tubes with task-specific labels (either pure semantic categories, or both semantic categories and instance identities), which not only significantly simplifies video segmentation models, but also advances state-of-the-art results on multiple video segmentation benchmarks

Xuan Yang, Liangzhe Yuan, Kimberly Wilber et al. · deepmind

Dense prediction tasks, such as semantic segmentation, depth estimation, and surface normal prediction, can be easily formulated as per-pixel classification (discrete outputs) or regression (continuous outputs). This per-pixel prediction paradigm has remained popular due to the prevalence of fully convolutional networks. However, on the recent frontier of segmentation task, the community has been witnessing a shift of paradigm from per-pixel prediction to cluster-prediction with the emergence of transformer architectures, particularly the mask transformers, which directly predicts a label for a mask instead of a pixel. Despite this shift, methods based on the per-pixel prediction paradigm still dominate the benchmarks on the other dense prediction tasks that require continuous outputs, such as depth estimation and surface normal prediction. Motivated by the success of DORN and AdaBins in depth estimation, achieved by discretizing the continuous output space, we propose to generalize the cluster-prediction based method to general dense prediction tasks. This allows us to unify dense prediction tasks with the mask transformer framework. Remarkably, the resulting model PolyMaX demonstrates state-of-the-art performance on three benchmarks of NYUD-v2 dataset. We hope our simple yet effective design can inspire more research on exploiting mask transformers for more dense prediction tasks. Code and model will be made available.

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

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

Jieru Mei, Alex Zihao Zhu, Xinchen Yan et al.

Panoptic image segmentation is the computer vision task of finding groups of pixels in an image and assigning semantic classes and object instance identifiers to them. Research in image segmentation has become increasingly popular due to its critical applications in robotics and autonomous driving. The research community thereby relies on publicly available benchmark dataset to advance the state-of-the-art in computer vision. Due to the high costs of densely labeling the images, however, there is a shortage of publicly available ground truth labels that are suitable for panoptic segmentation. The high labeling costs also make it challenging to extend existing datasets to the video domain and to multi-camera setups. We therefore present the Waymo Open Dataset: Panoramic Video Panoptic Segmentation Dataset, a large-scale dataset that offers high-quality panoptic segmentation labels for autonomous driving. We generate our dataset using the publicly available Waymo Open Dataset, leveraging the diverse set of camera images. Our labels are consistent over time for video processing and consistent across multiple cameras mounted on the vehicles for full panoramic scene understanding. Specifically, we offer labels for 28 semantic categories and 2,860 temporal sequences that were captured by five cameras mounted on autonomous vehicles driving in three different geographical locations, leading to a total of 100k labeled camera images. To the best of our knowledge, this makes our dataset an order of magnitude larger than existing datasets that offer video panoptic segmentation labels. We further propose a new benchmark for Panoramic Video Panoptic Segmentation and establish a number of strong baselines based on the DeepLab family of models. We will make the benchmark and the code publicly available. Find the dataset at https://waymo.com/open.

Chen Wei, Chenxi Liu, Siyuan Qiao et al.

We demonstrate text as a strong cross-modal interface. Rather than relying on deep embeddings to connect image and language as the interface representation, our approach represents an image as text, from which we enjoy the interpretability and flexibility inherent to natural language. We employ an autoencoder that uses a pre-trained text-to-image diffusion model for decoding. The encoder is trained to transform an input image into text, which is then fed into the fixed text-to-image diffusion decoder to reconstruct the original input -- a process we term De-Diffusion. Experiments validate both the precision and comprehensiveness of De-Diffusion text representing images, such that it can be readily ingested by off-the-shelf text-to-image tools and LLMs for diverse multi-modal tasks. For example, a single De-Diffusion model can generalize to provide transferable prompts for different text-to-image tools, and also achieves a new state of the art on open-ended vision-language tasks by simply prompting large language models with few-shot examples.

Alex Zihao Zhu, Jieru Mei, Siyuan Qiao et al.

Semantic segmentation, which aims to classify every pixel in an image, is a key task in machine perception, with many applications across robotics and autonomous driving. Due to the high dimensionality of this task, most existing approaches use local operations, such as convolutions, to generate per-pixel features. However, these methods are typically unable to effectively leverage global context information due to the high computational costs of operating on a dense image. In this work, we propose a solution to this issue by leveraging the idea of superpixels, an over-segmentation of the image, and applying them with a modern transformer framework. In particular, our model learns to decompose the pixel space into a spatially low dimensional superpixel space via a series of local cross-attentions. We then apply multi-head self-attention to the superpixels to enrich the superpixel features with global context and then directly produce a class prediction for each superpixel. Finally, we directly project the superpixel class predictions back into the pixel space using the associations between the superpixels and the image pixel features. Reasoning in the superpixel space allows our method to be substantially more computationally efficient compared to convolution-based decoder methods. Yet, our method achieves state-of-the-art performance in semantic segmentation due to the rich superpixel features generated by the global self-attention mechanism. Our experiments on Cityscapes and ADE20K demonstrate that our method matches the state of the art in terms of accuracy, while outperforming in terms of model parameters and latency.

Seed, Baisheng Li, Banggu Wu et al.

We introduce Virtual Width Networks (VWN), a framework that delivers the benefits of wider representations without incurring the quadratic cost of increasing the hidden size. VWN decouples representational width from backbone width, expanding the embedding space while keeping backbone compute nearly constant. In our large-scale experiment, an 8-times expansion accelerates optimization by over 2 times for next-token and 3 times for next-2-token prediction. The advantage amplifies over training as both the loss gap grows and the convergence-speedup ratio increases, showing that VWN is not only token-efficient but also increasingly effective with scale. Moreover, we identify an approximately log-linear scaling relation between virtual width and loss reduction, offering an initial empirical basis and motivation for exploring virtual-width scaling as a new dimension of large-model efficiency.

Yunshui Li, Yiyuan Ma, Shen Yan et al.

Model merging has emerged as a promising technique for enhancing large language models, though its application in large-scale pre-training remains relatively unexplored. In this paper, we present a comprehensive investigation of model merging techniques during the pre-training process. Through extensive experiments with both dense and Mixture-of-Experts (MoE) architectures ranging from millions to over 100 billion parameters, we demonstrate that merging checkpoints trained with constant learning rates not only achieves significant performance improvements but also enables accurate prediction of annealing behavior. These improvements lead to both more efficient model development and significantly lower training costs. Our detailed ablation studies on merging strategies and hyperparameters provide new insights into the underlying mechanisms while uncovering novel applications. Through comprehensive experimental analysis, we offer the open-source community practical pre-training guidelines for effective model merging.

Gemini Team, Petko Georgiev, Ving Ian Lei et al. · deepmind, mila

In this report, we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. The family includes two new models: (1) an updated Gemini 1.5 Pro, which exceeds the February version on the great majority of capabilities and benchmarks; (2) Gemini 1.5 Flash, a more lightweight variant designed for efficiency with minimal regression in quality. Gemini 1.5 models achieve near-perfect recall on long-context retrieval tasks across modalities, improve the state-of-the-art in long-document QA, long-video QA and long-context ASR, and match or surpass Gemini 1.0 Ultra's state-of-the-art performance across a broad set of benchmarks. Studying the limits of Gemini 1.5's long-context ability, we find continued improvement in next-token prediction and near-perfect retrieval (>99%) up to at least 10M tokens, a generational leap over existing models such as Claude 3.0 (200k) and GPT-4 Turbo (128k). Finally, we highlight real-world use cases, such as Gemini 1.5 collaborating with professionals on completing their tasks achieving 26 to 75% time savings across 10 different job categories, as well as surprising new capabilities of large language models at the frontier; when given a grammar manual for Kalamang, a language with fewer than 200 speakers worldwide, the model learns to translate English to Kalamang at a similar level to a person who learned from the same content.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Ang Lv, Jin Ma, Yiyuan Ma et al.

Mixture-of-Experts (MoE) models lack explicit constraints to ensure the router's decisions align well with the experts' capabilities, which ultimately limits model performance. To address this, we propose expert-router coupling (ERC) loss, a lightweight auxiliary loss that tightly couples the router's decisions with expert capabilities. Our approach treats each expert's router embedding as a proxy token for the tokens assigned to that expert, and feeds perturbed router embeddings through the experts to obtain intermediate activations. The ERC loss enforces two constraints on these activations: (1) Each expert must exhibit higher activation for its own proxy token than for the proxy tokens of any other expert. (2) Each proxy token must elicit stronger activation from its corresponding expert than from any other expert. These constraints jointly ensure that each router embedding faithfully represents its corresponding expert's capability, while each expert specializes in processing the tokens actually routed to it. The ERC loss is computationally efficient, operating only on $n^2$ activations, where $n$ is the number of experts. This represents a fixed cost independent of batch size, unlike prior coupling methods that scale with the number of tokens (often millions per batch). Through pre-training MoE-LLMs ranging from 3B to 15B parameters and extensive analysis on trillions of tokens, we demonstrate the effectiveness of the ERC loss. Moreover, the ERC loss offers flexible control and quantitative tracking of expert specialization levels during training, providing valuable insights into MoEs.

Mark Weber, Huiyu Wang, Siyuan Qiao et al.

DeepLab2 is a TensorFlow library for deep labeling, aiming to provide a state-of-the-art and easy-to-use TensorFlow codebase for general dense pixel prediction problems in computer vision. DeepLab2 includes all our recently developed DeepLab model variants with pretrained checkpoints as well as model training and evaluation code, allowing the community to reproduce and further improve upon the state-of-art systems. To showcase the effectiveness of DeepLab2, our Panoptic-DeepLab employing Axial-SWideRNet as network backbone achieves 68.0% PQ or 83.5% mIoU on Cityscaspes validation set, with only single-scale inference and ImageNet-1K pretrained checkpoints. We hope that publicly sharing our library could facilitate future research on dense pixel labeling tasks and envision new applications of this technology. Code is made publicly available at \url{https://github.com/google-research/deeplab2}.

Yuhui Xu, Lingxi Xie, Cihang Xie et al.

Batch normalization (BN) is a fundamental unit in modern deep networks, in which a linear transformation module was designed for improving BN's flexibility of fitting complex data distributions. In this paper, we demonstrate properly enhancing this linear transformation module can effectively improve the ability of BN. Specifically, rather than using a single neuron, we propose to additionally consider each neuron's neighborhood for calculating the outputs of the linear transformation. Our method, named BNET, can be implemented with 2-3 lines of code in most deep learning libraries. Despite the simplicity, BNET brings consistent performance gains over a wide range of backbones and visual benchmarks. Moreover, we verify that BNET accelerates the convergence of network training and enhances spatial information by assigning the important neurons with larger weights accordingly. The code is available at https://github.com/yuhuixu1993/BNET.

Hao Ding, Siyuan Qiao, Alan Yuille et al.

The key to a successful cascade architecture for precise instance segmentation is to fully leverage the relationship between bounding box detection and mask segmentation across multiple stages. Although modern instance segmentation cascades achieve leading performance, they mainly make use of a unidirectional relationship, i.e., mask segmentation can benefit from iteratively refined bounding box detection. In this paper, we investigate an alternative direction, i.e., how to take the advantage of precise mask segmentation for bounding box detection in a cascade architecture. We propose a Deeply Shape-guided Cascade (DSC) for instance segmentation, which iteratively imposes the shape guidances extracted from mask prediction at the previous stage on bounding box detection at current stage. It forms a bi-directional relationship between the two tasks by introducing three key components: (1) Initial shape guidance: A mask-supervised Region Proposal Network (mPRN) with the ability to generate class-agnostic masks; (2) Explicit shape guidance: A mask-guided region-of-interest (RoI) feature extractor, which employs mask segmentation at previous stage to focus feature extraction at current stage within a region aligned well with the shape of the instance-of-interest rather than a rectangular RoI; (3) Implicit shape guidance: A feature fusion operation which feeds intermediate mask features at previous stage to the bounding box head at current stage. Experimental results show that DSC outperforms the state-of-the-art instance segmentation cascade, Hybrid Task Cascade (HTC), by a large margin and achieves 51.8 box AP and 45.5 mask AP on COCO test-dev. The code is released at: https://github.com/hding2455/DSC.

Siyuan Qiao, Huiyu Wang, Chenxi Liu et al.

In this paper, we study normalization methods for neural networks from the perspective of elimination singularity. Elimination singularities correspond to the points on the training trajectory where neurons become consistently deactivated. They cause degenerate manifolds in the loss landscape which will slow down training and harm model performances. We show that channel-based normalizations (e.g. Layer Normalization and Group Normalization) are unable to guarantee a far distance from elimination singularities, in contrast with Batch Normalization which by design avoids models from getting too close to them. To address this issue, we propose BatchChannel Normalization (BCN), which uses batch knowledge to avoid the elimination singularities in the training of channel-normalized models. Unlike Batch Normalization, BCN is able to run in both large-batch and micro-batch training settings. The effectiveness of BCN is verified on many tasks, including image classification, object detection, instance segmentation, and semantic segmentation. The code is here: https://github.com/joe-siyuan-qiao/Batch-Channel-Normalization.

Zhishuai Zhang, Wei Shen, Siyuan Qiao et al.

Recent anchor-based deep face detectors have achieved promising performance, but they are still struggling to detect hard faces, such as small, blurred and partially occluded faces. A reason is that they treat all images and faces equally, without putting more effort on hard ones; however, many training images only contain easy faces, which are less helpful to achieve better performance on hard images. In this paper, we propose that the robustness of a face detector against hard faces can be improved by learning small faces on hard images. Our intuitions are (1) hard images are the images which contain at least one hard face, thus they facilitate training robust face detectors; (2) most hard faces are small faces and other types of hard faces can be easily converted to small faces by shrinking. We build an anchor-based deep face detector, which only output a single feature map with small anchors, to specifically learn small faces and train it by a novel hard image mining strategy. Extensive experiments have been conducted on WIDER FACE, FDDB, Pascal Faces, and AFW datasets to show the effectiveness of our method. Our method achieves APs of 95.7, 94.9 and 89.7 on easy, medium and hard WIDER FACE val dataset respectively, which surpass the previous state-of-the-arts, especially on the hard subset. Code and model are available at https://github.com/bairdzhang/smallhardface.

Kai Zhang, Yi Luan, Hexiang Hu et al. · microsoft-research

Image retrieval, i.e., finding desired images given a reference image, inherently encompasses rich, multi-faceted search intents that are difficult to capture solely using image-based measures. Recent works leverage text instructions to allow users to more freely express their search intents. However, they primarily focus on image pairs that are visually similar and/or can be characterized by a small set of pre-defined relations. The core thesis of this paper is that text instructions can enable retrieving images with richer relations beyond visual similarity. To show this, we introduce MagicLens, a series of self-supervised image retrieval models that support open-ended instructions. MagicLens is built on a key novel insight: image pairs that naturally occur on the same web pages contain a wide range of implicit relations (e.g., inside view of), and we can bring those implicit relations explicit by synthesizing instructions via foundation models. Trained on 36.7M (query image, instruction, target image) triplets with rich semantic relations mined from the web, MagicLens achieves results comparable with or better than prior best on eight benchmarks of various image retrieval tasks, while maintaining high parameter efficiency with a significantly smaller model size. Additional human analyses on a 1.4M-image unseen corpus further demonstrate the diversity of search intents supported by MagicLens. Code and models are publicly available at https://open-vision-language.github.io/MagicLens/.

Lijie Fan, Luming Tang, Siyang Qin et al. · deepmind

We present UniFluid, a unified autoregressive framework for joint visual generation and understanding leveraging continuous visual tokens. Our unified autoregressive architecture processes multimodal image and text inputs, generating discrete tokens for text and continuous tokens for image. We find though there is an inherent trade-off between the image generation and understanding task, a carefully tuned training recipe enables them to improve each other. By selecting an appropriate loss balance weight, the unified model achieves results comparable to or exceeding those of single-task baselines on both tasks. Furthermore, we demonstrate that employing stronger pre-trained LLMs and random-order generation during training is important to achieve high-fidelity image generation within this unified framework. Built upon the Gemma model series, UniFluid exhibits competitive performance across both image generation and understanding, demonstrating strong transferability to various downstream tasks, including image editing for generation, as well as visual captioning and question answering for understanding.

Yuwei Zhang, Kumar Ayush, Siyuan Qiao et al.

We present SensorLM, a family of sensor-language foundation models that enable wearable sensor data understanding with natural language. Despite its pervasive nature, aligning and interpreting sensor data with language remains challenging due to the lack of paired, richly annotated sensor-text descriptions in uncurated, real-world wearable data. We introduce a hierarchical caption generation pipeline designed to capture statistical, structural, and semantic information from sensor data. This approach enabled the curation of the largest sensor-language dataset to date, comprising over 59.7 million hours of data from more than 103,000 people. Furthermore, SensorLM extends prominent multimodal pretraining architectures (e.g., CLIP, CoCa) and recovers them as specific variants within a generic architecture. Extensive experiments on real-world tasks in human activity analysis and healthcare verify the superior performance of SensorLM over state-of-the-art in zero-shot recognition, few-shot learning, and cross-modal retrieval. SensorLM also demonstrates intriguing capabilities including scaling behaviors, label efficiency, sensor captioning, and zero-shot generalization to unseen tasks.

Yaqiong Li, Peng Zhang, Hansu Gu et al.

Although there have been automated approaches and tools supporting toxicity censorship for social posts, most of them focus on detection. Toxicity censorship is a complex process, wherein detection is just an initial task and a user can have further needs such as rationale understanding and content modification. For this problem, we conduct a needfinding study to investigate people's diverse needs in toxicity censorship and then build a ChatGPT-based censorship tool named DeMod accordingly. DeMod is equipped with the features of explainable Detection and personalized Modification, providing fine-grained detection results, detailed explanations, and personalized modification suggestions. We also implemented the tool and recruited 35 Weibo users for evaluation. The results suggest DeMod's multiple strengths like the richness of functionality, the accuracy of censorship, and ease of use. Based on the findings, we further propose several insights into the design of content censorship systems.

Bohong Wu, Mengzhao Chen, Xiang Luo et al.

Large Language Models (LLMs) are powerful but often too slow and costly for real-world use during inference. Looped transformers save on parameters by reusing the same weights for multiple computational steps, or "loops." However, this approach has a major flaw: the loops run one after another, causing inference latency and memory requirements to increase with each added loop. This makes them impractical for fast applications. To solve this problem, we introduce the Parallel Loop Transformer (PLT). PLT is a new architecture that delivers the performance benefits of a deep, looped model but with the low latency of a standard, non-looped model. PLT works using two key techniques. First, Cross-Loop Parallelism (CLP) breaks the sequential dependency by computing different loops for different tokens at the same time, all within a single pass. Second, to prevent memory costs from growing, we use an Efficient Representation Enhancement strategy. This method shares the memory (KV cache) from the first loop with all other loops. It then uses a Gated Sliding-Window Attention (G-SWA) to combine this shared global information with local information, maintaining high accuracy. Our experiments show that PLT achieves the high accuracy of a traditional looped model but with almost no extra latency or memory cost compared to a standard transformer.

Chen Zheng, Yuhang Cai, Deyi Liu et al. · bytedance

Modern large language models leverage Mixture-of-Experts (MoE) architectures for efficient scaling, but face a critical challenge: functionally similar experts are often selected simultaneously, creating redundant computation and limiting effective model capacity. Existing auxiliary balance loss methods improve token distribution but fail to address the underlying expert diversity problem. We introduce GatePro, a novel parameter-free method that directly promotes expert selection diversity. GatePro identifies the most similar expert pairs and introduces localized competition mechanisms, preventing redundant expert co-activation while maintaining natural expert specialization. Our comprehensive evaluation demonstrates GatePro's effectiveness across model scales and benchmarks. Analysis demonstrates GatePro's ability to achieve enhanced expert diversity, where experts develop more distinct and complementary capabilities, avoiding functional redundancy. This approach can be deployed hot-swappable during any training phase without additional learnable parameters, offering a practical solution for improving MoE effectiveness.

Chen Zheng, Yiyuan Ma, Yuan Yang et al.

The development of alignment and reasoning capabilities in large language models has seen remarkable progress through two paradigms: instruction tuning and reinforcement learning from human feedback (RLHF) alignment paradigm, and distillation-based reasoning fine-tuning paradigm. While both approaches prove effective independently, the third paradigm of applying RLHF to distillation-trained models presents significant challenges. Our investigation reveals two critical phenomena that emerge in this paradigm: Sequence Length Collapse, where language generation dramatically reduces during early RLHF training, and the Reward Hockey Stick Curve, featuring severe reward score drops followed by gradual recovery. These instabilities fundamentally compromise the model's alignment and reasoning capabilities. To address these challenges, we propose Balanced Actor Initialization (BAI), a two-stage weighted model merging approach. BAI first merges instruction-following and distillation-based reasoning fine-tuned models, then further combines this intermediate model with the pretrained model to preserve foundational knowledge. Through comprehensive experiments across diverse benchmarks and detailed analysis of training experiments, we demonstrate that BAI resolves Sequence Length Collapse, mitigates the Reward Hockey Stick Curve, and enables continuous sequence length improvement during training. Additionally, our analysis reveals that balanced merging ratios achieve optimal trade-offs between training stability and reasoning capability preservation. Our work provides the effective solution for stable training in this third paradigm, enabling more capable reasoning models that combine distillation efficiency with RLHF alignment.

Zihao Huang, Yu Bao, Qiyang Min et al. · bytedance

While Mixture of Experts (MoE) models achieve remarkable efficiency by activating only subsets of parameters, they suffer from high memory access costs during inference. Memory-layer architectures offer an appealing alternative with very few memory access, but previous attempts like UltraMem have only matched the performance of 2-expert MoE models, falling significantly short of state-of-the-art 8-expert configurations. We present UltraMemV2, a redesigned memory-layer architecture that closes this performance gap. Our approach introduces five key improvements: integrating memory layers into every transformer block, simplifying value expansion with single linear projections, adopting FFN-based value processing from PEER, implementing principled parameter initialization, and rebalancing memory-to-FFN computation ratios. Through extensive evaluation, we demonstrate that UltraMemV2 achieves performance parity with 8-expert MoE models under same computation and parameters but significantly low memory access. Notably, UltraMemV2 shows superior performance on memory-intensive tasks, with improvements of +1.6 points on long-context memorization, +6.2 points on multi-round memorization, and +7.9 points on in-context learning. We validate our approach at scale with models up to 2.5B activated parameters from 120B total parameters, and establish that activation density has greater impact on performance than total sparse parameter count. Our work brings memory-layer architectures to performance parity with state-of-the-art MoE models, presenting a compelling alternative for efficient sparse computation.

Yaqiong Li, Peng Zhang, Lin Wang et al.

Risk perception is subjective, and youth's understanding of toxic content differs from that of adults. Although previous research has conducted extensive studies on toxicity detection in social media, the investigation of youth's unique toxicity, i.e., languages perceived as nontoxic by adults but toxic as youth, is ignored. To address this gap, we aim to explore: 1) What are the features of ``youth-toxicity'' languages in social media (RQ1); 2) Can existing toxicity detection techniques accurately detect these languages (RQ2). For these questions, we took Chinese youth as the research target, constructed the first Chinese ``youth-toxicity'' dataset, and then conducted extensive analysis. Our results suggest that youth's perception of these is associated with several contextual factors, like the source of an utterance and text-related features. Incorporating these meta information into current toxicity detection methods significantly improves accuracy overall. Finally, we propose several insights into future research on youth-centered toxicity detection.

Lin Yang, Shawn Xu, Andrew Sellergren et al.

Many clinical tasks require an understanding of specialized data, such as medical images and genomics, which is not typically found in general-purpose large multimodal models. Building upon Gemini's multimodal models, we develop several models within the new Med-Gemini family that inherit core capabilities of Gemini and are optimized for medical use via fine-tuning with 2D and 3D radiology, histopathology, ophthalmology, dermatology and genomic data. Med-Gemini-2D sets a new standard for AI-based chest X-ray (CXR) report generation based on expert evaluation, exceeding previous best results across two separate datasets by an absolute margin of 1% and 12%, where 57% and 96% of AI reports on normal cases, and 43% and 65% on abnormal cases, are evaluated as "equivalent or better" than the original radiologists' reports. We demonstrate the first ever large multimodal model-based report generation for 3D computed tomography (CT) volumes using Med-Gemini-3D, with 53% of AI reports considered clinically acceptable, although additional research is needed to meet expert radiologist reporting quality. Beyond report generation, Med-Gemini-2D surpasses the previous best performance in CXR visual question answering (VQA) and performs well in CXR classification and radiology VQA, exceeding SoTA or baselines on 17 of 20 tasks. In histopathology, ophthalmology, and dermatology image classification, Med-Gemini-2D surpasses baselines across 18 out of 20 tasks and approaches task-specific model performance. Beyond imaging, Med-Gemini-Polygenic outperforms the standard linear polygenic risk score-based approach for disease risk prediction and generalizes to genetically correlated diseases for which it has never been trained. Although further development and evaluation are necessary in the safety-critical medical domain, our results highlight the potential of Med-Gemini across a wide range of medical tasks.

Gemini Team, Rohan Anil, Sebastian Borgeaud et al.

This report introduces a new family of multimodal models, Gemini, that exhibit remarkable capabilities across image, audio, video, and text understanding. The Gemini family consists of Ultra, Pro, and Nano sizes, suitable for applications ranging from complex reasoning tasks to on-device memory-constrained use-cases. Evaluation on a broad range of benchmarks shows that our most-capable Gemini Ultra model advances the state of the art in 30 of 32 of these benchmarks - notably being the first model to achieve human-expert performance on the well-studied exam benchmark MMLU, and improving the state of the art in every one of the 20 multimodal benchmarks we examined. We believe that the new capabilities of the Gemini family in cross-modal reasoning and language understanding will enable a wide variety of use cases. We discuss our approach toward post-training and deploying Gemini models responsibly to users through services including Gemini, Gemini Advanced, Google AI Studio, and Cloud Vertex AI.

Rohan Anil, Andrew M. Dai, Orhan Firat et al.

We introduce PaLM 2, a new state-of-the-art language model that has better multilingual and reasoning capabilities and is more compute-efficient than its predecessor PaLM. PaLM 2 is a Transformer-based model trained using a mixture of objectives. Through extensive evaluations on English and multilingual language, and reasoning tasks, we demonstrate that PaLM 2 has significantly improved quality on downstream tasks across different model sizes, while simultaneously exhibiting faster and more efficient inference compared to PaLM. This improved efficiency enables broader deployment while also allowing the model to respond faster, for a more natural pace of interaction. PaLM 2 demonstrates robust reasoning capabilities exemplified by large improvements over PaLM on BIG-Bench and other reasoning tasks. PaLM 2 exhibits stable performance on a suite of responsible AI evaluations, and enables inference-time control over toxicity without additional overhead or impact on other capabilities. Overall, PaLM 2 achieves state-of-the-art performance across a diverse set of tasks and capabilities. When discussing the PaLM 2 family, it is important to distinguish between pre-trained models (of various sizes), fine-tuned variants of these models, and the user-facing products that use these models. In particular, user-facing products typically include additional pre- and post-processing steps. Additionally, the underlying models may evolve over time. Therefore, one should not expect the performance of user-facing products to exactly match the results reported in this report.

Chenglin Yang, Siyuan Qiao, Adam Kortylewski et al.

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

Siyuan Qiao, Yukun Zhu, Hartwig Adam et al.

In this paper, we present ViP-DeepLab, a unified model attempting to tackle the long-standing and challenging inverse projection problem in vision, which we model as restoring the point clouds from perspective image sequences while providing each point with instance-level semantic interpretations. Solving this problem requires the vision models to predict the spatial location, semantic class, and temporally consistent instance label for each 3D point. ViP-DeepLab approaches it by jointly performing monocular depth estimation and video panoptic segmentation. We name this joint task as Depth-aware Video Panoptic Segmentation, and propose a new evaluation metric along with two derived datasets for it, which will be made available to the public. On the individual sub-tasks, ViP-DeepLab also achieves state-of-the-art results, outperforming previous methods by 5.1% VPQ on Cityscapes-VPS, ranking 1st on the KITTI monocular depth estimation benchmark, and 1st on KITTI MOTS pedestrian. The datasets and the evaluation codes are made publicly available.

Liang-Chieh Chen, Huiyu Wang, Siyuan Qiao

The Wide Residual Networks (Wide-ResNets), a shallow but wide model variant of the Residual Networks (ResNets) by stacking a small number of residual blocks with large channel sizes, have demonstrated outstanding performance on multiple dense prediction tasks. However, since proposed, the Wide-ResNet architecture has barely evolved over the years. In this work, we revisit its architecture design for the recent challenging panoptic segmentation task, which aims to unify semantic segmentation and instance segmentation. A baseline model is obtained by incorporating the simple and effective Squeeze-and-Excitation and Switchable Atrous Convolution to the Wide-ResNets. Its network capacity is further scaled up or down by adjusting the width (i.e., channel size) and depth (i.e., number of layers), resulting in a family of SWideRNets (short for Scaling Wide Residual Networks). We demonstrate that such a simple scaling scheme, coupled with grid search, identifies several SWideRNets that significantly advance state-of-the-art performance on panoptic segmentation datasets in both the fast model regime and strong model regime.

Siyuan Qiao, Liang-Chieh Chen, Alan Yuille

Many modern object detectors demonstrate outstanding performances by using the mechanism of looking and thinking twice. In this paper, we explore this mechanism in the backbone design for object detection. At the macro level, we propose Recursive Feature Pyramid, which incorporates extra feedback connections from Feature Pyramid Networks into the bottom-up backbone layers. At the micro level, we propose Switchable Atrous Convolution, which convolves the features with different atrous rates and gathers the results using switch functions. Combining them results in DetectoRS, which significantly improves the performances of object detection. On COCO test-dev, DetectoRS achieves state-of-the-art 55.7% box AP for object detection, 48.5% mask AP for instance segmentation, and 50.0% PQ for panoptic segmentation. The code is made publicly available.

Mingqing Xiao, Adam Kortylewski, Ruihai Wu et al.

Despite deep convolutional neural networks' great success in object classification, it suffers from severe generalization performance drop under occlusion due to the inconsistency between training and testing data. Because of the large variance of occluders, our goal is a model trained on occlusion-free data while generalizable to occlusion conditions. In this work, we integrate prototypes, partial matching and top-down attention regulation into deep neural networks to realize robust object classification under occlusion. We first introduce prototype learning as its regularization encourages compact data clusters, which enables better generalization ability under inconsistent conditions. Then, attention map at intermediate layer based on feature dictionary and activation scale is estimated for partial matching, which sifts irrelevant information out when comparing features with prototypes. Further, inspired by neuroscience research that reveals the important role of feedback connection for object recognition under occlusion, a top-down feedback attention regulation is introduced into convolution layers, purposefully reducing the contamination by occlusion during feature extraction stage. Our experiment results on partially occluded MNIST and vehicles from the PASCAL3D+ dataset demonstrate that the proposed network significantly improves the robustness of current deep neural networks under occlusion. Our code will be released.

Siyuan Qiao, Huiyu Wang, Chenxi Liu et al.

Batch Normalization (BN) has become an out-of-box technique to improve deep network training. However, its effectiveness is limited for micro-batch training, i.e., each GPU typically has only 1-2 images for training, which is inevitable for many computer vision tasks, e.g., object detection and semantic segmentation, constrained by memory consumption. To address this issue, we propose Weight Standardization (WS) and Batch-Channel Normalization (BCN) to bring two success factors of BN into micro-batch training: 1) the smoothing effects on the loss landscape and 2) the ability to avoid harmful elimination singularities along the training trajectory. WS standardizes the weights in convolutional layers to smooth the loss landscape by reducing the Lipschitz constants of the loss and the gradients; BCN combines batch and channel normalizations and leverages estimated statistics of the activations in convolutional layers to keep networks away from elimination singularities. We validate WS and BCN on comprehensive computer vision tasks, including image classification, object detection, instance segmentation, video recognition and semantic segmentation. All experimental results consistently show that WS and BCN improve micro-batch training significantly. Moreover, using WS and BCN with micro-batch training is even able to match or outperform the performances of BN with large-batch training.

Siyuan Qiao, Zhe Lin, Jianming Zhang et al.

In this paper, we study the problem of improving computational resource utilization of neural networks. Deep neural networks are usually over-parameterized for their tasks in order to achieve good performances, thus are likely to have underutilized computational resources. This observation motivates a lot of research topics, e.g. network pruning, architecture search, etc. As models with higher computational costs (e.g. more parameters or more computations) usually have better performances, we study the problem of improving the resource utilization of neural networks so that their potentials can be further realized. To this end, we propose a novel optimization method named Neural Rejuvenation. As its name suggests, our method detects dead neurons and computes resource utilization in real time, rejuvenates dead neurons by resource reallocation and reinitialization, and trains them with new training schemes. By simply replacing standard optimizers with Neural Rejuvenation, we are able to improve the performances of neural networks by a very large margin while using similar training efforts and maintaining their original resource usages.

Xutong Ren, Lingxi Xie, Chen Wei et al.

Computer vision is difficult, partly because the desired mathematical function connecting input and output data is often complex, fuzzy and thus hard to learn. Coarse-to-fine (C2F) learning is a promising direction, but it remains unclear how it is applied to a wide range of vision problems. This paper presents a generalized C2F framework by making two technical contributions. First, we provide a unified way of C2F propagation, in which the coarse prediction (a class vector, a detected box, a segmentation mask, etc.) is encoded into a dense (pixel-level) matrix and concatenated to the original input, so that the fine model takes the same design of the coarse model but sees additional information. Second, we present a progressive training strategy which starts with feeding the ground-truth instead of the coarse output into the fine model, and gradually increases the fraction of coarse output, so that at the end of training the fine model is ready for testing. We also relate our approach to curriculum learning by showing that data difficulty keeps increasing during the training process. We apply our framework to three vision tasks including image classification, object localization and semantic segmentation, and demonstrate consistent accuracy gain compared to the baseline training strategy.

Chenglin Yang, Lingxi Xie, Siyuan Qiao et al.

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

Yan Wang, Lingxi Xie, Siyuan Qiao et al.

Convolution is spatially-symmetric, i.e., the visual features are independent of its position in the image, which limits its ability to utilize contextual cues for visual recognition. This paper addresses this issue by introducing a recalibration process, which refers to the surrounding region of each neuron, computes an importance value and multiplies it to the original neural response. Our approach is named multi-scale spatially-asymmetric recalibration (MS-SAR), which extracts visual cues from surrounding regions at multiple scales, and designs a weighting scheme which is asymmetric in the spatial domain. MS-SAR is implemented in an efficient way, so that only small fractions of extra parameters and computations are required. We apply MS-SAR to several popular building blocks, including the residual block and the densely-connected block, and demonstrate its superior performance in both CIFAR and ILSVRC2012 classification tasks.

Siyuan Qiao, Wei Shen, Zhishuai Zhang et al.

In this paper, we study the problem of semi-supervised image recognition, which is to learn classifiers using both labeled and unlabeled images. We present Deep Co-Training, a deep learning based method inspired by the Co-Training framework. The original Co-Training learns two classifiers on two views which are data from different sources that describe the same instances. To extend this concept to deep learning, Deep Co-Training trains multiple deep neural networks to be the different views and exploits adversarial examples to encourage view difference, in order to prevent the networks from collapsing into each other. As a result, the co-trained networks provide different and complementary information about the data, which is necessary for the Co-Training framework to achieve good results. We test our method on SVHN, CIFAR-10/100 and ImageNet datasets, and our method outperforms the previous state-of-the-art methods by a large margin.

Zhishuai Zhang, Siyuan Qiao, Cihang Xie et al.

We propose a novel single shot object detection network named Detection with Enriched Semantics (DES). Our motivation is to enrich the semantics of object detection features within a typical deep detector, by a semantic segmentation branch and a global activation module. The segmentation branch is supervised by weak segmentation ground-truth, i.e., no extra annotation is required. In conjunction with that, we employ a global activation module which learns relationship between channels and object classes in a self-supervised manner. Comprehensive experimental results on both PASCAL VOC and MS COCO detection datasets demonstrate the effectiveness of the proposed method. In particular, with a VGG16 based DES, we achieve an mAP of 81.7 on VOC2007 test and an mAP of 32.8 on COCO test-dev with an inference speed of 31.5 milliseconds per image on a Titan Xp GPU. With a lower resolution version, we achieve an mAP of 79.7 on VOC2007 with an inference speed of 13.0 milliseconds per image.

Siyuan Qiao, Zhishuai Zhang, Wei Shen et al.

Depth is one of the keys that make neural networks succeed in the task of large-scale image recognition. The state-of-the-art network architectures usually increase the depths by cascading convolutional layers or building blocks. In this paper, we present an alternative method to increase the depth. Our method is by introducing computation orderings to the channels within convolutional layers or blocks, based on which we gradually compute the outputs in a channel-wise manner. The added orderings not only increase the depths and the learning capacities of the networks without any additional computation costs, but also eliminate the overlap singularities so that the networks are able to converge faster and perform better. Experiments show that the networks based on our method achieve the state-of-the-art performances on CIFAR and ImageNet datasets.

Boyang Deng, Qing Liu, Siyuan Qiao et al.

Convolutional neural networks (CNNs) are one of the driving forces for the advancement of computer vision. Despite their promising performances on many tasks, CNNs still face major obstacles on the road to achieving ideal machine intelligence. One is that CNNs are complex and hard to interpret. Another is that standard CNNs require large amounts of annotated data, which is sometimes hard to obtain, and it is desirable to learn to recognize objects from few examples. In this work, we address these limitations of CNNs by developing novel, flexible, and interpretable models for few-shot learning. Our models are based on the idea of encoding objects in terms of visual concepts (VCs), which are interpretable visual cues represented by the feature vectors within CNNs. We first adapt the learning of VCs to the few-shot setting, and then uncover two key properties of feature encoding using VCs, which we call category sensitivity and spatial pattern. Motivated by these properties, we present two intuitive models for the problem of few-shot learning. Experiments show that our models achieve competitive performances, while being more flexible and interpretable than alternative state-of-the-art few-shot learning methods. We conclude that using VCs helps expose the natural capability of CNNs for few-shot learning.

Siyuan Qiao, Chenxi Liu, Wei Shen et al.

In this paper, we are interested in the few-shot learning problem. In particular, we focus on a challenging scenario where the number of categories is large and the number of examples per novel category is very limited, e.g. 1, 2, or 3. Motivated by the close relationship between the parameters and the activations in a neural network associated with the same category, we propose a novel method that can adapt a pre-trained neural network to novel categories by directly predicting the parameters from the activations. Zero training is required in adaptation to novel categories, and fast inference is realized by a single forward pass. We evaluate our method by doing few-shot image recognition on the ImageNet dataset, which achieves the state-of-the-art classification accuracy on novel categories by a significant margin while keeping comparable performance on the large-scale categories. We also test our method on the MiniImageNet dataset and it strongly outperforms the previous state-of-the-art methods.

Siyuan Qiao, Wei Shen, Weichao Qiu et al.

Motivated by product detection in supermarkets, this paper studies the problem of object proposal generation in supermarket images and other natural images. We argue that estimation of object scales in images is helpful for generating object proposals, especially for supermarket images where object scales are usually within a small range. Therefore, we propose to estimate object scales of images before generating object proposals. The proposed method for predicting object scales is called ScaleNet. To validate the effectiveness of ScaleNet, we build three supermarket datasets, two of which are real-world datasets used for testing and the other one is a synthetic dataset used for training. In short, we extend the previous state-of-the-art object proposal methods by adding a scale prediction phase. The resulted method outperforms the previous state-of-the-art on the supermarket datasets by a large margin. We also show that the approach works for object proposal on other natural images and it outperforms the previous state-of-the-art object proposal methods on the MS COCO dataset. The supermarket datasets, the virtual supermarkets, and the tools for creating more synthetic datasets will be made public.