Lijuan Wang

Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu et al. · microsoft-research, uw

In this paper, we design and train a Generative Image-to-text Transformer, GIT, to unify vision-language tasks such as image/video captioning and question answering. While generative models provide a consistent network architecture between pre-training and fine-tuning, existing work typically contains complex structures (uni/multi-modal encoder/decoder) and depends on external modules such as object detectors/taggers and optical character recognition (OCR). In GIT, we simplify the architecture as one image encoder and one text decoder under a single language modeling task. We also scale up the pre-training data and the model size to boost the model performance. Without bells and whistles, our GIT establishes new state of the arts on 12 challenging benchmarks with a large margin. For instance, our model surpasses the human performance for the first time on TextCaps (138.2 vs. 125.5 in CIDEr). Furthermore, we present a new scheme of generation-based image classification and scene text recognition, achieving decent performance on standard benchmarks. Codes are released at \url{https://github.com/microsoft/GenerativeImage2Text}.

Zi-Yi Dou, Aishwarya Kamath, Zhe Gan et al. · microsoft-research

Vision-language (VL) pre-training has recently received considerable attention. However, most existing end-to-end pre-training approaches either only aim to tackle VL tasks such as image-text retrieval, visual question answering (VQA) and image captioning that test high-level understanding of images, or only target region-level understanding for tasks such as phrase grounding and object detection. We present FIBER (Fusion-In-the-Backbone-based transformER), a new VL model architecture that can seamlessly handle both these types of tasks. Instead of having dedicated transformer layers for fusion after the uni-modal backbones, FIBER pushes multimodal fusion deep into the model by inserting cross-attention into the image and text backbones, bringing gains in terms of memory and performance. In addition, unlike previous work that is either only pre-trained on image-text data or on fine-grained data with box-level annotations, we present a two-stage pre-training strategy that uses both these kinds of data efficiently: (i) coarse-grained pre-training based on image-text data; followed by (ii) fine-grained pre-training based on image-text-box data. We conduct comprehensive experiments on a wide range of VL tasks, ranging from VQA, image captioning, and retrieval, to phrase grounding, referring expression comprehension, and object detection. Using deep multimodal fusion coupled with the two-stage pre-training, FIBER provides consistent performance improvements over strong baselines across all tasks, often outperforming methods using magnitudes more data. Code is available at https://github.com/microsoft/FIBER.

Haotian Zhang, Pengchuan Zhang, Xiaowei Hu et al. · microsoft-research

We present GLIPv2, a grounded VL understanding model, that serves both localization tasks (e.g., object detection, instance segmentation) and Vision-Language (VL) understanding tasks (e.g., VQA, image captioning). GLIPv2 elegantly unifies localization pre-training and Vision-Language Pre-training (VLP) with three pre-training tasks: phrase grounding as a VL reformulation of the detection task, region-word contrastive learning as a novel region-word level contrastive learning task, and the masked language modeling. This unification not only simplifies the previous multi-stage VLP procedure but also achieves mutual benefits between localization and understanding tasks. Experimental results show that a single GLIPv2 model (all model weights are shared) achieves near SoTA performance on various localization and understanding tasks. The model also shows (1) strong zero-shot and few-shot adaption performance on open-vocabulary object detection tasks and (2) superior grounding capability on VL understanding tasks. Code will be released at https://github.com/microsoft/GLIP.

Sheng Shen, Chunyuan Li, Xiaowei Hu et al. · berkeley, gatech

The new generation of state-of-the-art computer vision systems are trained from natural language supervision, ranging from simple object category names to descriptive captions. This form of supervision ensures high generality and usability of the learned visual models, due to the broad concept coverage achieved via large-scale data collection process. Alternatively, we argue that learning with external knowledge is a promising way which leverages a much more structured source of supervision and offers sample efficiency. We propose K-LITE, a simple strategy to leverage external knowledge for building transferable visual systems: In training, it enriches entities in text with WordNet and Wiktionary knowledge, leading to an efficient and scalable approach to learning image representations that uses knowledge about the visual concepts. In evaluation, the text is also augmented with external knowledge and then used to reference learned visual concepts (or describe new ones) to enable zero-shot and few-shot transfer of the pre-trained models. We study the performance of K-LITE on two important computer vision problems, image classification and object detection, benchmarking on 20 and 13 different existing datasets, respectively. The proposed knowledge-augmented models show significant improvement in transfer learning performance over existing methods. Our code is available at https://github.com/microsoft/klite.

Fuxiao Liu, Kevin Lin, Linjie Li et al. · microsoft-research, uw

Despite the promising progress in multi-modal tasks, current large multi-modal models (LMMs) are prone to hallucinating inconsistent descriptions with respect to the associated image and human instructions. This paper addresses this issue by introducing the first large and diverse visual instruction tuning dataset, named Large-scale Robust Visual (LRV)-Instruction. Our dataset comprises 400k visual instructions generated by GPT4, covering 16 vision-and-language tasks with open-ended instructions and answers. Unlike existing studies that primarily focus on positive instruction samples, we design LRV-Instruction to include both positive and negative instructions for more robust visual instruction tuning. Our negative instructions are designed at three semantic levels: (i) Nonexistent Object Manipulation, (ii) Existent Object Manipulation and (iii) Knowledge Manipulation. To efficiently measure the hallucination generated by LMMs, we propose GPT4-Assisted Visual Instruction Evaluation (GAVIE), a stable approach to evaluate visual instruction tuning like human experts. GAVIE does not require human-annotated groundtruth answers and can adapt to diverse instruction formats. We conduct comprehensive experiments to investigate the hallucination of LMMs. Our results demonstrate existing LMMs exhibit significant hallucinations when presented with our negative instructions, particularly Existent Object and Knowledge Manipulation instructions. Moreover, we successfully mitigate hallucination by finetuning MiniGPT4 and mPLUG-Owl on LRV-Instruction while improving performance on several public datasets compared to state-of-the-art methods. Additionally, we observed that a balanced ratio of positive and negative instances in the training data leads to a more robust model. Code and data are available at https://github.com/FuxiaoLiu/LRV-Instruction.

Jialian Wu, Jianfeng Wang, Zhengyuan Yang et al. · microsoft-research

This paper presents a Generative RegIon-to-Text transformer, GRiT, for object understanding. The spirit of GRiT is to formulate object understanding as <region, text> pairs, where region locates objects and text describes objects. For example, the text in object detection denotes class names while that in dense captioning refers to descriptive sentences. Specifically, GRiT consists of a visual encoder to extract image features, a foreground object extractor to localize objects, and a text decoder to generate open-set object descriptions. With the same model architecture, GRiT can understand objects via not only simple nouns, but also rich descriptive sentences including object attributes or actions. Experimentally, we apply GRiT to object detection and dense captioning tasks. GRiT achieves 60.4 AP on COCO 2017 test-dev for object detection and 15.5 mAP on Visual Genome for dense captioning. Code is available at https://github.com/JialianW/GRiT

Chenfei Wu, Jian Liang, Xiaowei Hu et al. · microsoft-research

In this paper, we present NUWA-Infinity, a generative model for infinite visual synthesis, which is defined as the task of generating arbitrarily-sized high-resolution images or long-duration videos. An autoregressive over autoregressive generation mechanism is proposed to deal with this variable-size generation task, where a global patch-level autoregressive model considers the dependencies between patches, and a local token-level autoregressive model considers dependencies between visual tokens within each patch. A Nearby Context Pool (NCP) is introduced to cache-related patches already generated as the context for the current patch being generated, which can significantly save computation costs without sacrificing patch-level dependency modeling. An Arbitrary Direction Controller (ADC) is used to decide suitable generation orders for different visual synthesis tasks and learn order-aware positional embeddings. Compared to DALL-E, Imagen and Parti, NUWA-Infinity can generate high-resolution images with arbitrary sizes and support long-duration video generation additionally. Compared to NUWA, which also covers images and videos, NUWA-Infinity has superior visual synthesis capabilities in terms of resolution and variable-size generation. The GitHub link is https://github.com/microsoft/NUWA. The homepage link is https://nuwa-infinity.microsoft.com.

Linjie Li, Zhe Gan, Kevin Lin et al. · microsoft-research, uw

Unified vision-language frameworks have greatly advanced in recent years, most of which adopt an encoder-decoder architecture to unify image-text tasks as sequence-to-sequence generation. However, existing video-language (VidL) models still require task-specific designs in model architecture and training objectives for each task. In this work, we explore a unified VidL framework LAVENDER, where Masked Language Modeling (MLM) is used as the common interface for all pre-training and downstream tasks. Such unification leads to a simplified model architecture, where only a lightweight MLM head, instead of a decoder with much more parameters, is needed on top of the multimodal encoder. Surprisingly, experimental results show that this unified framework achieves competitive performance on 14 VidL benchmarks, covering video question answering, text-to-video retrieval and video captioning. Extensive analyses further demonstrate the advantage of LAVENDER over existing VidL methods in: (i) supporting all downstream tasks with just a single set of parameter values when multi-task finetuned; (ii) few-shot generalization on various downstream tasks; and (iii) enabling zero-shot evaluation on video question answering tasks. Code is available at https://github.com/microsoft/LAVENDER.

Yi-Lin Sung, Linjie Li, Kevin Lin et al. · microsoft-research, uw

Model merging (e.g., via interpolation or task arithmetic) fuses multiple models trained on different tasks to generate a multi-task solution. The technique has been proven successful in previous studies, where the models are trained on similar tasks and with the same initialization. In this paper, we expand on this concept to a multimodal setup by merging transformers trained on different modalities. Furthermore, we conduct our study for a novel goal where we can merge vision, language, and cross-modal transformers of a modality-specific architecture to create a parameter-efficient modality-agnostic architecture. Through comprehensive experiments, we systematically investigate the key factors impacting model performance after merging, including initialization, merging mechanisms, and model architectures. We also propose two metrics that assess the distance between weights to be merged and can serve as an indicator of the merging outcomes. Our analysis leads to an effective training recipe for matching the performance of the modality-agnostic baseline (i.e., pre-trained from scratch) via model merging. Our method also outperforms naive merging significantly on various tasks, with improvements of 3% on VQA, 7% on COCO retrieval, 25% on NLVR2, 14% on Flickr30k and 3% on ADE20k. Our code is available at https://github.com/ylsung/vl-merging

An Yan, Zhengyuan Yang, Wanrong Zhu et al. · microsoft-research

We present MM-Navigator, a GPT-4V-based agent for the smartphone graphical user interface (GUI) navigation task. MM-Navigator can interact with a smartphone screen as human users, and determine subsequent actions to fulfill given instructions. Our findings demonstrate that large multimodal models (LMMs), specifically GPT-4V, excel in zero-shot GUI navigation through its advanced screen interpretation, action reasoning, and precise action localization capabilities. We first benchmark MM-Navigator on our collected iOS screen dataset. According to human assessments, the system exhibited a 91\% accuracy rate in generating reasonable action descriptions and a 75\% accuracy rate in executing the correct actions for single-step instructions on iOS. Additionally, we evaluate the model on a subset of an Android screen navigation dataset, where the model outperforms previous GUI navigators in a zero-shot fashion. Our benchmark and detailed analyses aim to lay a robust groundwork for future research into the GUI navigation task. The project page is at https://github.com/zzxslp/MM-Navigator.

Tan Wang, Linjie Li, Kevin Lin et al. · microsoft-research, uw

Generative AI has made significant strides in computer vision, particularly in text-driven image/video synthesis (T2I/T2V). Despite the notable advancements, it remains challenging in human-centric content synthesis such as realistic dance generation. Current methodologies, primarily tailored for human motion transfer, encounter difficulties when confronted with real-world dance scenarios (e.g., social media dance), which require to generalize across a wide spectrum of poses and intricate human details. In this paper, we depart from the traditional paradigm of human motion transfer and emphasize two additional critical attributes for the synthesis of human dance content in social media contexts: (i) Generalizability: the model should be able to generalize beyond generic human viewpoints as well as unseen human subjects, backgrounds, and poses; (ii) Compositionality: it should allow for the seamless composition of seen/unseen subjects, backgrounds, and poses from different sources. To address these challenges, we introduce DISCO, which includes a novel model architecture with disentangled control to improve the compositionality of dance synthesis, and an effective human attribute pre-training for better generalizability to unseen humans. Extensive qualitative and quantitative results demonstrate that DisCc can generate high-quality human dance images and videos with diverse appearances and flexible motions. Code is available at https://disco-dance.github.io/.

Zixin Zhu, Yixuan Wei, Jianfeng Wang et al. · microsoft-research

The image captioning task is typically realized by an auto-regressive method that decodes the text tokens one by one. We present a diffusion-based captioning model, dubbed the name DDCap, to allow more decoding flexibility. Unlike image generation, where the output is continuous and redundant with a fixed length, texts in image captions are categorical and short with varied lengths. Therefore, naively applying the discrete diffusion model to text decoding does not work well, as shown in our experiments. To address the performance gap, we propose several key techniques including best-first inference, concentrated attention mask, text length prediction, and image-free training. On COCO without additional caption pre-training, it achieves a CIDEr score of 117.8, which is +5.0 higher than the auto-regressive baseline with the same architecture in the controlled setting. It also performs +26.8 higher CIDEr score than the auto-regressive baseline (230.3 v.s.203.5) on a caption infilling task. With 4M vision-language pre-training images and the base-sized model, we reach a CIDEr score of 125.1 on COCO, which is competitive to the best well-developed auto-regressive frameworks. The code is available at https://github.com/buxiangzhiren/DDCap.

Tan Wang, Kevin Lin, Linjie Li et al. · microsoft-research, uw

This study explores the concept of equivariance in vision-language foundation models (VLMs), focusing specifically on the multimodal similarity function that is not only the major training objective but also the core delivery to support downstream tasks. Unlike the existing image-text similarity objective which only categorizes matched pairs as similar and unmatched pairs as dissimilar, equivariance also requires similarity to vary faithfully according to the semantic changes. This allows VLMs to generalize better to nuanced and unseen multimodal compositions. However, modeling equivariance is challenging as the ground truth of semantic change is difficult to collect. For example, given an image-text pair about a dog, it is unclear to what extent the similarity changes when the pixel is changed from dog to cat? To this end, we propose EqSim, a regularization loss that can be efficiently calculated from any two matched training pairs and easily pluggable into existing image-text retrieval fine-tuning. Meanwhile, to further diagnose the equivariance of VLMs, we present a new challenging benchmark EqBen. Compared to the existing evaluation sets, EqBen is the first to focus on "visual-minimal change". Extensive experiments show the lack of equivariance in current VLMs and validate the effectiveness of EqSim. Code is available at https://github.com/Wangt-CN/EqBen.

Jielin Qiu, Jiacheng Zhu, William Han et al. · cmu, microsoft-research

Multimodal summarization with multimodal output (MSMO) has emerged as a promising research direction. Nonetheless, numerous limitations exist within existing public MSMO datasets, including insufficient maintenance, data inaccessibility, limited size, and the absence of proper categorization, which pose significant challenges. To address these challenges and provide a comprehensive dataset for this new direction, we have meticulously curated the \textbf{MMSum} dataset. Our new dataset features (1) Human-validated summaries for both video and textual content, providing superior human instruction and labels for multimodal learning. (2) Comprehensively and meticulously arranged categorization, spanning 17 principal categories and 170 subcategories to encapsulate a diverse array of real-world scenarios. (3) Benchmark tests performed on the proposed dataset to assess various tasks and methods, including \textit{video summarization}, \textit{text summarization}, and \textit{multimodal summarization}. To champion accessibility and collaboration, we will release the \textbf{MMSum} dataset and the data collection tool as fully open-source resources, fostering transparency and accelerating future developments. Our project website can be found at~\url{https://mmsum-dataset.github.io/}

Chung-Ching Lin, Jiang Wang, Kun Luo et al. · microsoft-research, uw

The most recent efforts in video matting have focused on eliminating trimap dependency since trimap annotations are expensive and trimap-based methods are less adaptable for real-time applications. Despite the latest tripmap-free methods showing promising results, their performance often degrades when dealing with highly diverse and unstructured videos. We address this limitation by introducing Adaptive Matting for Dynamic Videos, termed AdaM, which is a framework designed for simultaneously differentiating foregrounds from backgrounds and capturing alpha matte details of human subjects in the foreground. Two interconnected network designs are employed to achieve this goal: (1) an encoder-decoder network that produces alpha mattes and intermediate masks which are used to guide the transformer in adaptively decoding foregrounds and backgrounds, and (2) a transformer network in which long- and short-term attention combine to retain spatial and temporal contexts, facilitating the decoding of foreground details. We benchmark and study our methods on recently introduced datasets, showing that our model notably improves matting realism and temporal coherence in complex real-world videos and achieves new best-in-class generalizability. Further details and examples are available at https://github.com/microsoft/AdaM.

Chenglei Si, Zhe Gan, Zhengyuan Yang et al. · microsoft-research

Large language models (LLMs) show impressive abilities via few-shot prompting. Commercialized APIs such as OpenAI GPT-3 further increase their use in real-world language applications. However, the crucial problem of how to improve the reliability of GPT-3 is still under-explored. While reliability is a broad and vaguely defined term, we decompose reliability into four main facets that correspond to the existing framework of ML safety and are well-recognized to be important: generalizability, social biases, calibration, and factuality. Our core contribution is to establish simple and effective prompts that improve GPT-3's reliability as it: 1) generalizes out-of-distribution, 2) balances demographic distribution and uses natural language instructions to reduce social biases, 3) calibrates output probabilities, and 4) updates the LLM's factual knowledge and reasoning chains. With appropriate prompts, GPT-3 is more reliable than smaller-scale supervised models on all these facets. We release all processed datasets, evaluation scripts, and model predictions. Our systematic empirical study not only sheds new insights on the reliability of prompting LLMs, but more importantly, our prompting strategies can help practitioners more reliably use LLMs like GPT-3.

Xueyan Zou, Zi-Yi Dou, Jianwei Yang et al. · microsoft-research

We present X-Decoder, a generalized decoding model that can predict pixel-level segmentation and language tokens seamlessly. X-Decodert takes as input two types of queries: (i) generic non-semantic queries and (ii) semantic queries induced from text inputs, to decode different pixel-level and token-level outputs in the same semantic space. With such a novel design, X-Decoder is the first work that provides a unified way to support all types of image segmentation and a variety of vision-language (VL) tasks. Further, our design enables seamless interactions across tasks at different granularities and brings mutual benefits by learning a common and rich pixel-level visual-semantic understanding space, without any pseudo-labeling. After pretraining on a mixed set of a limited amount of segmentation data and millions of image-text pairs, X-Decoder exhibits strong transferability to a wide range of downstream tasks in both zero-shot and finetuning settings. Notably, it achieves (1) state-of-the-art results on open-vocabulary segmentation and referring segmentation on eight datasets; (2) better or competitive finetuned performance to other generalist and specialist models on segmentation and VL tasks; and (3) flexibility for efficient finetuning and novel task composition (e.g., referring captioning and image editing). Code, demo, video, and visualization are available at https://x-decoder-vl.github.io.

Zhengyuan Yang, Linjie Li, Jianfeng Wang et al. · microsoft-research, uw

We propose MM-REACT, a system paradigm that integrates ChatGPT with a pool of vision experts to achieve multimodal reasoning and action. In this paper, we define and explore a comprehensive list of advanced vision tasks that are intriguing to solve, but may exceed the capabilities of existing vision and vision-language models. To achieve such advanced visual intelligence, MM-REACT introduces a textual prompt design that can represent text descriptions, textualized spatial coordinates, and aligned file names for dense visual signals such as images and videos. MM-REACT's prompt design allows language models to accept, associate, and process multimodal information, thereby facilitating the synergetic combination of ChatGPT and various vision experts. Zero-shot experiments demonstrate MM-REACT's effectiveness in addressing the specified capabilities of interests and its wide application in different scenarios that require advanced visual understanding. Furthermore, we discuss and compare MM-REACT's system paradigm with an alternative approach that extends language models for multimodal scenarios through joint finetuning. Code, demo, video, and visualization are available at https://multimodal-react.github.io/

Zhe Gan, Linjie Li, Chunyuan Li et al. · microsoft-research

This paper surveys vision-language pre-training (VLP) methods for multimodal intelligence that have been developed in the last few years. We group these approaches into three categories: ($i$) VLP for image-text tasks, such as image captioning, image-text retrieval, visual question answering, and visual grounding; ($ii$) VLP for core computer vision tasks, such as (open-set) image classification, object detection, and segmentation; and ($iii$) VLP for video-text tasks, such as video captioning, video-text retrieval, and video question answering. For each category, we present a comprehensive review of state-of-the-art methods, and discuss the progress that has been made and challenges still being faced, using specific systems and models as case studies. In addition, for each category, we discuss advanced topics being actively explored in the research community, such as big foundation models, unified modeling, in-context few-shot learning, knowledge, robustness, and computer vision in the wild, to name a few.

Zhengyuan Yang, Jianfeng Wang, Zhe Gan et al. · microsoft-research, uw

Recently, large-scale text-to-image (T2I) models have shown impressive performance in generating high-fidelity images, but with limited controllability, e.g., precisely specifying the content in a specific region with a free-form text description. In this paper, we propose an effective technique for such regional control in T2I generation. We augment T2I models' inputs with an extra set of position tokens, which represent the quantized spatial coordinates. Each region is specified by four position tokens to represent the top-left and bottom-right corners, followed by an open-ended natural language regional description. Then, we fine-tune a pre-trained T2I model with such new input interface. Our model, dubbed as ReCo (Region-Controlled T2I), enables the region control for arbitrary objects described by open-ended regional texts rather than by object labels from a constrained category set. Empirically, ReCo achieves better image quality than the T2I model strengthened by positional words (FID: 8.82->7.36, SceneFID: 15.54->6.51 on COCO), together with objects being more accurately placed, amounting to a 20.40% region classification accuracy improvement on COCO. Furthermore, we demonstrate that ReCo can better control the object count, spatial relationship, and region attributes such as color/size, with the free-form regional description. Human evaluation on PaintSkill shows that ReCo is +19.28% and +17.21% more accurate in generating images with correct object count and spatial relationship than the T2I model.

Shengming Yin, Chenfei Wu, Huan Yang et al. · microsoft-research, pku

In this paper, we propose NUWA-XL, a novel Diffusion over Diffusion architecture for eXtremely Long video generation. Most current work generates long videos segment by segment sequentially, which normally leads to the gap between training on short videos and inferring long videos, and the sequential generation is inefficient. Instead, our approach adopts a ``coarse-to-fine'' process, in which the video can be generated in parallel at the same granularity. A global diffusion model is applied to generate the keyframes across the entire time range, and then local diffusion models recursively fill in the content between nearby frames. This simple yet effective strategy allows us to directly train on long videos (3376 frames) to reduce the training-inference gap, and makes it possible to generate all segments in parallel. To evaluate our model, we build FlintstonesHD dataset, a new benchmark for long video generation. Experiments show that our model not only generates high-quality long videos with both global and local coherence, but also decreases the average inference time from 7.55min to 26s (by 94.26\%) at the same hardware setting when generating 1024 frames. The homepage link is \url{https://msra-nuwa.azurewebsites.net/}

Tsu-Jui Fu, Linjie Li, Zhe Gan et al. · microsoft-research, uw

Masked visual modeling (MVM) has been recently proven effective for visual pre-training. While similar reconstructive objectives on video inputs (e.g., masked frame modeling) have been explored in video-language (VidL) pre-training, previous studies fail to find a truly effective MVM strategy that can largely benefit the downstream performance. In this work, we systematically examine the potential of MVM in the context of VidL learning. Specifically, we base our study on a fully end-to-end VIdeO-LanguagE Transformer (VIOLET), where the supervision from MVM training can be backpropagated to the video pixel space. In total, eight different reconstructive targets of MVM are explored, from low-level pixel values and oriented gradients to high-level depth maps, optical flow, discrete visual tokens, and latent visual features. We conduct comprehensive experiments and provide insights into the factors leading to effective MVM training, resulting in an enhanced model VIOLETv2. Empirically, we show VIOLETv2 pre-trained with MVM objective achieves notable improvements on 13 VidL benchmarks, ranging from video question answering, video captioning, to text-to-video retrieval.

Jinghao Zhou, Li Dong, Zhe Gan et al. · microsoft-research

Contrastive language-image pre-training (CLIP) serves as a de-facto standard to align images and texts. Nonetheless, the loose correlation between images and texts of web-crawled data renders the contrastive objective data inefficient and craving for a large training batch size. In this work, we explore the validity of non-contrastive language-image pre-training (nCLIP), and study whether nice properties exhibited in visual self-supervised models can emerge. We empirically observe that the non-contrastive objective nourishes representation learning while sufficiently underperforming under zero-shot recognition. Based on the above study, we further introduce xCLIP, a multi-tasking framework combining CLIP and nCLIP, and show that nCLIP aids CLIP in enhancing feature semantics. The synergy between two objectives lets xCLIP enjoy the best of both worlds: superior performance in both zero-shot transfer and representation learning. Systematic evaluation is conducted spanning a wide variety of downstream tasks including zero-shot classification, out-of-domain classification, retrieval, visual representation learning, and textual representation learning, showcasing a consistent performance gain and validating the effectiveness of xCLIP.

Kevin Lin, Faisal Ahmed, Linjie Li et al. · microsoft-research, uw

We present MM-VID, an integrated system that harnesses the capabilities of GPT-4V, combined with specialized tools in vision, audio, and speech, to facilitate advanced video understanding. MM-VID is designed to address the challenges posed by long-form videos and intricate tasks such as reasoning within hour-long content and grasping storylines spanning multiple episodes. MM-VID uses a video-to-script generation with GPT-4V to transcribe multimodal elements into a long textual script. The generated script details character movements, actions, expressions, and dialogues, paving the way for large language models (LLMs) to achieve video understanding. This enables advanced capabilities, including audio description, character identification, and multimodal high-level comprehension. Experimental results demonstrate the effectiveness of MM-VID in handling distinct video genres with various video lengths. Additionally, we showcase its potential when applied to interactive environments, such as video games and graphic user interfaces.

Jaemin Cho, Linjie Li, Zhengyuan Yang et al. · allen-ai, microsoft-research

Spatial control is a core capability in controllable image generation. Advancements in layout-guided image generation have shown promising results on in-distribution (ID) datasets with similar spatial configurations. However, it is unclear how these models perform when facing out-of-distribution (OOD) samples with arbitrary, unseen layouts. In this paper, we propose LayoutBench, a diagnostic benchmark for layout-guided image generation that examines four categories of spatial control skills: number, position, size, and shape. We benchmark two recent representative layout-guided image generation methods and observe that the good ID layout control may not generalize well to arbitrary layouts in the wild (e.g., objects at the boundary). Next, we propose IterInpaint, a new baseline that generates foreground and background regions step-by-step via inpainting, demonstrating stronger generalizability than existing models on OOD layouts in LayoutBench. We perform quantitative and qualitative evaluation and fine-grained analysis on the four LayoutBench skills to pinpoint the weaknesses of existing models. We show comprehensive ablation studies on IterInpaint, including training task ratio, crop&paste vs. repaint, and generation order. Lastly, we evaluate the zero-shot performance of different pretrained layout-guided image generation models on LayoutBench-COCO, our new benchmark for OOD layouts with real objects, where our IterInpaint consistently outperforms SOTA baselines in all four splits. Project website: https://layoutbench.github.io

Zhengyuan Yang, Jianfeng Wang, Linjie Li et al. · microsoft-research, uw

We introduce ``Idea to Image,'' a system that enables multimodal iterative self-refinement with GPT-4V(ision) for automatic image design and generation. Humans can quickly identify the characteristics of different text-to-image (T2I) models via iterative explorations. This enables them to efficiently convert their high-level generation ideas into effective T2I prompts that can produce good images. We investigate if systems based on large multimodal models (LMMs) can develop analogous multimodal self-refinement abilities that enable exploring unknown models or environments via self-refining tries. Idea2Img cyclically generates revised T2I prompts to synthesize draft images, and provides directional feedback for prompt revision, both conditioned on its memory of the probed T2I model's characteristics. The iterative self-refinement brings Idea2Img various advantages over vanilla T2I models. Notably, Idea2Img can process input ideas with interleaved image-text sequences, follow ideas with design instructions, and generate images of better semantic and visual qualities. The user preference study validates the efficacy of multimodal iterative self-refinement on automatic image design and generation.

Chung-Ching Lin, Kevin Lin, Linjie Li et al. · microsoft-research, uw

We present a cross-modal Transformer-based framework, which jointly encodes video data and text labels for zero-shot action recognition (ZSAR). Our model employs a conceptually new pipeline by which visual representations are learned in conjunction with visual-semantic associations in an end-to-end manner. The model design provides a natural mechanism for visual and semantic representations to be learned in a shared knowledge space, whereby it encourages the learned visual embedding to be discriminative and more semantically consistent. In zero-shot inference, we devise a simple semantic transfer scheme that embeds semantic relatedness information between seen and unseen classes to composite unseen visual prototypes. Accordingly, the discriminative features in the visual structure could be preserved and exploited to alleviate the typical zero-shot issues of information loss, semantic gap, and the hubness problem. Under a rigorous zero-shot setting of not pre-training on additional datasets, the experiment results show our model considerably improves upon the state of the arts in ZSAR, reaching encouraging top-1 accuracy on UCF101, HMDB51, and ActivityNet benchmark datasets. Code will be made available.

Kevin Lin, Chung-Ching Lin, Lin Liang et al. · microsoft-research, uw

Traditional methods of reconstructing 3D human pose and mesh from single images rely on paired image-mesh datasets, which can be difficult and expensive to obtain. Due to this limitation, model scalability is constrained as well as reconstruction performance. Towards addressing the challenge, we introduce Mesh Pre-Training (MPT), an effective pre-training strategy that leverages large amounts of MoCap data to effectively perform pre-training at scale. We introduce the use of MoCap-generated heatmaps as input representations to the mesh regression transformer and propose a Masked Heatmap Modeling approach for improving pre-training performance. This study demonstrates that pre-training using the proposed MPT allows our models to perform effective inference without requiring fine-tuning. We further show that fine-tuning the pre-trained MPT model considerably improves the accuracy of human mesh reconstruction from single images. Experimental results show that MPT outperforms previous state-of-the-art methods on Human3.6M and 3DPW datasets. As a further application, we benchmark and study MPT on the task of 3D hand reconstruction, showing that our generic pre-training scheme generalizes well to hand pose estimation and achieves promising reconstruction performance.

Jie An, Zhengyuan Yang, Linjie Li et al. · microsoft-research

This work investigates a challenging task named open-domain interleaved image-text generation, which generates interleaved texts and images following an input query. We propose a new interleaved generation framework based on prompting large-language models (LLMs) and pre-trained text-to-image (T2I) models, namely OpenLEAF. In OpenLEAF, the LLM generates textual descriptions, coordinates T2I models, creates visual prompts for generating images, and incorporates global contexts into the T2I models. This global context improves the entity and style consistencies of images in the interleaved generation. For model assessment, we first propose to use large multi-modal models (LMMs) to evaluate the entity and style consistencies of open-domain interleaved image-text sequences. According to the LMM evaluation on our constructed evaluation set, the proposed interleaved generation framework can generate high-quality image-text content for various domains and applications, such as how-to question answering, storytelling, graphical story rewriting, and webpage/poster generation tasks. Moreover, we validate the effectiveness of the proposed LMM evaluation technique with human assessment. We hope our proposed framework, benchmark, and LMM evaluation could help establish the intriguing interleaved image-text generation task.

Xiaodong Wang, Chenfei Wu, Shengming Yin et al. · microsoft-research, pku

3D photography renders a static image into a video with appealing 3D visual effects. Existing approaches typically first conduct monocular depth estimation, then render the input frame to subsequent frames with various viewpoints, and finally use an inpainting model to fill those missing/occluded regions. The inpainting model plays a crucial role in rendering quality, but it is normally trained on out-of-domain data. To reduce the training and inference gap, we propose a novel self-supervised diffusion model as the inpainting module. Given a single input image, we automatically construct a training pair of the masked occluded image and the ground-truth image with random cycle-rendering. The constructed training samples are closely aligned to the testing instances, without the need of data annotation. To make full use of the masked images, we design a Masked Enhanced Block (MEB), which can be easily plugged into the UNet and enhance the semantic conditions. Towards real-world animation, we present a novel task: out-animation, which extends the space and time of input objects. Extensive experiments on real datasets show that our method achieves competitive results with existing SOTA methods.

Ying Jin, Yinpeng Chen, Lijuan Wang et al.

Human-Object Interaction (HOI) recognition is challenging due to two factors: (1) significant imbalance across classes and (2) requiring multiple labels per image. This paper shows that these two challenges can be effectively addressed by improving the classifier with the backbone architecture untouched. Firstly, we encode the semantic correlation among classes into the classification head by initializing the weights with language embeddings of HOIs. As a result, the performance is boosted significantly, especially for the few-shot subset. Secondly, we propose a new loss named LSE-Sign to enhance multi-label learning on a long-tailed dataset. Our simple yet effective method enables detection-free HOI classification, outperforming the state-of-the-arts that require object detection and human pose by a clear margin. Moreover, we transfer the classification model to instance-level HOI detection by connecting it with an off-the-shelf object detector. We achieve state-of-the-art without additional fine-tuning.

An Yan, Zhengyuan Yang, Junda Wu et al. · microsoft-research

Set-of-Mark (SoM) Prompting unleashes the visual grounding capability of GPT-4V, by enabling the model to associate visual objects with tags inserted on the image. These tags, marked with alphanumerics, can be indexed via text tokens for easy reference. Despite the extraordinary performance from GPT-4V, we observe that other Multimodal Large Language Models (MLLMs) struggle to understand these visual tags. To promote the learning of SoM prompting for open-source models, we propose a new learning paradigm: "list items one by one," which asks the model to enumerate and describe all visual tags placed on the image following the alphanumeric orders of tags. By integrating our curated dataset with other visual instruction tuning datasets, we are able to equip existing MLLMs with the SoM prompting ability. Furthermore, we evaluate our finetuned SoM models on five MLLM benchmarks. We find that this new dataset, even in a relatively small size (10k-30k images with tags), significantly enhances visual reasoning capabilities and reduces hallucinations for MLLMs. Perhaps surprisingly, these improvements persist even when the visual tags are omitted from input images during inference. This suggests the potential of "list items one by one" as a new paradigm for training MLLMs, which strengthens the object-text alignment through the use of visual tags in the training stage. Finally, we conduct analyses by probing trained models to understand the working mechanism of SoM. Our code and data are available at \url{https://github.com/zzxslp/SoM-LLaVA}.

Alex Jinpeng Wang, Dongxing Mao, Jiawei Zhang et al. · microsoft-research

Text-conditioned image generation has gained significant attention in recent years and are processing increasingly longer and comprehensive text prompt. In everyday life, dense and intricate text appears in contexts like advertisements, infographics, and signage, where the integration of both text and visuals is essential for conveying complex information. However, despite these advances, the generation of images containing long-form text remains a persistent challenge, largely due to the limitations of existing datasets, which often focus on shorter and simpler text. To address this gap, we introduce TextAtlas5M, a novel dataset specifically designed to evaluate long-text rendering in text-conditioned image generation. Our dataset consists of 5 million long-text generated and collected images across diverse data types, enabling comprehensive evaluation of large-scale generative models on long-text image generation. We further curate 3000 human-improved test set TextAtlasEval across 3 data domains, establishing one of the most extensive benchmarks for text-conditioned generation. Evaluations suggest that the TextAtlasEval benchmarks present significant challenges even for the most advanced proprietary models (e.g. GPT4o with DallE-3), while their open-source counterparts show an even larger performance gap. These evidences position TextAtlas5M as a valuable dataset for training and evaluating future-generation text-conditioned image generation models.

Xuehai He, Weixi Feng, Kaizhi Zheng et al.

Multimodal Language Language Models (MLLMs) demonstrate the emerging abilities of "world models" -- interpreting and reasoning about complex real-world dynamics. To assess these abilities, we posit videos are the ideal medium, as they encapsulate rich representations of real-world dynamics and causalities. To this end, we introduce MMWorld, a new benchmark for multi-discipline, multi-faceted multimodal video understanding. MMWorld distinguishes itself from previous video understanding benchmarks with two unique advantages: (1) multi-discipline, covering various disciplines that often require domain expertise for comprehensive understanding; (2) multi-faceted reasoning, including explanation, counterfactual thinking, future prediction, etc. MMWorld consists of a human-annotated dataset to evaluate MLLMs with questions about the whole videos and a synthetic dataset to analyze MLLMs within a single modality of perception. Together, MMWorld encompasses 1,910 videos across seven broad disciplines and 69 subdisciplines, complete with 6,627 question-answer pairs and associated captions. The evaluation includes 2 proprietary and 10 open-source MLLMs, which struggle on MMWorld (e.g., GPT-4V performs the best with only 52.3\% accuracy), showing large room for improvement. Further ablation studies reveal other interesting findings such as models' different skill sets from humans. We hope MMWorld can serve as an essential step towards world model evaluation in videos.

Liunian Harold Li, Pengchuan Zhang, Haotian Zhang et al.

This paper presents a grounded language-image pre-training (GLIP) model for learning object-level, language-aware, and semantic-rich visual representations. GLIP unifies object detection and phrase grounding for pre-training. The unification brings two benefits: 1) it allows GLIP to learn from both detection and grounding data to improve both tasks and bootstrap a good grounding model; 2) GLIP can leverage massive image-text pairs by generating grounding boxes in a self-training fashion, making the learned representation semantic-rich. In our experiments, we pre-train GLIP on 27M grounding data, including 3M human-annotated and 24M web-crawled image-text pairs. The learned representations demonstrate strong zero-shot and few-shot transferability to various object-level recognition tasks. 1) When directly evaluated on COCO and LVIS (without seeing any images in COCO during pre-training), GLIP achieves 49.8 AP and 26.9 AP, respectively, surpassing many supervised baselines. 2) After fine-tuned on COCO, GLIP achieves 60.8 AP on val and 61.5 AP on test-dev, surpassing prior SoTA. 3) When transferred to 13 downstream object detection tasks, a 1-shot GLIP rivals with a fully-supervised Dynamic Head. Code is released at https://github.com/microsoft/GLIP.

Kevin Lin, Linjie Li, Chung-Ching Lin et al.

The canonical approach to video captioning dictates a caption generation model to learn from offline-extracted dense video features. These feature extractors usually operate on video frames sampled at a fixed frame rate and are often trained on image/video understanding tasks, without adaption to video captioning data. In this work, we present SwinBERT, an end-to-end transformer-based model for video captioning, which takes video frame patches directly as inputs, and outputs a natural language description. Instead of leveraging multiple 2D/3D feature extractors, our method adopts a video transformer to encode spatial-temporal representations that can adapt to variable lengths of video input without dedicated design for different frame rates. Based on this model architecture, we show that video captioning can benefit significantly from more densely sampled video frames as opposed to previous successes with sparsely sampled video frames for video-and-language understanding tasks (e.g., video question answering). Moreover, to avoid the inherent redundancy in consecutive video frames, we propose adaptively learning a sparse attention mask and optimizing it for task-specific performance improvement through better long-range video sequence modeling. Through extensive experiments on 5 video captioning datasets, we show that SwinBERT achieves across-the-board performance improvements over previous methods, often by a large margin. The learned sparse attention masks in addition push the limit to new state of the arts, and can be transferred between different video lengths and between different datasets. Code is available at https://github.com/microsoft/SwinBERT

Zi-Yi Dou, Yichong Xu, Zhe Gan et al.

Vision-and-language (VL) pre-training has proven to be highly effective on various VL downstream tasks. While recent work has shown that fully transformer-based VL models can be more efficient than previous region-feature-based methods, their performance on downstream tasks often degrades significantly. In this paper, we present METER, a Multimodal End-to-end TransformER framework, through which we investigate how to design and pre-train a fully transformer-based VL model in an end-to-end manner. Specifically, we dissect the model designs along multiple dimensions: vision encoders (e.g., CLIP-ViT, Swin transformer), text encoders (e.g., RoBERTa, DeBERTa), multimodal fusion module (e.g., merged attention vs. co-attention), architectural design (e.g., encoder-only vs. encoder-decoder), and pre-training objectives (e.g., masked image modeling). We conduct comprehensive experiments and provide insights on how to train a performant VL transformer. METER achieves an accuracy of 77.64% on the VQAv2 test-std set using only 4M images for pre-training, surpassing the state-of-the-art region-feature-based model by 1.04%, and outperforming the previous best fully transformer-based model by 1.6%. Notably, when further scaled up, our best VQA model achieves an accuracy of 80.54%. Code and pre-trained models are released at https://github.com/zdou0830/METER.

Kevin Lin, Lijuan Wang, Zicheng Liu

We present a graph-convolution-reinforced transformer, named Mesh Graphormer, for 3D human pose and mesh reconstruction from a single image. Recently both transformers and graph convolutional neural networks (GCNNs) have shown promising progress in human mesh reconstruction. Transformer-based approaches are effective in modeling non-local interactions among 3D mesh vertices and body joints, whereas GCNNs are good at exploiting neighborhood vertex interactions based on a pre-specified mesh topology. In this paper, we study how to combine graph convolutions and self-attentions in a transformer to model both local and global interactions. Experimental results show that our proposed method, Mesh Graphormer, significantly outperforms the previous state-of-the-art methods on multiple benchmarks, including Human3.6M, 3DPW, and FreiHAND datasets. Code and pre-trained models are available at https://github.com/microsoft/MeshGraphormer

Tianlong Chen, Yu Cheng, Zhe Gan et al.

Recent advances in computer vision take advantage of adversarial data augmentation to ameliorate the generalization ability of classification models. Here, we present an effective and efficient alternative that advocates adversarial augmentation on intermediate feature embeddings, instead of relying on computationally-expensive pixel-level perturbations. We propose Adversarial Feature Augmentation and Normalization (A-FAN), which (i) first augments visual recognition models with adversarial features that integrate flexible scales of perturbation strengths, (ii) then extracts adversarial feature statistics from batch normalization, and re-injects them into clean features through feature normalization. We validate the proposed approach across diverse visual recognition tasks with representative backbone networks, including ResNets and EfficientNets for classification, Faster-RCNN for detection, and Deeplab V3+ for segmentation. Extensive experiments show that A-FAN yields consistent generalization improvement over strong baselines across various datasets for classification, detection and segmentation tasks, such as CIFAR-10, CIFAR-100, ImageNet, Pascal VOC2007, Pascal VOC2012, COCO2017, and Cityspaces. Comprehensive ablation studies and detailed analyses also demonstrate that adding perturbations to specific modules and layers of classification/detection/segmentation backbones yields optimal performance. Codes and pre-trained models will be made available at: https://github.com/VITA-Group/CV_A-FAN.

Jianfeng Wang, Xi Yin, Lijuan Wang et al.

In this paper, we propose an algorithm, named hashing-based non-maximum suppression (HNMS) to efficiently suppress the non-maximum boxes for object detection. Non-maximum suppression (NMS) is an essential component to suppress the boxes at closely located locations with similar shapes. The time cost tends to be huge when the number of boxes becomes large, especially for crowded scenes. The basic idea of HNMS is to firstly map each box to a discrete code (hash cell) and then remove the boxes with lower confidences if they are in the same cell. Considering the intersection-over-union (IoU) as the metric, we propose a simple yet effective hashing algorithm, named IoUHash, which guarantees that the boxes within the same cell are close enough by a lower IoU bound. For two-stage detectors, we replace NMS in region proposal network with HNMS, and observe significant speed-up with comparable accuracy. For one-stage detectors, HNMS is used as a pre-filter to speed up the suppression with a large margin. Extensive experiments are conducted on CARPK, SKU-110K, CrowdHuman datasets to demonstrate the efficiency and effectiveness of HNMS. Code is released at \url{https://github.com/microsoft/hnms.git}.

Yunzhuo Hao, Jiawei Gu, Huichen Will Wang et al. · microsoft-research, uw

The ability to organically reason over and with both text and images is a pillar of human intelligence, yet the ability of Multimodal Large Language Models (MLLMs) to perform such multimodal reasoning remains under-explored. Existing benchmarks often emphasize text-dominant reasoning or rely on shallow visual cues, failing to adequately assess integrated visual and textual reasoning. We introduce EMMA (Enhanced MultiModal reAsoning), a benchmark targeting organic multimodal reasoning across mathematics, physics, chemistry, and coding. EMMA tasks demand advanced cross-modal reasoning that cannot be addressed by reasoning independently in each modality, offering an enhanced test suite for MLLMs' reasoning capabilities. Our evaluation of state-of-the-art MLLMs on EMMA reveals significant limitations in handling complex multimodal and multi-step reasoning tasks, even with advanced techniques like Chain-of-Thought prompting and test-time compute scaling underperforming. These findings underscore the need for improved multimodal architectures and training paradigms to close the gap between human and model reasoning in multimodality.

Yuyang Zhao, Chung-Ching Lin, Kevin Lin et al. · microsoft-research

Recent developments in 2D visual generation have been remarkably successful. However, 3D and 4D generation remain challenging in real-world applications due to the lack of large-scale 4D data and effective model design. In this paper, we propose to jointly investigate general 3D and 4D generation by leveraging camera and object movements commonly observed in daily life. Due to the lack of real-world 4D data in the community, we first propose a data curation pipeline to obtain camera poses and object motion strength from videos. Based on this pipeline, we introduce a large-scale real-world 4D scene dataset: CamVid-30K. By leveraging all the 3D and 4D data, we develop our framework, GenXD, which allows us to produce any 3D or 4D scene. We propose multiview-temporal modules, which disentangle camera and object movements, to seamlessly learn from both 3D and 4D data. Additionally, GenXD employs masked latent conditions to support a variety of conditioning views. GenXD can generate videos that follow the camera trajectory as well as consistent 3D views that can be lifted into 3D representations. We perform extensive evaluations across various real-world and synthetic datasets, demonstrating GenXD's effectiveness and versatility compared to previous methods in 3D and 4D generation.

Zecheng Tang, Chenfei Wu, Zekai Zhang et al.

To leverage LLMs for visual synthesis, traditional methods convert raster image information into discrete grid tokens through specialized visual modules, while disrupting the model's ability to capture the true semantic representation of visual scenes. This paper posits that an alternative representation of images, vector graphics, can effectively surmount this limitation by enabling a more natural and semantically coherent segmentation of the image information. Thus, we introduce StrokeNUWA, a pioneering work exploring a better visual representation ''stroke tokens'' on vector graphics, which is inherently visual semantics rich, naturally compatible with LLMs, and highly compressed. Equipped with stroke tokens, StrokeNUWA can significantly surpass traditional LLM-based and optimization-based methods across various metrics in the vector graphic generation task. Besides, StrokeNUWA achieves up to a 94x speedup in inference over the speed of prior methods with an exceptional SVG code compression ratio of 6.9%.

Alex Jinpeng Wang, Linjie Li, Kevin Qinghong Lin et al. · microsoft-research

In the evolution of Vision-Language Pre-training, shifting from short-text comprehension to encompassing extended textual contexts is pivotal. Recent autoregressive vision-language models like \cite{flamingo, palme}, leveraging the long-context capability of Large Language Models, have excelled in few-shot text generation tasks but face challenges in alignment tasks. Addressing this gap, we introduce the contrastive loss into text generation models, presenting the COntrastive-Streamlined MultimOdal framework (\ModelName), strategically partitioning the language model into dedicated unimodal text processing and adept multimodal data handling components. \ModelName, our unified framework, merges unimodal and multimodal elements, enhancing model performance for tasks involving textual and visual data while notably reducing learnable parameters. However, these models demand extensive long-text datasets, yet the availability of high-quality long-text video datasets remains limited. To bridge this gap, this work introduces \VideoDatasetName, an inaugural interleaved video-text dataset featuring comprehensive captions, marking a significant step forward. Demonstrating its impact, we illustrate how \VideoDatasetName{} enhances model performance in image-text tasks. With 34% learnable parameters and utilizing 72\% of the available data, our model demonstrates significant superiority over OpenFlamingo~\cite{openflamingo}. For instance, in the 4-shot flickr captioning task, performance notably improves from 57.2% to 65.\%. The contributions of \ModelName{} and \VideoDatasetName{} are underscored by notable performance gains across 14 diverse downstream datasets encompassing both image-text and video-text tasks.

Jie An, Zhengyuan Yang, Jianfeng Wang et al. · microsoft-research

We introduce a Cascaded Diffusion Model (Cas-DM) that improves a Denoising Diffusion Probabilistic Model (DDPM) by effectively incorporating additional metric functions in training. Metric functions such as the LPIPS loss have been proven highly effective in consistency models derived from the score matching. However, for the diffusion counterparts, the methodology and efficacy of adding extra metric functions remain unclear. One major challenge is the mismatch between the noise predicted by a DDPM at each step and the desired clean image that the metric function works well on. To address this problem, we propose Cas-DM, a network architecture that cascades two network modules to effectively apply metric functions to the diffusion model training. The first module, similar to a standard DDPM, learns to predict the added noise and is unaffected by the metric function. The second cascaded module learns to predict the clean image, thereby facilitating the metric function computation. Experiment results show that the proposed diffusion model backbone enables the effective use of the LPIPS loss, leading to state-of-the-art image quality (FID, sFID, IS) on various established benchmarks.

Cheng-Han Chiang, Xiaofei Wang, Linjie Li et al. · microsoft-research

Spoken Language Models (SLMs) are designed to take speech inputs and produce spoken responses. However, current SLMs lack the ability to perform an internal, unspoken thinking process before responding. In contrast, humans typically engage in complex mental reasoning internally, enabling them to communicate ideas clearly and concisely. Thus, integrating an unspoken thought process into SLMs is highly desirable. While naively generating a complete chain-of-thought (CoT) reasoning before starting to talk can enable thinking for SLMs, this induces additional latency for the speech response, as the CoT reasoning can be arbitrarily long. To solve this issue, we propose Stitch, a novel generation method that alternates between the generation of unspoken reasoning chunks and spoken response chunks. Since the audio duration of a chunk of spoken response is much longer than the time to generate the tokens in a chunk of spoken response, we use the remaining free time to generate the unspoken reasoning tokens. When a chunk of audio is played to the user, the model continues to generate the next unspoken reasoning chunk, achieving simultaneous thinking and talking. Remarkably, Stitch matches the latency of baselines that cannot generate unspoken CoT by design while outperforming those baselines by 15% on math reasoning datasets; Stitch also performs equally well on non-reasoning datasets as those baseline models. Some animations and demonstrations are on the project page: https://d223302.github.io/STITCH.

Ming Li, Zhengyuan Yang, Xiyao Wang et al.

Large reasoning models (LRMs) achieve strong reasoning performance by emitting long chains of thought. Yet, these verbose traces slow down inference and often drift into unnecessary detail, known as the overthinking phenomenon. To better understand LRMs' behavior, we systematically analyze the token-level misalignment between reasoning and non-reasoning models. While it is expected that their primary difference lies in the stylistic "thinking cues", LRMs uniquely exhibit two pivotal, previously under-explored phenomena: a Global Misalignment Rebound, where their divergence from non-reasoning models persists or even grows as response length increases, and more critically, a Local Misalignment Diminish, where the misalignment concentrates at the "thinking cues" each sentence starts with but rapidly declines in the remaining of the sentence. Motivated by the Local Misalignment Diminish, we propose FoReaL-Decoding, a collaborative fast-slow thinking decoding method for cost-quality trade-off. In FoReaL-Decoding, a Leading model leads the first few tokens for each sentence, and then a weaker draft model completes the following tokens to the end of each sentence. FoReaL-Decoding adopts a stochastic gate to smoothly interpolate between the small and the large model. On four popular math-reasoning benchmarks (AIME24, GPQA-Diamond, MATH500, AMC23), FoReaL-Decoding reduces theoretical FLOPs by 30 to 50% and trims CoT length by up to 40%, while preserving 86 to 100% of model performance. These results establish FoReaL-Decoding as a simple, plug-and-play route to controllable cost-quality trade-offs in reasoning-centric tasks.

Bingbing Wen, Zhengyuan Yang, Jianfeng Wang et al.

In this paper, we build a visual dialogue dataset, named InfoVisDial, which provides rich informative answers in each round even with external knowledge related to the visual content. Different from existing datasets where the answer is compact and short, InfoVisDial contains long free-form answers with rich information in each round of dialogue. For effective data collection, the key idea is to bridge the large-scale multimodal model (e.g., GIT) and the language models (e.g., GPT-3). GIT can describe the image content even with scene text, while GPT-3 can generate informative dialogue based on the image description and appropriate prompting techniques. With such automatic pipeline, we can readily generate informative visual dialogue data at scale. Then, we ask human annotators to rate the generated dialogues to filter the low-quality conversations.Human analyses show that InfoVisDial covers informative and diverse dialogue topics: $54.4\%$ of the dialogue rounds are related to image scene texts, and $36.7\%$ require external knowledge. Each round's answer is also long and open-ended: $87.3\%$ of answers are unique with an average length of $8.9$, compared with $27.37\%$ and $2.9$ in VisDial. Last, we propose a strong baseline by adapting the GIT model for the visual dialogue task and fine-tune the model on InfoVisDial. Hopefully, our work can motivate more effort on this direction.

Xueyan Zou, Linjie Li, Jianfeng Wang et al. · microsoft-research

Foundation models possess strong capabilities in reasoning and memorizing across modalities. To further unleash the power of foundation models, we present FIND, a generalized interface for aligning foundation models' embeddings with unified image and dataset-level understanding spanning modality and granularity. As shown in the teaser figure, a lightweight transformer interface without tuning any foundation model weights is enough for segmentation, grounding, and retrieval in an interleaved manner. The proposed interface has the following favorable attributes: (1) Generalizable. It applies to various tasks spanning retrieval, segmentation, etc., under the same architecture and weights. (2) Interleavable. With the benefit of multi-task multi-modal training, the proposed interface creates an interleaved shared embedding space. (3) Extendable. The proposed interface is adaptive to new tasks, and new models. In light of the interleaved embedding space, we introduce FIND-Bench, which introduces new training and evaluation annotations to the COCO dataset for interleaved segmentation and retrieval. We are the first work aligning foundations models' embeddings for interleave understanding. Meanwhile, our approach achieves state-of-the-art performance on FIND-Bench and competitive performance on standard retrieval and segmentation settings.

Haibin Wu, Yuxuan Hu, Ruchao Fan et al.

Speech language models (Speech LMs) enable end-to-end speech-text modelling within a single model, offering a promising direction for spoken dialogue systems. The choice of speech-text jointly decoding paradigm plays a critical role in performance, efficiency, and alignment quality. In this work, we systematically compare representative joint speech-text decoding strategies-including the interleaved, and parallel generation paradigms-under a controlled experimental setup using the same base language model, speech tokenizer and training data. Our results show that the interleaved approach achieves the best alignment. However it suffers from slow inference due to long token sequence length. To address this, we propose a novel early-stop interleaved (ESI) pattern that not only significantly accelerates decoding but also yields slightly better performance. Additionally, we curate high-quality question answering (QA) datasets to further improve speech QA performance.