Unified multimodal Large Language Models (LLMs) that can both understand and generate visual content hold immense potential. However, existing open-source models often suffer from a performance trade-off between these capabilities. We present Manzano, a simple and scalable unified framework that substantially reduces this tension by coupling a hybrid image tokenizer with a well-curated training recipe. A single shared vision encoder feeds two lightweight adapters that produce continuous embeddings for image-to-text understanding and discrete tokens for text-to-image generation within a common semantic space. A unified autoregressive LLM predicts high-level semantics in the form of text and image tokens, with an auxiliary diffusion decoder subsequently translating the image tokens into pixels. The architecture, together with a unified training recipe over understanding and generation data, enables scalable joint learning of both capabilities. Manzano achieves state-of-the-art results among unified models, and is competitive with specialist models, particularly on text-rich evaluation. Our studies show minimal task conflicts and consistent gains from scaling model size, validating our design choice of a hybrid tokenizer.

Graphical user interface (GUI) agents autonomously complete tasks across platforms (\eg, Linux) by sequentially decomposing user instructions into action proposals that iteratively interact with visual elements in the evolving environment. However, two main challenges arise: i) planning (\ie, the action proposal sequence) under expansive action space, where selecting an appropriate plan is non-trivial, as many valid ones may exist; ii) accurately grounding actions in complex and high-resolution interfaces, \ie, precisely interacting with visual targets. This paper investigates the aforementioned challenges with our \textbf{G}UI \textbf{T}est-time Scaling \textbf{A}gent, namely \name. First, we conduct test-time scaling to select the most appropriate action proposal: at each step, multiple candidate proposals are sampled and evaluated and selected by a judge model. It trades off computation for better decision quality by concurrent sampling. Second, we propose a model that improves grounding of the selected action proposals to its corresponding visual elements. Our key insight is that reinforcement learning (RL) facilitates grounding through inherent objective alignments, rewarding successful clicks on interface elements. Experimentally, \name achieves state-of-the-art performance on both grounding and agent task execution benchmarks.

With the development of location-based social networks, more and more users post their check-ins on social sites. Matching accounts on different social sites based on users' spatiotemporal trajectories has become a hot issue in current research. Existing methods do not take into account the temporal characteristics, spatial characteristics, and sequential characteristics of the user's trajectory at the same time, resulting in the inability to obtain accurate trajectory features. To solve this problem, we propose a new user identification approach based on the spatiotemporal characteristics of user trajectories, which consists of three parts: 1) convert trajectories into graph structures that preserve both spatial connectivity and access order, then use a graph attention network (GAT) to extract spatial representations-enabling effective modeling of users' mobility structural patterns; 2) design a time-slice enhanced long short-term memory (LSTM) model to extract temporal periodicity and sequential features, which is robust to sparse trajectories with uneven time intervals; and 3) adopt a contrastive learning framework with a modified SimCLR to mine hard negative samples, solving the problem of positive-negative sample imbalance and enhancing the discriminability of trajectory representations. Extensive experiments on real datasets demonstrate that the proposed approach is superior to baseline methods in terms of performance and efficiency.

We introduce two multilingual, multimodal foundation language models that power Apple Intelligence features across Apple devices and services: i a 3B-parameter on-device model optimized for Apple silicon through architectural innovations such as KV-cache sharing and 2-bit quantization-aware training; and ii a scalable server model built on a novel Parallel-Track Mixture-of-Experts PT-MoE transformer that combines track parallelism, mixture-of-experts sparse computation, and interleaved global-local attention to deliver high quality with competitive cost on Apple's Private Cloud Compute platform. Both models are trained on large-scale multilingual and multimodal datasets sourced via responsible web crawling, licensed corpora, and high-quality synthetic data, then further refined with supervised fine-tuning and reinforcement learning on a new asynchronous platform. The resulting models support several additional languages while understanding images and executing tool calls. In public benchmarks and human evaluations, both the server model and the on-device model match or surpass comparably sized open baselines. A new Swift-centric Foundation Models framework exposes guided generation, constrained tool calling, and LoRA adapter fine-tuning, allowing developers to integrate these capabilities with a few lines of code. The latest advancements in Apple Intelligence models are grounded in our Responsible AI approach with safeguards like content filtering and locale-specific evaluation, as well as our commitment to protecting our users' privacy with innovations like Private Cloud Compute.

Large multimodal models (LMMs) with advanced video analysis capabilities have recently garnered significant attention. However, most evaluations rely on traditional methods like multiple-choice question answering in benchmarks such as VideoMME and LongVideoBench, which are prone to lack the depth needed to capture the complex demands of real-world users. To address this limitation-and due to the prohibitive cost and slow pace of human annotation for video tasks-we introduce VideoAutoArena, an arena-style benchmark inspired by LMSYS Chatbot Arena's framework, designed to automatically assess LMMs' video analysis abilities. VideoAutoArena utilizes user simulation to generate open-ended, adaptive questions that rigorously assess model performance in video understanding. The benchmark features an automated, scalable evaluation framework, incorporating a modified ELO Rating System for fair and continuous comparisons across multiple LMMs. To validate our automated judging system, we construct a "gold standard" using a carefully curated subset of human annotations, demonstrating that our arena strongly aligns with human judgment while maintaining scalability. Additionally, we introduce a fault-driven evolution strategy, progressively increasing question complexity to push models toward handling more challenging video analysis scenarios. Experimental results demonstrate that VideoAutoArena effectively differentiates among state-of-the-art LMMs, providing insights into model strengths and areas for improvement. To further streamline our evaluation, we introduce VideoAutoBench as an auxiliary benchmark, where human annotators label winners in a subset of VideoAutoArena battles. We use GPT-4o as a judge to compare responses against these human-validated answers. Together, VideoAutoArena and VideoAutoBench offer a cost-effective, and scalable framework for evaluating LMMs in user-centric video analysis.

Current video diffusion models (VDMs) mostly rely on text conditions, limiting control over video appearance and geometry. This study introduces a new model, MoonShot, conditioning on both image and text for enhanced control. It features the Multimodal Video Block (MVB), integrating the motion-aware dual cross-attention layer for precise appearance and motion alignment with provided prompts, and the spatiotemporal attention layer for large motion dynamics. It can also incorporate pre-trained Image ControlNet modules for geometry conditioning without extra video training. Experiments show our model significantly improves visual quality and motion fidelity, and its versatility allows for applications in personalized video generation, animation, and editing, making it a foundational tool for controllable video creation. More video results can be found here .

This paper studies zero-shot object recognition using event camera data. Guided by CLIP, which is pre-trained on RGB images, existing approaches achieve zero-shot object recognition by optimizing embedding similarities between event data and RGB images respectively encoded by an event encoder and the CLIP image encoder. Alternatively, several methods learn RGB frame reconstructions from event data for the CLIP image encoder. However, they often result in suboptimal zero-shot performance. This study develops an event encoder without relying on additional reconstruction networks. We theoretically analyze the performance bottlenecks of previous approaches: the embedding optimization objectives are prone to suffer from the spatial sparsity of event data, causing semantic misalignments between the learned event embedding space and the CLIP text embedding space. To mitigate the issue, we explore a scalar-wise modulation strategy. Furthermore, to scale up the number of events and RGB data pairs for training, we also study a pipeline for synthesizing event data from static RGB images in mass. Experimentally, we demonstrate an attractive scaling property in the number of parameters and synthesized data. We achieve superior zero-shot object recognition performance on extensive standard benchmark datasets, even compared with past supervised learning approaches. For example, our model with a ViT/B-16 backbone achieves 47.84% zero-shot accuracy on the N-ImageNet dataset.

Digital agents for automating tasks across different platforms by directly manipulating the GUIs are increasingly important. For these agents, grounding from language instructions to target elements remains a significant challenge due to reliance on HTML or AXTree inputs. In this paper, we introduce Aria-UI, a large multimodal model specifically designed for GUI grounding. Aria-UI adopts a pure-vision approach, eschewing reliance on auxiliary inputs. To adapt to heterogeneous planning instructions, we propose a scalable data pipeline that synthesizes diverse and high-quality instruction samples for grounding. To handle dynamic contexts in task performing, Aria-UI incorporates textual and text-image interleaved action histories, enabling robust context-aware reasoning for grounding. Aria-UI sets new state-of-the-art results across offline and online agent benchmarks, outperforming both vision-only and AXTree-reliant baselines. We release all training data and model checkpoints to foster further research at https://ariaui.github.io.

Solving expert-level multimodal tasks is a key milestone towards general intelligence. As the capabilities of multimodal large language models (MLLMs) continue to improve, evaluation of such advanced multimodal intelligence becomes necessary yet challenging. In this work, we introduce ProBench, a benchmark of open-ended user queries that require professional expertise and advanced reasoning. ProBench consists of 4,000 high-quality samples independently submitted by professionals based on their daily productivity demands. It spans across 10 fields and 56 sub-fields, including science, arts, humanities, coding, mathematics, and creative writing. Experimentally, we evaluate and compare 24 latest models using MLLM-as-a-Judge. Our results reveal that although the best open-source models rival the proprietary ones, ProBench presents significant challenges in visual perception, textual understanding, domain knowledge and advanced reasoning, thus providing valuable directions for future multimodal AI research efforts.

Recent research has achieved significant advancements in visual reasoning tasks through learning image-to-language projections and leveraging the impressive reasoning abilities of Large Language Models (LLMs). This paper introduces an efficient and effective framework that integrates multiple modalities (images, 3D, audio and video) to a frozen LLM and demonstrates an emergent ability for cross-modal reasoning (2+ modality inputs). Our approach explores two distinct projection mechanisms: Q-Formers and Linear Projections (LPs). Through extensive experimentation across all four modalities on 16 benchmarks, we explore both methods and assess their adaptability in integrated and separate cross-modal reasoning. The Q-Former projection demonstrates superior performance in single modality scenarios and adaptability in joint versus discriminative reasoning involving two or more modalities. However, it exhibits lower generalization capabilities than linear projection in contexts where task-modality data are limited. To enable this framework, we devise a scalable pipeline that automatically generates high-quality, instruction-tuning datasets from readily available captioning data across different modalities, and contribute 24K QA data for audio and 250K QA data for 3D. To facilitate further research in cross-modal reasoning, we introduce the DisCRn (Discriminative Cross-modal Reasoning (DisCRn)) benchmark comprising 9K audio-video QA samples and 28K image-3D QA samples that require the model to reason discriminatively across disparate input modalities. Code and data is available at https://github.com/salesforce/LAVIS/tree/main/projects/xinstructblip.

Information comes in diverse modalities. Multimodal native AI models are essential to integrate real-world information and deliver comprehensive understanding. While proprietary multimodal native models exist, their lack of openness imposes obstacles for adoptions, let alone adaptations. To fill this gap, we introduce Aria, an open multimodal native model with best-in-class performance across a wide range of multimodal, language, and coding tasks. Aria is a mixture-of-expert model with 3.9B and 3.5B activated parameters per visual token and text token, respectively. It outperforms Pixtral-12B and Llama3.2-11B, and is competitive against the best proprietary models on various multimodal tasks. We pre-train Aria from scratch following a 4-stage pipeline, which progressively equips the model with strong capabilities in language understanding, multimodal understanding, long context window, and instruction following. We open-source the model weights along with a codebase that facilitates easy adoptions and adaptations of Aria in real-world applications.

Vision-language pre-training and instruction tuning have demonstrated general-purpose capabilities in 2D visual reasoning tasks by aligning visual encoders with state-of-the-art large language models (LLMs). In this paper, we introduce a simple, yet effective, cross-modality framework built atop frozen LLMs that allows the integration of various modalities without extensive modality-specific customization. To facilitate instruction-modality fine-tuning, we collect high-quality instruction tuning data in an automatic and scalable manner, composed of 24K QA samples for audio and 250K QA samples for 3D. Leveraging instruction-aware representations, our model performs comparably with leading-edge counterparts without the need of extensive modality-specific pre-training or customization. Furthermore, our approach demonstrates cross-modal reasoning abilities across two or more input modalities, despite each modality projection being trained individually. To study the model's cross-modal abilities, we contribute a novel Discriminative Cross-modal Reasoning (DisCRn) evaluation task, comprising 9K audio-video QA samples and 28K image-3D QA samples that require the model to reason discriminatively across disparate input modalities.

Relative positional encoding is widely used in vanilla and linear transformers to represent positional information. However, existing encoding methods of a vanilla transformer are not always directly applicable to a linear transformer, because the latter requires a decomposition of the query and key representations into separate kernel functions. Nevertheless, principles for designing encoding methods suitable for linear transformers remain understudied. In this work, we put together a variety of existing linear relative positional encoding approaches under a canonical form and further propose a family of linear relative positional encoding algorithms via unitary transformation. Our formulation leads to a principled framework that can be used to develop new relative positional encoding methods that preserve linear space-time complexity. Equipped with different models, the proposed linearized relative positional encoding (LRPE) family derives effective encoding for various applications. Experiments show that compared with existing methods, LRPE achieves state-of-the-art performance in language modeling, text classification, and image classification. Meanwhile, it emphasizes a general paradigm for designing broadly more relative positional encoding methods that are applicable to linear transformers. The code is available at https://github.com/OpenNLPLab/Lrpe.

Large multimodal models (LMMs) are processing increasingly longer and richer inputs. Albeit the progress, few public benchmark is available to measure such development. To mitigate this gap, we introduce LongVideoBench, a question-answering benchmark that features video-language interleaved inputs up to an hour long. Our benchmark includes 3,763 varying-length web-collected videos with their subtitles across diverse themes, designed to comprehensively evaluate LMMs on long-term multimodal understanding. To achieve this, we interpret the primary challenge as to accurately retrieve and reason over detailed multimodal information from long inputs. As such, we formulate a novel video question-answering task termed referring reasoning. Specifically, as part of the question, it contains a referring query that references related video contexts, called referred context. The model is then required to reason over relevant video details from the referred context. Following the paradigm of referring reasoning, we curate 6,678 human-annotated multiple-choice questions in 17 fine-grained categories, establishing one of the most comprehensive benchmarks for long-form video understanding. Evaluations suggest that the LongVideoBench presents significant challenges even for the most advanced proprietary models (e.g. GPT-4o, Gemini-1.5-Pro, GPT-4-Turbo), while their open-source counterparts show an even larger performance gap. In addition, our results indicate that model performance on the benchmark improves only when they are capable of processing more frames, positioning LongVideoBench as a valuable benchmark for evaluating future-generation long-context LMMs.

Sequence modeling has important applications in natural language processing and computer vision. Recently, the transformer-based models have shown strong performance on various sequence modeling tasks, which rely on attention to capture pairwise token relations, and position embedding to inject positional information. While showing good performance, the transformer models are inefficient to scale to long input sequences, mainly due to the quadratic space-time complexity of attention. To overcome this inefficiency, we propose to model sequences with a relative position encoded Toeplitz matrix and use a Toeplitz matrix-vector production trick to reduce the space-time complexity of the sequence modeling to log linear. A lightweight sub-network called relative position encoder is proposed to generate relative position coefficients with a fixed budget of parameters, enabling the proposed Toeplitz neural network to deal with varying sequence lengths. In addition, despite being trained on 512-token sequences, our model can extrapolate input sequence length up to 14K tokens in inference with consistent performance. Extensive experiments on autoregressive and bidirectional language modeling, image modeling, and the challenging Long-Range Arena benchmark show that our method achieves better performance than its competitors in most downstream tasks while being significantly faster. The code is available at https://github.com/OpenNLPLab/Tnn.

Relative position encoding is widely used in vanilla and linear transformers to represent positional information. However, the existing encoding methods of a vanilla transformer are not always directly applicable to a linear transformer, because the latter requires a decomposition of the query and key representations into separate kernel functions. Nevertheless, principles to design encoding methods suitable for linear transformers remain under-studied. In this work, we put together a variety of existing encoding methods under a canonical form and further propose a family of relative positional encodings via unitary transformation. Our formulation leads to a principled framework that can be used to develop new relative positional encoding methods that preserve linear space-time complexity. Equipping with different parameters, the proposed unitary relative positional encoding family (URPE) derives effective encoding for various applications. Experiments show that compared with existing encoding methods, unitary encoding achieves competitive performance on language modeling and various challenging downstream tasks, such as machine translation and text classification. In the meantime, it highlights a general paradigm to design broadly more relative positional encoding methods, applicable inclusively to linear and vanilla transformers.

We introduce LAVIS, an open-source deep learning library for LAnguage-VISion research and applications. LAVIS aims to serve as a one-stop comprehensive library that brings recent advancements in the language-vision field accessible for researchers and practitioners, as well as fertilizing future research and development. It features a unified interface to easily access state-of-the-art image-language, video-language models and common datasets. LAVIS supports training, evaluation and benchmarking on a rich variety of tasks, including multimodal classification, retrieval, captioning, visual question answering, dialogue and pre-training. In the meantime, the library is also highly extensible and configurable, facilitating future development and customization. In this technical report, we describe design principles, key components and functionalities of the library, and also present benchmarking results across common language-vision tasks. The library is available at: https://github.com/salesforce/LAVIS.

The cost of vision-and-language pre-training has become increasingly prohibitive due to end-to-end training of large-scale models. This paper proposes BLIP-2, a generic and efficient pre-training strategy that bootstraps vision-language pre-training from off-the-shelf frozen pre-trained image encoders and frozen large language models. BLIP-2 bridges the modality gap with a lightweight Querying Transformer, which is pre-trained in two stages. The first stage bootstraps vision-language representation learning from a frozen image encoder. The second stage bootstraps vision-to-language generative learning from a frozen language model. BLIP-2 achieves state-of-the-art performance on various vision-language tasks, despite having significantly fewer trainable parameters than existing methods. For example, our model outperforms Flamingo80B by 8.7% on zero-shot VQAv2 with 54x fewer trainable parameters. We also demonstrate the model's emerging capabilities of zero-shot image-to-text generation that can follow natural language instructions.

Large language models (LLMs) have demonstrated excellent zero-shot generalization to new language tasks. However, effective utilization of LLMs for zero-shot visual question-answering (VQA) remains challenging, primarily due to the modality disconnect and task disconnect between the LLM and VQA tasks. End-to-end training on multimodal data may bridge the disconnects, but is inflexible and computationally expensive. To address this issue, we propose Img2LLM, a plug-and-play module that provides LLM prompts to enable LLMs to perform zero-shot VQA tasks without end-to-end training. We develop LLM-agnostic models describe image content as exemplar question-answer pairs, which prove to be effective LLM prompts. Img2LLM offers the following benefits: 1) It achieves comparable or better performance than methods relying on end-to-end training. For example, we outperform Flamingo [3] by 5.6% on VQAv2. On the challenging A-OKVQA dataset, our method outperforms few-shot methods by as much as 20%. 2) It flexibly interfaces with a wide range of LLMs to perform VQA. 3) It eliminates the need to specialize LLMs using end-to-end finetuning and serve highly specialized LLMs to end users, thereby reducing cost. Code is available via the LAVIS [28] framework at https://github.com/salesforce/LAVIS/ tree/main/projects/img2llm-vqa.