Kai Kang

Mingze Xu, Mingfei Gao, Zhe Gan et al.

We propose SlowFast-LLaVA (or SF-LLaVA for short), a training-free video large language model (LLM) that can jointly capture detailed spatial semantics and long-range temporal context without exceeding the token budget of commonly used LLMs. This is realized by using a two-stream SlowFast design of inputs for Video LLMs to aggregate features from sampled frames in an effective way. Specifically, the Slow pathway extracts features at a low frame rate while keeping as much spatial detail as possible (e.g., with 12x24 tokens), and the Fast pathway operates on a high frame rate but uses a larger spatial pooling stride (e.g., downsampling 6x) to focus on the motion cues. As a result, this design allows us to adequately capture both spatial and temporal features that are beneficial for detailed video understanding. Experimental results show that SF-LLaVA outperforms existing training-free methods on a wide range of video tasks. On some benchmarks, it achieves comparable or even better performance compared to state-of-the-art Video LLMs that are fine-tuned on video datasets. Code has been made available at: https://github.com/apple/ml-slowfast-llava.

Vincent Perot, Kai Kang, Florian Luisier et al.

Large Language Models (LLM) have revolutionized Natural Language Processing (NLP), improving state-of-the-art and exhibiting emergent capabilities across various tasks. However, their application in extracting information from visually rich documents, which is at the core of many document processing workflows and involving the extraction of key entities from semi-structured documents, has not yet been successful. The main obstacles to adopting LLMs for this task include the absence of layout encoding within LLMs, which is critical for high quality extraction, and the lack of a grounding mechanism to localize the predicted entities within the document. In this paper, we introduce Language Model-based Document Information Extraction and Localization (LMDX), a methodology to reframe the document information extraction task for a LLM. LMDX enables extraction of singular, repeated, and hierarchical entities, both with and without training data, while providing grounding guarantees and localizing the entities within the document. Finally, we apply LMDX to the PaLM 2-S and Gemini Pro LLMs and evaluate it on VRDU and CORD benchmarks, setting a new state-of-the-art and showing how LMDX enables the creation of high quality, data-efficient parsers.

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.

Chengsi Yao, Ge Wang, Kai Kang et al.

Generative Control Policies (GCPs) show immense promise in robotic manipulation but struggle to simultaneously model stable global motions and high-frequency local corrections. While modern architectures extract multi-scale spatial features, their underlying Probability Flow ODEs apply a uniform temporal integration schedule. Compressed to a single step for real-time Receding Horizon Control (RHC), uniform ODE solvers mathematically smooth over sparse, high-frequency transients entangled within low-frequency steady states. To decouple these dynamics without accumulating pipelined errors, we introduce KoopmanFlow, a parameter-efficient generative policy guided by a Koopman-inspired structural inductive bias. Operating in a unified multimodal latent space with visual context, KoopmanFlow bifurcates generation at the terminal stage. Because visual conditioning occurs before spectral decomposition, both branches are visually guided yet temporally specialized. A macroscopic branch anchors slow-varying trajectories via single-step Consistency Training, while a transient branch uses Flow Matching to isolate high-frequency residuals stimulated by sudden visual cues (e.g., contacts or occlusions). Guided by an explicit spectral prior and optimized via a novel asymmetric consistency objective, KoopmanFlow establishes a fused co-training mechanism. This allows the variant branch to absorb localized dynamics without multi-stage error accumulation. Extensive experiments show KoopmanFlow significantly outperforms state-of-the-art baselines in contact-rich tasks requiring agile disturbance rejection. By trading a surplus latency buffer for a richer structural prior, KoopmanFlow achieves superior control fidelity and parameter efficiency within real-time deployment limits.

Yuzhuo Chen, Zehua Ma, Jianhua Wang et al.

In controllable image synthesis, generating coherent and consistent images from multiple references with spatial layout awareness remains an open challenge. We present LAMIC, a Layout-Aware Multi-Image Composition framework that, for the first time, extends single-reference diffusion models to multi-reference scenarios in a training-free manner. Built upon the MMDiT model, LAMIC introduces two plug-and-play attention mechanisms: 1) Group Isolation Attention (GIA) to enhance entity disentanglement; and 2) Region-Modulated Attention (RMA) to enable layout-aware generation. To comprehensively evaluate model capabilities, we further introduce three metrics: 1) Inclusion Ratio (IN-R) and Fill Ratio (FI-R) for assessing layout control; and 2) Background Similarity (BG-S) for measuring background consistency. Extensive experiments show that LAMIC achieves state-of-the-art performance across most major metrics: it consistently outperforms existing multi-reference baselines in ID-S, BG-S, IN-R and AVG scores across all settings, and achieves the best DPG in complex composition tasks. These results demonstrate LAMIC's superior abilities in identity keeping, background preservation, layout control, and prompt-following, all achieved without any training or fine-tuning, showcasing strong zero-shot generalization ability. By inheriting the strengths of advanced single-reference models and enabling seamless extension to multi-image scenarios, LAMIC establishes a new training-free paradigm for controllable multi-image composition. As foundation models continue to evolve, LAMIC's performance is expected to scale accordingly. Our implementation is available at: https://github.com/Suchenl/LAMIC.

Erik Daxberger, Nina Wenzel, David Griffiths et al.

Multimodal large language models (MLLMs) excel at 2D visual understanding but remain limited in their ability to reason about 3D space. In this work, we leverage large-scale high-quality 3D scene data with open-set annotations to introduce 1) a novel supervised fine-tuning dataset and 2) a new evaluation benchmark, focused on indoor scenes. Our Cubify Anything VQA (CA-VQA) data covers diverse spatial tasks including spatial relationship prediction, metric size and distance estimation, and 3D grounding. We show that CA-VQA enables us to train MM-Spatial, a strong generalist MLLM that also achieves state-of-the-art performance on 3D spatial understanding benchmarks, including our own. We show how incorporating metric depth and multi-view inputs (provided in CA-VQA) can further improve 3D understanding, and demonstrate that data alone allows our model to achieve depth perception capabilities comparable to dedicated monocular depth estimation models.

Mingze Xu, Mingfei Gao, Shiyu Li et al.

We introduce SlowFast-LLaVA-1.5 (abbreviated as SF-LLaVA-1.5), a family of video large language models (LLMs) offering a token-efficient solution for long-form video understanding. We incorporate the two-stream SlowFast mechanism into a streamlined training pipeline, and perform joint video-image training on a carefully curated data mixture of only publicly available datasets. Our primary focus is on highly efficient model scales (1B and 3B), demonstrating that even relatively small Video LLMs can achieve state-of-the-art performance on video understanding, meeting the demand for mobile-friendly models. Experimental results demonstrate that SF-LLaVA-1.5 achieves superior performance on a wide range of video and image tasks, with robust results at all model sizes (ranging from 1B to 7B). Notably, SF-LLaVA-1.5 achieves state-of-the-art results in long-form video understanding (e.g., LongVideoBench and MLVU) and excels at small scales across various video benchmarks.

Kai Kang, Feng Liu

Renewable energy (RE) generation exhibits pronounced seasonality and variability, and neglecting these features can lead to significant underestimation of long-term power system risks in power supply. While long-term dispatch strategies are essential for evaluating and mitigating tail risks, they are often excluded from existing models due to their complexity. This paper proposes a long-term tail risk assessment and mitigation framework for renewable power systems, explicitly embedding dispatch strategies. A representative scenario generation method is designed, combining multi-timescale Copula modeling to capture RE's long-range variability and correlation. Building on these scenarios, an evolution-based risk assessment model is established, where Conditional Value-at-Risk (CVaR) is employed as a robust metric to quantify tail risks. Finally, a controlled evolution-based risk mitigation scheme is introduced to refine long-term dispatch strategies for mitigating tail risks. Case studies on a modified IEEE-39 bus system incorporating real-world data substantiate the efficacy of the proposed method.

Mingyang Li, Viktor Schlegel, Tingting Mu et al.

Automated clinical coding involves mapping unstructured text from Electronic Health Records (EHRs) to standardized code systems such as the International Classification of Diseases (ICD). While recent advances in deep learning have significantly improved the accuracy and efficiency of ICD coding, the lack of explainability in these models remains a major limitation, undermining trust and transparency. Current explorations about explainability largely rely on attention-based techniques and qualitative assessments by physicians, yet lack systematic evaluation using consistent criteria on high-quality rationale datasets, as well as dedicated approaches explicitly trained to generate rationales for further enhancing explanation. In this work, we conduct a comprehensive evaluation of the explainability of the rationales for ICD coding through two key lenses: faithfulness that evaluates how well explanations reflect the model's actual reasoning and plausibility that measures how consistent the explanations are with human expert judgment. To facilitate the evaluation of plausibility, we construct a new rationale-annotated dataset, offering denser annotations with diverse granularity and aligns better with current clinical practice, and conduct evaluation across three types of rationales of ICD coding. Encouraged by the promising plausibility of LLM-generated rationales for ICD coding, we further propose new rationale learning methods to improve the quality of model-generated rationales, where rationales produced by prompting LLMs with/without annotation examples are used as distant supervision signals. We empirically find that LLM-generated rationales align most closely with those of human experts. Moreover, incorporating few-shot human-annotated examples not only further improves rationale generation but also enhances rationale-learning approaches.

Justin Lazarow, Kai Kang, Afshin Dehghan

We revisit scene-level 3D object detection as the output of an object-centric framework capable of both localization and mapping using 3D oriented boxes as the underlying geometric primitive. While existing 3D object detection approaches operate globally and implicitly rely on the a priori existence of metric camera poses, our method, Rooms from Motion (RfM) operates on a collection of un-posed images. By replacing the standard 2D keypoint-based matcher of structure-from-motion with an object-centric matcher based on image-derived 3D boxes, we estimate metric camera poses, object tracks, and finally produce a global, semantic 3D object map. When a priori pose is available, we can significantly improve map quality through optimization of global 3D boxes against individual observations. RfM shows strong localization performance and subsequently produces maps of higher quality than leading point-based and multi-view 3D object detection methods on CA-1M and ScanNet++, despite these global methods relying on overparameterization through point clouds or dense volumes. Rooms from Motion achieves a general, object-centric representation which not only extends the work of Cubify Anything to full scenes but also allows for inherently sparse localization and parametric mapping proportional to the number of objects in a scene.

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.

Liangchen Song, Liangliang Cao, Hongyu Xu et al.

The techniques for 3D indoor scene capturing are widely used, but the meshes produced leave much to be desired. In this paper, we propose "RoomDreamer", which leverages powerful natural language to synthesize a new room with a different style. Unlike existing image synthesis methods, our work addresses the challenge of synthesizing both geometry and texture aligned to the input scene structure and prompt simultaneously. The key insight is that a scene should be treated as a whole, taking into account both scene texture and geometry. The proposed framework consists of two significant components: Geometry Guided Diffusion and Mesh Optimization. Geometry Guided Diffusion for 3D Scene guarantees the consistency of the scene style by applying the 2D prior to the entire scene simultaneously. Mesh Optimization improves the geometry and texture jointly and eliminates the artifacts in the scanned scene. To validate the proposed method, real indoor scenes scanned with smartphones are used for extensive experiments, through which the effectiveness of our method is demonstrated.

Kai Kang, Fangzhen Lin

A class or taxonomic hierarchy is often manually constructed, and part of our knowledge about the world. In this paper, we propose a novel algorithm for automatically acquiring a class hierarchy from a classifier which is often a large neural network these days. The information that we need from a classifier is its confusion matrix which contains, for each pair of base classes, the number of errors the classifier makes by mistaking one for another. Our algorithm produces surprisingly good hierarchies for some well-known deep neural network models trained on the CIFAR-10 dataset, a neural network model for predicting the native language of a non-native English speaker, a neural network model for detecting the language of a written text, and a classifier for identifying music genre. In the literature, such class hierarchies have been used to provide interpretability to the neural networks. We also discuss some other potential uses of the acquired hierarchies.

Qiyu Hu, Yunlong Cai, Kai Kang et al.

In this paper, we propose an end-to-end deep learning-based joint transceiver design algorithm for millimeter wave (mmWave) massive multiple-input multiple-output (MIMO) systems, which consists of deep neural network (DNN)-aided pilot training, channel feedback, and hybrid analog-digital (HAD) precoding. Specifically, we develop a DNN architecture that maps the received pilots into feedback bits at the receiver, and then further maps the feedback bits into the hybrid precoder at the transmitter. To reduce the signaling overhead and channel state information (CSI) mismatch caused by the transmission delay, a two-timescale DNN composed of a long-term DNN and a short-term DNN is developed. The analog precoders are designed by the long-term DNN based on the CSI statistics and updated once in a frame consisting of a number of time slots. In contrast, the digital precoders are optimized by the short-term DNN at each time slot based on the estimated low-dimensional equivalent CSI matrices. A two-timescale training method is also developed for the proposed DNN with a binary layer. We then analyze the generalization ability and signaling overhead for the proposed DNN based algorithm. Simulation results show that our proposed technique significantly outperforms conventional schemes in terms of bit-error rate performance with reduced signaling overhead and shorter pilot sequences.

Vincent Casser, Kai Kang, Hanspeter Pfister et al.

High-resolution connectomics data allows for the identification of dysfunctional mitochondria which are linked to a variety of diseases such as autism or bipolar. However, manual analysis is not feasible since datasets can be petabytes in size. We present a fully automatic mitochondria detector based on a modified U-Net architecture that yields high accuracy and fast processing times. We evaluate our method on multiple real-world connectomics datasets, including an improved version of the EPFL mitochondria benchmark. Our results show an Jaccard index of up to 0.90 with inference times lower than 16ms for a 512x512px image tile. This speed is faster than the acquisition speed of modern electron microscopes, enabling mitochondria detection in real-time. Our detector ranks first for real-time detection when compared to previous works and data, results, and code are openly available.

Kai Kang, Hongsheng Li, Tong Xiao et al.

Object detection in videos has drawn increasing attention recently with the introduction of the large-scale ImageNet VID dataset. Different from object detection in static images, temporal information in videos is vital for object detection. To fully utilize temporal information, state-of-the-art methods are based on spatiotemporal tubelets, which are essentially sequences of associated bounding boxes across time. However, the existing methods have major limitations in generating tubelets in terms of quality and efficiency. Motion-based methods are able to obtain dense tubelets efficiently, but the lengths are generally only several frames, which is not optimal for incorporating long-term temporal information. Appearance-based methods, usually involving generic object tracking, could generate long tubelets, but are usually computationally expensive. In this work, we propose a framework for object detection in videos, which consists of a novel tubelet proposal network to efficiently generate spatiotemporal proposals, and a Long Short-term Memory (LSTM) network that incorporates temporal information from tubelet proposals for achieving high object detection accuracy in videos. Experiments on the large-scale ImageNet VID dataset demonstrate the effectiveness of the proposed framework for object detection in videos.

Kai Kang, Wanli Ouyang, Hongsheng Li et al.

Deep Convolution Neural Networks (CNNs) have shown impressive performance in various vision tasks such as image classification, object detection and semantic segmentation. For object detection, particularly in still images, the performance has been significantly increased last year thanks to powerful deep networks (e.g. GoogleNet) and detection frameworks (e.g. Regions with CNN features (R-CNN)). The lately introduced ImageNet task on object detection from video (VID) brings the object detection task into the video domain, in which objects' locations at each frame are required to be annotated with bounding boxes. In this work, we introduce a complete framework for the VID task based on still-image object detection and general object tracking. Their relations and contributions in the VID task are thoroughly studied and evaluated. In addition, a temporal convolution network is proposed to incorporate temporal information to regularize the detection results and shows its effectiveness for the task.

Kai Kang, Hongsheng Li, Junjie Yan et al.

The state-of-the-art performance for object detection has been significantly improved over the past two years. Besides the introduction of powerful deep neural networks such as GoogleNet and VGG, novel object detection frameworks such as R-CNN and its successors, Fast R-CNN and Faster R-CNN, play an essential role in improving the state-of-the-art. Despite their effectiveness on still images, those frameworks are not specifically designed for object detection from videos. Temporal and contextual information of videos are not fully investigated and utilized. In this work, we propose a deep learning framework that incorporates temporal and contextual information from tubelets obtained in videos, which dramatically improves the baseline performance of existing still-image detection frameworks when they are applied to videos. It is called T-CNN, i.e. tubelets with convolutional neueral networks. The proposed framework won the recently introduced object-detection-from-video (VID) task with provided data in the ImageNet Large-Scale Visual Recognition Challenge 2015 (ILSVRC2015).

Kai Kang, Xiaogang Wang

In this paper, we propose a fast fully convolutional neural network (FCNN) for crowd segmentation. By replacing the fully connected layers in CNN with 1 by 1 convolution kernels, FCNN takes whole images as inputs and directly outputs segmentation maps by one pass of forward propagation. It has the property of translation invariance like patch-by-patch scanning but with much lower computation cost. Once FCNN is learned, it can process input images of any sizes without warping them to a standard size. These attractive properties make it extendable to other general image segmentation problems. Based on FCNN, a multi-stage deep learning is proposed to integrate appearance and motion cues for crowd segmentation. Both appearance filters and motion filers are pretrained stage-by-stage and then jointly optimized. Different combination methods are investigated. The effectiveness of our approach and component-wise analysis are evaluated on two crowd segmentation datasets created by us, which include image frames from 235 and 11 scenes, respectively. They are currently the largest crowd segmentation datasets and will be released to the public.