Jianhua Han

Kai Chen, Yunhao Gou, Runhui Huang et al.

GPT-4o, an omni-modal model that enables vocal conversations with diverse emotions and tones, marks a milestone for omni-modal foundation models. However, empowering Large Language Models to perceive and generate images, texts, and speeches end-to-end with publicly available data remains challenging for the open-source community. Existing vision-language models rely on external tools for speech processing, while speech-language models still suffer from limited or totally without vision-understanding capabilities. To address this gap, we propose the EMOVA (EMotionally Omni-present Voice Assistant), to enable Large Language Models with end-to-end speech abilities while maintaining the leading vision-language performance. With a semantic-acoustic disentangled speech tokenizer, we surprisingly notice that omni-modal alignment can further enhance vision-language and speech abilities compared with the bi-modal aligned counterparts. Moreover, a lightweight style module is introduced for the flexible speech style controls including emotions and pitches. For the first time, EMOVA achieves state-of-the-art performance on both the vision-language and speech benchmarks, and meanwhile, supporting omni-modal spoken dialogue with vivid emotions.

Lewei Yao, Jianhua Han, Youpeng Wen et al.

Open-world object detection, as a more general and challenging goal, aims to recognize and localize objects described by arbitrary category names. The recent work GLIP formulates this problem as a grounding problem by concatenating all category names of detection datasets into sentences, which leads to inefficient interaction between category names. This paper presents DetCLIP, a paralleled visual-concept pre-training method for open-world detection by resorting to knowledge enrichment from a designed concept dictionary. To achieve better learning efficiency, we propose a novel paralleled concept formulation that extracts concepts separately to better utilize heterogeneous datasets (i.e., detection, grounding, and image-text pairs) for training. We further design a concept dictionary~(with descriptions) from various online sources and detection datasets to provide prior knowledge for each concept. By enriching the concepts with their descriptions, we explicitly build the relationships among various concepts to facilitate the open-domain learning. The proposed concept dictionary is further used to provide sufficient negative concepts for the construction of the word-region alignment loss\, and to complete labels for objects with missing descriptions in captions of image-text pair data. The proposed framework demonstrates strong zero-shot detection performances, e.g., on the LVIS dataset, our DetCLIP-T outperforms GLIP-T by 9.9% mAP and obtains a 13.5% improvement on rare categories compared to the fully-supervised model with the same backbone as ours.

Guian Fang, Wenbiao Yan, Yuanfan Guo et al.

Text-to-image diffusion models have significantly advanced in conditional image generation. However, these models usually struggle with accurately rendering images featuring humans, resulting in distorted limbs and other anomalies. This issue primarily stems from the insufficient recognition and evaluation of limb qualities in diffusion models. To address this issue, we introduce AbHuman, the first large-scale synthesized human benchmark focusing on anatomical anomalies. This benchmark consists of 56K synthesized human images, each annotated with detailed, bounding-box level labels identifying 147K human anomalies in 18 different categories. Based on this, the recognition of human anomalies can be established, which in turn enhances image generation through traditional techniques such as negative prompting and guidance. To further boost the improvement, we propose HumanRefiner, a novel plug-and-play approach for the coarse-to-fine refinement of human anomalies in text-to-image generation. Specifically, HumanRefiner utilizes a self-diagnostic procedure to detect and correct issues related to both coarse-grained abnormal human poses and fine-grained anomaly levels, facilitating pose-reversible diffusion generation. Experimental results on the AbHuman benchmark demonstrate that HumanRefiner significantly reduces generative discrepancies, achieving a 2.9x improvement in limb quality compared to the state-of-the-art open-source generator SDXL and a 1.4x improvement over DALL-E 3 in human evaluations. Our data and code are available at https://github.com/Enderfga/HumanRefiner.

Lewei Yao, Jianhua Han, Xiaodan Liang et al.

This paper presents DetCLIPv2, an efficient and scalable training framework that incorporates large-scale image-text pairs to achieve open-vocabulary object detection (OVD). Unlike previous OVD frameworks that typically rely on a pre-trained vision-language model (e.g., CLIP) or exploit image-text pairs via a pseudo labeling process, DetCLIPv2 directly learns the fine-grained word-region alignment from massive image-text pairs in an end-to-end manner. To accomplish this, we employ a maximum word-region similarity between region proposals and textual words to guide the contrastive objective. To enable the model to gain localization capability while learning broad concepts, DetCLIPv2 is trained with a hybrid supervision from detection, grounding and image-text pair data under a unified data formulation. By jointly training with an alternating scheme and adopting low-resolution input for image-text pairs, DetCLIPv2 exploits image-text pair data efficiently and effectively: DetCLIPv2 utilizes 13X more image-text pairs than DetCLIP with a similar training time and improves performance. With 13M image-text pairs for pre-training, DetCLIPv2 demonstrates superior open-vocabulary detection performance, e.g., DetCLIPv2 with Swin-T backbone achieves 40.4% zero-shot AP on the LVIS benchmark, which outperforms previous works GLIP/GLIPv2/DetCLIP by 14.4/11.4/4.5% AP, respectively, and even beats its fully-supervised counterpart by a large margin.

Kaican Li, Kai Chen, Haoyu Wang et al.

Contemporary deep-learning object detection methods for autonomous driving usually assume prefixed categories of common traffic participants, such as pedestrians and cars. Most existing detectors are unable to detect uncommon objects and corner cases (e.g., a dog crossing a street), which may lead to severe accidents in some situations, making the timeline for the real-world application of reliable autonomous driving uncertain. One main reason that impedes the development of truly reliably self-driving systems is the lack of public datasets for evaluating the performance of object detectors on corner cases. Hence, we introduce a challenging dataset named CODA that exposes this critical problem of vision-based detectors. The dataset consists of 1500 carefully selected real-world driving scenes, each containing four object-level corner cases (on average), spanning more than 30 object categories. On CODA, the performance of standard object detectors trained on large-scale autonomous driving datasets significantly drops to no more than 12.8% in mAR. Moreover, we experiment with the state-of-the-art open-world object detector and find that it also fails to reliably identify the novel objects in CODA, suggesting that a robust perception system for autonomous driving is probably still far from reach. We expect our CODA dataset to facilitate further research in reliable detection for real-world autonomous driving. Our dataset will be released at https://coda-dataset.github.io.

Quande Liu, Youpeng Wen, Jianhua Han et al.

To bridge the gap between supervised semantic segmentation and real-world applications that acquires one model to recognize arbitrary new concepts, recent zero-shot segmentation attracts a lot of attention by exploring the relationships between unseen and seen object categories, yet requiring large amounts of densely-annotated data with diverse base classes. In this paper, we propose a new open-world semantic segmentation pipeline that makes the first attempt to learn to segment semantic objects of various open-world categories without any efforts on dense annotations, by purely exploiting the image-caption data that naturally exist on the Internet. Our method, Vision-language-driven Semantic Segmentation (ViL-Seg), employs an image and a text encoder to generate visual and text embeddings for the image-caption data, with two core components that endow its segmentation ability: First, the image encoder is jointly trained with a vision-based contrasting and a cross-modal contrasting, which encourage the visual embeddings to preserve both fine-grained semantics and high-level category information that are crucial for the segmentation task. Furthermore, an online clustering head is devised over the image encoder, which allows to dynamically segment the visual embeddings into distinct semantic groups such that they can be classified by comparing with various text embeddings to complete our segmentation pipeline. Experiments show that without using any data with dense annotations, our method can directly segment objects of arbitrary categories, outperforming zero-shot segmentation methods that require data labeling on three benchmark datasets.

Fan Yan, Ming Nie, Xinyue Cai et al.

We present ONCE-3DLanes, a real-world autonomous driving dataset with lane layout annotation in 3D space. Conventional 2D lane detection from a monocular image yields poor performance of following planning and control tasks in autonomous driving due to the case of uneven road. Predicting the 3D lane layout is thus necessary and enables effective and safe driving. However, existing 3D lane detection datasets are either unpublished or synthesized from a simulated environment, severely hampering the development of this field. In this paper, we take steps towards addressing these issues. By exploiting the explicit relationship between point clouds and image pixels, a dataset annotation pipeline is designed to automatically generate high-quality 3D lane locations from 2D lane annotations in 211K road scenes. In addition, we present an extrinsic-free, anchor-free method, called SALAD, regressing the 3D coordinates of lanes in image view without converting the feature map into the bird's-eye view (BEV). To facilitate future research on 3D lane detection, we benchmark the dataset and provide a novel evaluation metric, performing extensive experiments of both existing approaches and our proposed method. The aim of our work is to revive the interest of 3D lane detection in a real-world scenario. We believe our work can lead to the expected and unexpected innovations in both academia and industry.

Yanxin Long, Youpeng Wen, Jianhua Han et al.

Benefiting from large-scale vision-language pre-training on image-text pairs, open-world detection methods have shown superior generalization ability under the zero-shot or few-shot detection settings. However, a pre-defined category space is still required during the inference stage of existing methods and only the objects belonging to that space will be predicted. To introduce a "real" open-world detector, in this paper, we propose a novel method named CapDet to either predict under a given category list or directly generate the category of predicted bounding boxes. Specifically, we unify the open-world detection and dense caption tasks into a single yet effective framework by introducing an additional dense captioning head to generate the region-grounded captions. Besides, adding the captioning task will in turn benefit the generalization of detection performance since the captioning dataset covers more concepts. Experiment results show that by unifying the dense caption task, our CapDet has obtained significant performance improvements (e.g., +2.1% mAP on LVIS rare classes) over the baseline method on LVIS (1203 classes). Besides, our CapDet also achieves state-of-the-art performance on dense captioning tasks, e.g., 15.44% mAP on VG V1.2 and 13.98% on the VG-COCO dataset.

Jianhua Han, Xiajun Deng, Xinyue Cai et al.

We present Laneformer, a conceptually simple yet powerful transformer-based architecture tailored for lane detection that is a long-standing research topic for visual perception in autonomous driving. The dominant paradigms rely on purely CNN-based architectures which often fail in incorporating relations of long-range lane points and global contexts induced by surrounding objects (e.g., pedestrians, vehicles). Inspired by recent advances of the transformer encoder-decoder architecture in various vision tasks, we move forwards to design a new end-to-end Laneformer architecture that revolutionizes the conventional transformers into better capturing the shape and semantic characteristics of lanes, with minimal overhead in latency. First, coupling with deformable pixel-wise self-attention in the encoder, Laneformer presents two new row and column self-attention operations to efficiently mine point context along with the lane shapes. Second, motivated by the appearing objects would affect the decision of predicting lane segments, Laneformer further includes the detected object instances as extra inputs of multi-head attention blocks in the encoder and decoder to facilitate the lane point detection by sensing semantic contexts. Specifically, the bounding box locations of objects are added into Key module to provide interaction with each pixel and query while the ROI-aligned features are inserted into Value module. Extensive experiments demonstrate our Laneformer achieves state-of-the-art performances on CULane benchmark, in terms of 77.1% F1 score. We hope our simple and effective Laneformer will serve as a strong baseline for future research in self-attention models for lane detection.

Shipeng Yan, Lanqing Hong, Hang Xu et al.

Vision-language pre-training (VLP) has attracted increasing attention recently. With a large amount of image-text pairs, VLP models trained with contrastive loss have achieved impressive performance in various tasks, especially the zero-shot generalization on downstream datasets. In practical applications, however, massive data are usually collected in a streaming fashion, requiring VLP models to continuously integrate novel knowledge from incoming data and retain learned knowledge. In this work, we focus on learning a VLP model with sequential chunks of image-text pair data. To tackle the catastrophic forgetting issue in this multi-modal continual learning setting, we first introduce pseudo text replay that generates hard negative texts conditioned on the training images in memory, which not only better preserves learned knowledge but also improves the diversity of negative samples in the contrastive loss. Moreover, we propose multi-modal knowledge distillation between images and texts to align the instance-wise prediction between old and new models. We incrementally pre-train our model on both the instance and class incremental splits of the Conceptual Caption dataset, and evaluate the model on zero-shot image classification and image-text retrieval tasks. Our method consistently outperforms the existing baselines with a large margin, which demonstrates its superiority. Notably, we realize an average performance boost of $4.60\%$ on image-classification downstream datasets for the class incremental split.

Xiwen Liang, Yangxin Wu, Jianhua Han et al.

Aiming towards a holistic understanding of multiple downstream tasks simultaneously, there is a need for extracting features with better transferability. Though many latest self-supervised pre-training methods have achieved impressive performance on various vision tasks under the prevailing pretrain-finetune paradigm, their generalization capacity to multi-task learning scenarios is yet to be explored. In this paper, we extensively investigate the transfer performance of various types of self-supervised methods, e.g., MoCo and SimCLR, on three downstream tasks, including semantic segmentation, drivable area segmentation, and traffic object detection, on the large-scale driving dataset BDD100K. We surprisingly find that their performances are sub-optimal or even lag far behind the single-task baseline, which may be due to the distinctions of training objectives and architectural design lied in the pretrain-finetune paradigm. To overcome this dilemma as well as avoid redesigning the resource-intensive pre-training stage, we propose a simple yet effective pretrain-adapt-finetune paradigm for general multi-task training, where the off-the-shelf pretrained models can be effectively adapted without increasing the training overhead. During the adapt stage, we utilize learnable multi-scale adapters to dynamically adjust the pretrained model weights supervised by multi-task objectives while leaving the pretrained knowledge untouched. Furthermore, we regard the vision-language pre-training model CLIP as a strong complement to the pretrain-adapt-finetune paradigm and propose a novel adapter named LV-Adapter, which incorporates language priors in the multi-task model via task-specific prompting and alignment between visual and textual features.

Zhili Liu, Jianhua Han, Lanqing Hong et al.

Self-supervised learning (SSL), especially contrastive methods, has raised attraction recently as it learns effective transferable representations without semantic annotations. A common practice for self-supervised pre-training is to use as much data as possible. For a specific downstream task, however, involving irrelevant data in pre-training may degenerate the downstream performance, observed from our extensive experiments. On the other hand, for existing SSL methods, it is burdensome and infeasible to use different downstream-task-customized datasets in pre-training for different tasks. To address this issue, we propose a novel SSL paradigm called Scalable Dynamic Routing (SDR), which can be trained once and deployed efficiently to different downstream tasks with task-customized pre-trained models. Specifically, we construct the SDRnet with various sub-nets and train each sub-net with only one subset of the data by data-aware progressive training. When a downstream task arrives, we route among all the pre-trained sub-nets to get the best along with its corresponding weights. Experiment results show that our SDR can train 256 sub-nets on ImageNet simultaneously, which provides better transfer performance than a unified model trained on the full ImageNet, achieving state-of-the-art (SOTA) averaged accuracy over 11 downstream classification tasks and AP on PASCAL VOC detection task.

Runhui Huang, Yanxin Long, Jianhua Han et al.

Large-scale cross-modal pre-training paradigms have recently shown ubiquitous success on a wide range of downstream tasks, e.g., zero-shot classification, retrieval and image captioning. However, their successes highly rely on the scale and quality of web-crawled data that naturally contain incomplete and noisy information (e.g., wrong or irrelevant content). Existing works either design manual rules to clean data or generate pseudo-targets as auxiliary signals for reducing noise impact, which do not explicitly tackle both the incorrect and incomplete challenges simultaneously. In this paper, to automatically mitigate the impact of noise by solely mining over existing data, we propose a principled Noise-robust Language-Image Pre-training framework (NLIP) to stabilize pre-training via two schemes: noise-harmonization and noise-completion. First, in noise-harmonization scheme, NLIP estimates the noise probability of each pair according to the memorization effect of cross-modal transformers, then adopts noise-adaptive regularization to harmonize the cross-modal alignments with varying degrees. Second, in noise-completion scheme, to enrich the missing object information of text, NLIP injects a concept-conditioned cross-modal decoder to obtain semantic-consistent synthetic captions to complete noisy ones, which uses the retrieved visual concepts (i.e., objects' names) for the corresponding image to guide captioning generation. By collaboratively optimizing noise-harmonization and noise-completion schemes, our NLIP can alleviate the common noise effects during image-text pre-training in a more efficient way. Extensive experiments show the significant performance improvements of our NLIP using only 26M data over existing pre-trained models (e.g., CLIP, FILIP and BLIP) on 12 zero-shot classification datasets, MSCOCO image captioning and zero-shot image-text retrieval tasks.

Kaixin Cai, Pengzhen Ren, Jianhua Han et al.

Open-world semantic segmentation presently relies significantly on extensive image-text pair datasets, which often suffer from a lack of fine-grained pixel annotations on sufficient categories. The acquisition of such data is rendered economically prohibitive due to the substantial investments of both human labor and time. In light of the formidable image generation capabilities of diffusion models, we introduce a novel diffusion model-driven pipeline for automatically generating datasets tailored to the needs of open-world semantic segmentation, named "MagicSeg". Our MagicSeg initiates from class labels and proceeds to generate high-fidelity textual descriptions, which in turn serve as guidance for the diffusion model to generate images. Rather than only generating positive samples for each label, our process encompasses the simultaneous generation of corresponding negative images, designed to serve as paired counterfactual samples for contrastive training. Then, to provide a self-supervised signal for open-world segmentation pretraining, our MagicSeg integrates an open-vocabulary detection model and an interactive segmentation model to extract object masks as precise segmentation labels from images based on the provided category labels. By applying our dataset to the contrastive language-image pretraining model with the pseudo mask supervision and the auxiliary counterfactual contrastive training, the downstream model obtains strong performance on open-world semantic segmentation. We evaluate our model on PASCAL VOC, PASCAL Context, and COCO, achieving SOTA with performance of 62.9%, 26.7%, and 40.2%, respectively, demonstrating our dataset's effectiveness in enhancing open-world semantic segmentation capabilities. Project website: https://github.com/ckxhp/magicseg.

Yunshuang Nie, Bingqian Lin, Minzhe Niu et al.

Pre-trained Multi-modal Large Language Models (MLLMs) provide a knowledge-rich foundation for post-training by leveraging their inherent perception and reasoning capabilities to solve complex tasks. However, the lack of an efficient evaluation framework impedes the diagnosis of their performance bottlenecks. Current evaluation primarily relies on testing after supervised fine-tuning, which introduces laborious additional training and autoregressive decoding costs. Meanwhile, common pre-training metrics cannot quantify a model's perception and reasoning abilities in a disentangled manner. Furthermore, existing evaluation benchmarks are typically limited in scale or misaligned with pre-training objectives. Thus, we propose RADAR, an efficient ability-centric evaluation framework for Revealing Asymmetric Development of Abilities in MLLM pRe-training. RADAR involves two key components: (1) Soft Discrimination Score, a novel metric for robustly tracking ability development without fine-tuning, based on quantifying nuanced gradations of the model preference for the correct answer over distractors; and (2) Multi-Modal Mixture Benchmark, a new 15K+ sample benchmark for comprehensively evaluating pre-trained MLLMs' perception and reasoning abilities in a 0-shot manner, where we unify authoritative benchmark datasets and carefully collect new datasets, extending the evaluation scope and addressing the critical gaps in current benchmarks. With RADAR, we comprehensively reveal the asymmetric development of perceptual and reasoning capabilities in pretrained MLLMs across diverse factors, including data volume, model size, and pretraining strategy. Our RADAR underscores the need for a decomposed perspective on pre-training ability bottlenecks, informing targeted interventions to advance MLLMs efficiently. Our code is publicly available at https://github.com/Nieysh/RADAR.

Yihan Zeng, Chenhan Jiang, Jiageng Mao et al.

Contrastive Language-Image Pre-training, benefiting from large-scale unlabeled text-image pairs, has demonstrated great performance in open-world vision understanding tasks. However, due to the limited Text-3D data pairs, adapting the success of 2D Vision-Language Models (VLM) to the 3D space remains an open problem. Existing works that leverage VLM for 3D understanding generally resort to constructing intermediate 2D representations for the 3D data, but at the cost of losing 3D geometry information. To take a step toward open-world 3D vision understanding, we propose Contrastive Language-Image-Point Cloud Pretraining (CLIP$^2$) to directly learn the transferable 3D point cloud representation in realistic scenarios with a novel proxy alignment mechanism. Specifically, we exploit naturally-existed correspondences in 2D and 3D scenarios, and build well-aligned and instance-based text-image-point proxies from those complex scenarios. On top of that, we propose a cross-modal contrastive objective to learn semantic and instance-level aligned point cloud representation. Experimental results on both indoor and outdoor scenarios show that our learned 3D representation has great transfer ability in downstream tasks, including zero-shot and few-shot 3D recognition, which boosts the state-of-the-art methods by large margins. Furthermore, we provide analyses of the capability of different representations in real scenarios and present the optional ensemble scheme.

Haoyuan Li, Qihang Cao, Tao Tang et al.

Recent progress in spatial reasoning with Multimodal Large Language Models (MLLMs) increasingly leverages geometric priors from 3D encoders. However, most existing integration strategies remain passive: geometry is exposed as a global stream and fused in an indiscriminate manner, which often induces semantic-geometry misalignment and redundant signals. We propose GeoThinker, a framework that shifts the paradigm from passive fusion to active perception. Instead of feature mixing, GeoThinker enables the model to selectively retrieve geometric evidence conditioned on its internal reasoning demands. GeoThinker achieves this through Spatial-Grounded Fusion applied at carefully selected VLM layers, where semantic visual priors selectively query and integrate task-relevant geometry via frame-strict cross-attention, further calibrated by Importance Gating that biases per-frame attention toward task-relevant structures. Comprehensive evaluation results show that GeoThinker sets a new state-of-the-art in spatial intelligence, achieving a peak score of 72.6 on the VSI-Bench. Furthermore, GeoThinker demonstrates robust generalization and significantly improved spatial perception across complex downstream scenarios, including embodied referring and autonomous driving. Our results indicate that the ability to actively integrate spatial structures is essential for next-generation spatial intelligence. Code can be found at https://github.com/Li-Hao-yuan/GeoThinker.

Yanxin Long, Jianhua Han, Runhui Huang et al.

Inspired by the success of vision-language methods (VLMs) in zero-shot classification, recent works attempt to extend this line of work into object detection by leveraging the localization ability of pre-trained VLMs and generating pseudo labels for unseen classes in a self-training manner. However, since the current VLMs are usually pre-trained with aligning sentence embedding with global image embedding, the direct use of them lacks fine-grained alignment for object instances, which is the core of detection. In this paper, we propose a simple but effective fine-grained Visual-Text Prompt-driven self-training paradigm for Open-Vocabulary Detection (VTP-OVD) that introduces a fine-grained visual-text prompt adapting stage to enhance the current self-training paradigm with a more powerful fine-grained alignment. During the adapting stage, we enable VLM to obtain fine-grained alignment by using learnable text prompts to resolve an auxiliary dense pixel-wise prediction task. Furthermore, we propose a visual prompt module to provide the prior task information (i.e., the categories need to be predicted) for the vision branch to better adapt the pre-trained VLM to the downstream tasks. Experiments show that our method achieves the state-of-the-art performance for open-vocabulary object detection, e.g., 31.5% mAP on unseen classes of COCO.

Kai Chen, Chunwei Wang, Kuo Yang et al.

The rapid development of large language models (LLMs) has not only provided numerous opportunities but also presented significant challenges. This becomes particularly evident when LLMs inadvertently generate harmful or toxic content, either unintentionally or because of intentional inducement. Existing alignment methods usually direct LLMs toward the favorable outcomes by utilizing human-annotated, flawless instruction-response pairs. Conversely, this study proposes a novel alignment technique based on mistake analysis, which deliberately exposes LLMs to erroneous content to learn the reasons for mistakes and how to avoid them. In this case, mistakes are repurposed into valuable data for alignment, effectively helping to avoid the production of erroneous responses. Without external models or human annotations, our method leverages a model's intrinsic ability to discern undesirable mistakes and improves the safety of its generated responses. Experimental results reveal that our method outperforms existing alignment approaches in enhancing model safety while maintaining the overall utility.

Xiwen Liang, Minzhe Niu, Jianhua Han et al.

Multi-task learning has emerged as a powerful paradigm to solve a range of tasks simultaneously with good efficiency in both computation resources and inference time. However, these algorithms are designed for different tasks mostly not within the scope of autonomous driving, thus making it hard to compare multi-task methods in autonomous driving. Aiming to enable the comprehensive evaluation of present multi-task learning methods in autonomous driving, we extensively investigate the performance of popular multi-task methods on the large-scale driving dataset, which covers four common perception tasks, i.e., object detection, semantic segmentation, drivable area segmentation, and lane detection. We provide an in-depth analysis of current multi-task learning methods under different common settings and find out that the existing methods make progress but there is still a large performance gap compared with single-task baselines. To alleviate this dilemma in autonomous driving, we present an effective multi-task framework, VE-Prompt, which introduces visual exemplars via task-specific prompting to guide the model toward learning high-quality task-specific representations. Specifically, we generate visual exemplars based on bounding boxes and color-based markers, which provide accurate visual appearances of target categories and further mitigate the performance gap. Furthermore, we bridge transformer-based encoders and convolutional layers for efficient and accurate unified perception in autonomous driving. Comprehensive experimental results on the diverse self-driving dataset BDD100K show that the VE-Prompt improves the multi-task baseline and further surpasses single-task models.

Xinpeng Ding, Jianhua Han, Hang Xu et al.

Recent efforts to use natural language for interpretable driving focus mainly on planning, neglecting perception tasks. In this paper, we address this gap by introducing ROLISP (Risk Object Localization and Intention and Suggestion Prediction), which towards interpretable risk object detection and suggestion for ego car motions. Accurate ROLISP implementation requires extensive reasoning to identify critical traffic objects and infer their intentions, prompting us to explore the capabilities of multimodal large language models (MLLMs). However, the limited perception performance of CLIP-ViT vision encoders in existing MLLMs struggles with capturing essential visual perception information, e.g., high-resolution, multi-scale and visual-related inductive biases, which are important for autonomous driving. Addressing these challenges, we introduce HiLM-D, a resource-efficient framework that enhances visual information processing in MLLMs for ROLISP. Our method is motivated by the fact that the primary variations in autonomous driving scenarios are the motion trajectories rather than the semantic or appearance information (e.g., the shapes and colors) of objects. Hence, the visual process of HiLM-D is a two-stream framework: (i) a temporal reasoning stream, receiving low-resolution dynamic video content, to capture temporal semantics, and (ii) a spatial perception stream, receiving a single high-resolution frame, to capture holistic visual perception-related information. The spatial perception stream can be made very lightweight by a well-designed P-Adapter, which is lightweight, training-efficient, and easily integrated into existing MLLMs. Experiments on the DRAMA-ROLISP dataset show HiLM-D's significant improvements over current MLLMs, with a 3.7% in BLEU-4 for captioning and 8.7% in mIoU for detection.

Bowen Dong, Jiaxi Gu, Jianhua Han et al.

Existing open-world universal segmentation approaches usually leverage CLIP and pre-computed proposal masks to treat open-world segmentation tasks as proposal classification. However, 1) these works cannot handle universal segmentation in an end-to-end manner, and 2) the limited scale of panoptic datasets restricts the open-world segmentation ability on things classes. In this paper, we present Vision-Language Omni-Supervised Segmentation (VLOSS). VLOSS starts from a Mask2Former universal segmentation framework with CLIP text encoder. To improve the open-world segmentation ability, we leverage omni-supervised data (i.e., panoptic segmentation data, object detection data, and image-text pairs data) into training, thus enriching the open-world segmentation ability and achieving better segmentation accuracy. To better improve the training efficiency and fully release the power of omni-supervised data, we propose several advanced techniques, i.e., FPN-style encoder, switchable training technique, and positive classification loss. Benefiting from the end-to-end training manner with proposed techniques, VLOSS can be applied to various open-world segmentation tasks without further adaptation. Experimental results on different open-world panoptic and instance segmentation benchmarks demonstrate the effectiveness of VLOSS. Notably, with fewer parameters, our VLOSS with Swin-Tiny backbone surpasses MaskCLIP by ~2% in terms of mask AP on LVIS v1 dataset.

Qingping Zheng, Yuanfan Guo, Jiankang Deng et al.

Stable diffusion, a generative model used in text-to-image synthesis, frequently encounters resolution-induced composition problems when generating images of varying sizes. This issue primarily stems from the model being trained on pairs of single-scale images and their corresponding text descriptions. Moreover, direct training on images of unlimited sizes is unfeasible, as it would require an immense number of text-image pairs and entail substantial computational expenses. To overcome these challenges, we propose a two-stage pipeline named Any-Size-Diffusion (ASD), designed to efficiently generate well-composed images of any size, while minimizing the need for high-memory GPU resources. Specifically, the initial stage, dubbed Any Ratio Adaptability Diffusion (ARAD), leverages a selected set of images with a restricted range of ratios to optimize the text-conditional diffusion model, thereby improving its ability to adjust composition to accommodate diverse image sizes. To support the creation of images at any desired size, we further introduce a technique called Fast Seamless Tiled Diffusion (FSTD) at the subsequent stage. This method allows for the rapid enlargement of the ASD output to any high-resolution size, avoiding seaming artifacts or memory overloads. Experimental results on the LAION-COCO and MM-CelebA-HQ benchmarks demonstrate that ASD can produce well-structured images of arbitrary sizes, cutting down the inference time by 2x compared to the traditional tiled algorithm.

Xinchi Deng, Han Shi, Runhui Huang et al.

Cross-modal pre-training has shown impressive performance on a wide range of downstream tasks, benefiting from massive image-text pairs collected from the Internet. In practice, online data are growing constantly, highlighting the importance of the ability of pre-trained model to learn from data that is continuously growing. Existing works on cross-modal pre-training mainly focus on training a network with fixed architecture. However, it is impractical to limit the model capacity when considering the continuously growing nature of pre-training data in real-world applications. On the other hand, it is important to utilize the knowledge in the current model to obtain efficient training and better performance. To address the above issues, in this paper, we propose GrowCLIP, a data-driven automatic model growing algorithm for contrastive language-image pre-training with continuous image-text pairs as input. Specially, we adopt a dynamic growth space and seek out the optimal architecture at each growth step to adapt to online learning scenarios. And the shared encoder is proposed in our growth space to enhance the degree of cross-modal fusion. Besides, we explore the effect of growth in different dimensions, which could provide future references for the design of cross-modal model architecture. Finally, we employ parameter inheriting with momentum (PIM) to maintain the previous knowledge and address the issue of the local minimum dilemma. Compared with the existing methods, GrowCLIP improves 2.3% average top-1 accuracy on zero-shot image classification of 9 downstream tasks. As for zero-shot image retrieval, GrowCLIP can improve 1.2% for top-1 image-to-text recall on Flickr30K dataset.

Zhili Liu, Kai Chen, Yifan Zhang et al.

Text-to-image (T2I) diffusion models often inadvertently generate unwanted concepts such as watermarks and unsafe images. These concepts, termed as the "implicit concepts", could be unintentionally learned during training and then be generated uncontrollably during inference. Existing removal methods still struggle to eliminate implicit concepts primarily due to their dependency on the model's ability to recognize concepts it actually can not discern. To address this, we utilize the intrinsic geometric characteristics of implicit concepts and present the Geom-Erasing, a novel concept removal method based on the geometric-driven control. Specifically, once an unwanted implicit concept is identified, we integrate the existence and geometric information of the concept into the text prompts with the help of an accessible classifier or detector model. Subsequently, the model is optimized to identify and disentangle this information, which is then adopted as negative prompts during generation. Moreover, we introduce the Implicit Concept Dataset (ICD), a novel image-text dataset imbued with three typical implicit concepts (i.e., QR codes, watermarks, and text), reflecting real-life situations where implicit concepts are easily injected. Geom-Erasing effectively mitigates the generation of implicit concepts, achieving the state-of-the-art results on the Inappropriate Image Prompts (I2P) and our challenging Implicit Concept Dataset (ICD) benchmarks.

Runhui Huang, Xinpeng Ding, Chunwei Wang et al.

High-resolution inputs enable Large Vision-Language Models (LVLMs) to discern finer visual details, enhancing their comprehension capabilities. To reduce the training and computation costs caused by high-resolution input, one promising direction is to use sliding windows to slice the input into uniform patches, each matching the input size of the well-trained vision encoder. Although efficient, this slicing strategy leads to the fragmentation of original input, i.e., the continuity of contextual information and spatial geometry is lost across patches, adversely affecting performance in cross-patch context perception and position-specific tasks. To overcome these shortcomings, we introduce HiRes-LLaVA, a novel framework designed to efficiently process any size of high-resolution input without altering the original contextual and geometric information. HiRes-LLaVA comprises two innovative components: (i) a SliceRestore adapter that reconstructs sliced patches into their original form, efficiently extracting both global and local features via down-up-sampling and convolution layers, and (ii) a Self-Mining Sampler to compresses the vision tokens based on themselves, preserving the original context and positional information while reducing training overhead. To assess the ability of handling context fragmentation, we construct a new benchmark, EntityGrid-QA, consisting of edge-related and position-related tasks. Our comprehensive experiments demonstrate the superiority of HiRes-LLaVA on both existing public benchmarks and on EntityGrid-QA, particularly on document-oriented tasks, establishing new standards for handling high-resolution inputs.

Cong Wang, Jiaxi Gu, Panwen Hu et al.

Following the advancements in text-guided image generation technology exemplified by Stable Diffusion, video generation is gaining increased attention in the academic community. However, relying solely on text guidance for video generation has serious limitations, as videos contain much richer content than images, especially in terms of motion. This information can hardly be adequately described with plain text. Fortunately, in computer vision, various visual representations can serve as additional control signals to guide generation. With the help of these signals, video generation can be controlled in finer detail, allowing for greater flexibility for different applications. Integrating various controls, however, is nontrivial. In this paper, we propose a universal framework called EasyControl. By propagating and injecting condition features through condition adapters, our method enables users to control video generation with a single condition map. With our framework, various conditions including raw pixels, depth, HED, etc., can be integrated into different Unet-based pre-trained video diffusion models at a low practical cost. We conduct comprehensive experiments on public datasets, and both quantitative and qualitative results indicate that our method outperforms state-of-the-art methods. EasyControl significantly improves various evaluation metrics across multiple validation datasets compared to previous works. Specifically, for the sketch-to-video generation task, EasyControl achieves an improvement of 152.0 on FVD and 19.9 on IS, respectively, in UCF101 compared with VideoComposer. For fidelity, our model demonstrates powerful image retention ability, resulting in high FVD and IS in UCF101 and MSR-VTT compared to other image-to-video models.

Yi Zhu, Yanpeng Zhou, Chunwei Wang et al.

Currently, vision encoder models like Vision Transformers (ViTs) typically excel at image recognition tasks but cannot simultaneously support text recognition like human visual recognition. To address this limitation, we propose UNIT, a novel training framework aimed at UNifying Image and Text recognition within a single model. Starting with a vision encoder pre-trained with image recognition tasks, UNIT introduces a lightweight language decoder for predicting text outputs and a lightweight vision decoder to prevent catastrophic forgetting of the original image encoding capabilities. The training process comprises two stages: intra-scale pretraining and inter-scale finetuning. During intra-scale pretraining, UNIT learns unified representations from multi-scale inputs, where images and documents are at their commonly used resolution, to enable fundamental recognition capability. In the inter-scale finetuning stage, the model introduces scale-exchanged data, featuring images and documents at resolutions different from the most commonly used ones, to enhance its scale robustness. Notably, UNIT retains the original vision encoder architecture, making it cost-free in terms of inference and deployment. Experiments across multiple benchmarks confirm that our method significantly outperforms existing methods on document-related tasks (e.g., OCR and DocQA) while maintaining the performances on natural images, demonstrating its ability to substantially enhance text recognition without compromising its core image recognition capabilities.

Ming Nie, Dan Ding, Chunwei Wang et al.

Large language models (LLMs) have demonstrated exceptional capabilities in text understanding, which has paved the way for their expansion into video LLMs (Vid-LLMs) to analyze video data. However, current Vid-LLMs struggle to simultaneously retain high-quality frame-level semantic information (i.e., a sufficient number of tokens per frame) and comprehensive video-level temporal information (i.e., an adequate number of sampled frames per video). This limitation hinders the advancement of Vid-LLMs towards fine-grained video understanding. To address this issue, we introduce the SlowFocus mechanism, which significantly enhances the equivalent sampling frequency without compromising the quality of frame-level visual tokens. SlowFocus begins by identifying the query-related temporal segment based on the posed question, then performs dense sampling on this segment to extract local high-frequency features. A multi-frequency mixing attention module is further leveraged to aggregate these local high-frequency details with global low-frequency contexts for enhanced temporal comprehension. Additionally, to tailor Vid-LLMs to this innovative mechanism, we introduce a set of training strategies aimed at bolstering both temporal grounding and detailed temporal reasoning capabilities. Furthermore, we establish FineAction-CGR, a benchmark specifically devised to assess the ability of Vid-LLMs to process fine-grained temporal understanding tasks. Comprehensive experiments demonstrate the superiority of our mechanism across both existing public video understanding benchmarks and our proposed FineAction-CGR.

Runhui Huang, Jianhua Han, Guansong Lu et al.

Recently, large-scale diffusion models, e.g., Stable diffusion and DallE2, have shown remarkable results on image synthesis. On the other hand, large-scale cross-modal pre-trained models (e.g., CLIP, ALIGN, and FILIP) are competent for various downstream tasks by learning to align vision and language embeddings. In this paper, we explore the possibility of jointly modeling generation and discrimination. Specifically, we propose DiffDis to unify the cross-modal generative and discriminative pretraining into one single framework under the diffusion process. DiffDis first formulates the image-text discriminative problem as a generative diffusion process of the text embedding from the text encoder conditioned on the image. Then, we propose a novel dual-stream network architecture, which fuses the noisy text embedding with the knowledge of latent images from different scales for image-text discriminative learning. Moreover, the generative and discriminative tasks can efficiently share the image-branch network structure in the multi-modality model. Benefiting from diffusion-based unified training, DiffDis achieves both better generation ability and cross-modal semantic alignment in one architecture. Experimental results show that DiffDis outperforms single-task models on both the image generation and the image-text discriminative tasks, e.g., 1.65% improvement on average accuracy of zero-shot classification over 12 datasets and 2.42 improvement on FID of zero-shot image synthesis.

Xiwen Liang, Liang Ma, Shanshan Guo et al.

Understanding and following natural language instructions while navigating through complex, real-world environments poses a significant challenge for general-purpose robots. These environments often include obstacles and pedestrians, making it essential for autonomous agents to possess the capability of self-corrected planning to adjust their actions based on feedback from the surroundings. However, the majority of existing vision-and-language navigation (VLN) methods primarily operate in less realistic simulator settings and do not incorporate environmental feedback into their decision-making processes. To address this gap, we introduce a novel zero-shot framework called CorNav, utilizing a large language model for decision-making and comprising two key components: 1) incorporating environmental feedback for refining future plans and adjusting its actions, and 2) multiple domain experts for parsing instructions, scene understanding, and refining predicted actions. In addition to the framework, we develop a 3D simulator that renders realistic scenarios using Unreal Engine 5. To evaluate the effectiveness and generalization of navigation agents in a zero-shot multi-task setting, we create a benchmark called NavBench. Extensive experiments demonstrate that CorNav consistently outperforms all baselines by a significant margin across all tasks. On average, CorNav achieves a success rate of 28.1\%, surpassing the best baseline's performance of 20.5\%.

Ming Nie, Chunwei Wang, Jianhua Han et al.

Unified vision-language models have made significant progress in multimodal understanding and generation, yet they largely fall short in producing multimodal interleaved outputs, which is a crucial capability for tasks like visual storytelling and step-by-step visual reasoning. In this work, we propose a reinforcement learning-based post-training strategy to unlock this capability in existing unified models, without relying on large-scale multimodal interleaved datasets. We begin with a warm-up stage using a hybrid dataset comprising curated interleaved sequences and limited data for multimodal understanding and text-to-image generation, which exposes the model to interleaved generation patterns while preserving its pretrained capabilities. To further refine interleaved generation, we propose a unified policy optimization framework that extends Group Relative Policy Optimization (GRPO) to the multimodal setting. Our approach jointly models text and image generation within a single decoding trajectory and optimizes it with our novel hybrid rewards covering textual relevance, visual-text alignment, and structural fidelity. Additionally, we incorporate process-level rewards to provide step-wise guidance, enhancing training efficiency in complex multimodal tasks. Experiments on MMIE and InterleavedBench demonstrate that our approach significantly enhances the quality and coherence of multimodal interleaved generation.

Ming Nie, Renyuan Peng, Chunwei Wang et al.

Large vision-language models (VLMs) have garnered increasing interest in autonomous driving areas, due to their advanced capabilities in complex reasoning tasks essential for highly autonomous vehicle behavior. Despite their potential, research in autonomous systems is hindered by the lack of datasets with annotated reasoning chains that explain the decision-making processes in driving. To bridge this gap, we present Reason2Drive, a benchmark dataset with over 600K video-text pairs, aimed at facilitating the study of interpretable reasoning in complex driving environments. We distinctly characterize the autonomous driving process as a sequential combination of perception, prediction, and reasoning steps, and the question-answer pairs are automatically collected from a diverse range of open-source outdoor driving datasets, including nuScenes, Waymo and ONCE. Moreover, we introduce a novel aggregated evaluation metric to assess chain-based reasoning performance in autonomous systems, addressing the semantic ambiguities of existing metrics such as BLEU and CIDEr. Based on the proposed benchmark, we conduct experiments to assess various existing VLMs, revealing insights into their reasoning capabilities. Additionally, we develop an efficient approach to empower VLMs to leverage object-level perceptual elements in both feature extraction and prediction, further enhancing their reasoning accuracy. The code and dataset will be released.

Zhenquan Yao, Zitong Huang, Yihan Zeng et al.

Graphical User Interface (GUI) Agents, benefiting from recent advances in multimodal large language models (MLLM), have achieved significant development. However, due to the frequent updates of GUI applications, adapting to new tasks without forgetting old tasks in GUI continual learning remains an open problem. In this work, we reveal that while Supervised Fine-Tuning (SFT) facilitates fast adaptation, it often triggers knowledge overwriting, whereas Reinforcement Learning (RL) demonstrates an inherent resilience that shields prior interaction logic from erasure. Based on this insight, we propose a \textbf{C}ontinual \textbf{G}UI \textbf{L}earning (CGL) framework that dynamically balances adaptation efficiency and skill retention by enhancing the synergy between SFT and RL. Specifically, we introduce an SFT proportion adjustment mechanism guided by policy entropy to dynamically control the weight allocation between the SFT and RL training phases. To resolve explicit gradient interference, we further develop a specialized gradient surgery strategy. By projecting exploratory SFT gradients onto GRPO-based anchor gradients, our method explicitly clips the components of SFT gradients that conflict with GRPO. On top of that, we establish an AndroidControl-CL benchmark, which divides GUI applications into distinct task groups to effectively simulate and evaluate the performance of continual GUI learning. Experimental results demonstrate the effectiveness of our proposed CGL framework across continual learning scenarios. The benchmark, code, and model will be made publicly available.

Bingqian Lin, Yunshuang Nie, Ziming Wei et al.

Vision-and-Language Navigation (VLN), as a crucial research problem of Embodied AI, requires an embodied agent to navigate through complex 3D environments following natural language instructions. Recent research has highlighted the promising capacity of large language models (LLMs) in VLN by improving navigational reasoning accuracy and interpretability. However, their predominant use in an offline manner usually suffers from substantial domain gap between the VLN task and the LLM training corpus. This paper introduces a novel strategy called Navigational Chain-of-Thought (NavCoT), where we fulfill parameter-efficient in-domain training to enable self-guided navigational decision, leading to a significant mitigation of the domain gap in a cost-effective manner. Specifically, at each timestep, the LLM is prompted to forecast the navigational chain-of-thought by: 1) acting as a world model to imagine the next observation according to the instruction, 2) selecting the candidate observation that best aligns with the imagination, and 3) determining the action based on the reasoning from the prior steps. Through constructing formalized labels for training, the LLM can learn to generate desired and reasonable chain-of-thought outputs for improving the action decision. Experimental results across various training settings and popular VLN benchmarks (e.g., Room-to-Room (R2R), Room-across-Room (RxR), Room-for-Room (R4R)) show the significant superiority of NavCoT over the direct action prediction variants. Through simple parameter-efficient finetuning, our NavCoT outperforms a recent GPT4-based approach with ~7% relative improvement on the R2R dataset. We believe that NavCoT will help unlock more task-adaptive and scalable LLM-based embodied agents, which are helpful for developing real-world robotics applications. Code is available at https://github.com/expectorlin/NavCoT.

Kun Xiang, Zhili Liu, Zihao Jiang et al.

In this paper, we address the challenging task of multimodal mathematical reasoning by incorporating the ability of "slow thinking" into multimodal large language models (MLLMs). Our core idea is that different levels of reasoning abilities can be combined dynamically to tackle questions with different complexity. To this end, we propose a paradigm of Self-structured Chain of Thought (SCoT), which is composed of minimal semantic atomic steps. Different from existing methods that rely on structured templates or free-form paradigms, our method can not only generate cognitive CoT structures for various complex tasks but also mitigates the phenomenon of overthinking. To introduce structured reasoning capabilities into visual understanding models, we further design a novel AtomThink framework with four key modules, including (i) a data engine to generate high-quality multimodal reasoning paths; (ii) a supervised fine-tuning process with serialized inference data; (iii) a policy-guided multi-turn inference method; and (iv) an atomic capability metric to evaluate the single step utilization rate. We conduct extensive experiments to show that the proposed AtomThink significantly improves the performance of baseline MLLMs, achieving more than 10\% average accuracy gains on MathVista and MathVerse. Compared to state-of-the-art structured CoT approaches, our method not only achieves higher accuracy but also improves data utilization by 5 times and boosts inference efficiency by 85.3\%. Our code is now public available in https://github.com/Quinn777/AtomThink.

Xiuwei Chen, Wentao Hu, Hanhui Li et al.

Recent advances in multimodal large language models (MLLMs) have shown impressive reasoning capabilities. However, further enhancing existing MLLMs necessitates high-quality vision-language datasets with carefully curated task complexities, which are both costly and challenging to scale. Although recent self-improving models that iteratively refine themselves offer a feasible solution, they still suffer from two core challenges: (i) most existing methods augment visual or textual data separately, resulting in discrepancies in data complexity (e.g., over-simplified diagrams paired with redundant textual descriptions); and (ii) the evolution of data and models is also separated, leading to scenarios where models are exposed to tasks with mismatched difficulty levels. To address these issues, we propose C2-Evo, an automatic, closed-loop self-improving framework that jointly evolves both training data and model capabilities. Specifically, given a base dataset and a base model, C2-Evo enhances them by a cross-modal data evolution loop and a data-model evolution loop. The former loop expands the base dataset by generating complex multimodal problems that combine structured textual sub-problems with iteratively specified geometric diagrams, while the latter loop adaptively selects the generated problems based on the performance of the base model, to conduct supervised fine-tuning and reinforcement learning alternately. Consequently, our method continuously refines its model and training data, and consistently obtains considerable performance gains across multiple mathematical reasoning benchmarks. Our code, models, and datasets will be released.

Kun Xiang, Zhili Liu, Terry Jingchen Zhang et al.

In this paper, we address the challenging task of multimodal mathematical reasoning by incorporating the notion of ``slow thinking'' into multimodal large language models (MLLMs). Our core idea is that models can learn to adaptively use different levels of reasoning to tackle questions of different complexity. We propose a novel paradigm of Self-structured Chain of Thought (SCoT), which comprises of minimal semantic atomic steps. Different from existing methods that rely on structured templates or free-form paradigms, our method can not only generate cognitive CoT structures for various complex tasks but also mitigates the phenomena of overthinking for easier tasks. To introduce structured reasoning into visual cognition, we further design a novel AtomThink framework with four key modules, including (i) a data engine to generate high-quality multimodal reasoning paths; (ii) a supervised fine-tuning (SFT) process with serialized inference data; (iii) a policy-guided multi-turn inference method; and (iv) an atomic capability metric to evaluate the single step utilization rate. We conduct extensive experiments to show that the proposed AtomThink significantly improves the performance of baseline MLLMs, achieving more than 10\% average accuracy gains on MathVista and MathVerse. Compared to state-of-the-art structured CoT approaches, our method not only achieves higher accuracy but also improves data utilization by 5 times and boosts inference efficiency by 85.3\%. Our code is now public available in https://github.com/Quinn777/AtomThink.

Jiahui Gao, Renjie Pi, Jipeng Zhang et al.

Large language models (LLMs) have shown remarkable proficiency in human-level reasoning and generation capabilities, which encourages extensive research on their application in mathematical problem solving. However, current work has been largely focused on text-based mathematical problems, with limited investigation in problems involving geometric information. Addressing this gap, we aim to enable LLMs to solve geometric problems by understanding image input. We first analyze the limitations of current Multimodal Large Language Models (MLLMs) in this area: they struggle to accurately comprehending basic geometric elements and their relationships. To overcome these challenges, we take advantage of the unique characteristics of geometric problems (such as unique geometric logical form, and geometric scalability) and the capacity of the textual LLMs to build an enriched multimodal geometry dataset based on existing data. The augmented dataset, Geo170K, contains more than 170K geometric image-caption and question-answer pairs. Utilizing our constructed Geo170K dataset, we develop G-LLaVA, which demonstrates exceptional performance in solving geometric problems, significantly outperforming GPT-4-V on the MathVista benchmark with only 7B parameters.

Kun Xiang, Terry Jingchen Zhang, Yinya Huang et al.

The rapid advancement of embodied intelligence and world models has intensified efforts to integrate physical laws into AI systems, yet physical perception and symbolic physics reasoning have developed along separate trajectories without a unified bridging framework. This work provides a comprehensive overview of physical AI, establishing clear distinctions between theoretical physics reasoning and applied physical understanding while systematically examining how physics-grounded methods enhance AI's real-world comprehension across structured symbolic reasoning, embodied systems, and generative models. Through rigorous analysis of recent advances, we advocate for intelligent systems that ground learning in both physical principles and embodied reasoning processes, transcending pattern recognition toward genuine understanding of physical laws. Our synthesis envisions next-generation world models capable of explaining physical phenomena and predicting future states, advancing safe, generalizable, and interpretable AI systems. We maintain a continuously updated resource at https://github.com/AI4Phys/Awesome-AI-for-Physics.

Bingqian Lin, Yunshuang Nie, Khun Loun Zai et al.

Recent studies have revealed the potential of training open-source Large Language Models (LLMs) to unleash LLMs' reasoning ability for enhancing vision-language navigation (VLN) performance, and simultaneously mitigate the domain gap between LLMs' training corpus and the VLN task. However, these approaches predominantly adopt straightforward input-output mapping paradigms, causing the mapping learning difficult and the navigational decisions unexplainable. Chain-of-Thought (CoT) training is a promising way to improve both navigational decision accuracy and interpretability, while the complexity of the navigation task makes the perfect CoT labels unavailable and may lead to overfitting through pure CoT supervised fine-tuning. To address these issues, we propose EvolveNav, a novel sElf-improving embodied reasoning paradigm that realizes adaptable and generalizable navigational reasoning for boosting LLM-based vision-language Navigation. Specifically, EvolveNav involves a two-stage training process: (1) Formalized CoT Supervised Fine-Tuning, where we train the model with curated formalized CoT labels to first activate the model's navigational reasoning capabilities, and simultaneously increase the reasoning speed; (2) Self-Reflective Post-Training, where the model is iteratively trained with its own reasoning outputs as self-enriched CoT labels to enhance the supervision diversity. A self-reflective auxiliary task is also designed to encourage the model to learn correct reasoning patterns by contrasting with wrong ones. Experimental results under both task-specific and cross-task training paradigms demonstrate the consistent superiority of EvolveNav over previous LLM-based VLN approaches on various popular benchmarks, including R2R, REVERIE, CVDN, and SOON. Code is available at https://github.com/expectorlin/EvolveNav.

Lewei Yao, Renjie Pi, Jianhua Han et al.

Existing open-vocabulary object detectors typically require a predefined set of categories from users, significantly confining their application scenarios. In this paper, we introduce DetCLIPv3, a high-performing detector that excels not only at both open-vocabulary object detection, but also generating hierarchical labels for detected objects. DetCLIPv3 is characterized by three core designs: 1. Versatile model architecture: we derive a robust open-set detection framework which is further empowered with generation ability via the integration of a caption head. 2. High information density data: we develop an auto-annotation pipeline leveraging visual large language model to refine captions for large-scale image-text pairs, providing rich, multi-granular object labels to enhance the training. 3. Efficient training strategy: we employ a pre-training stage with low-resolution inputs that enables the object captioner to efficiently learn a broad spectrum of visual concepts from extensive image-text paired data. This is followed by a fine-tuning stage that leverages a small number of high-resolution samples to further enhance detection performance. With these effective designs, DetCLIPv3 demonstrates superior open-vocabulary detection performance, \eg, our Swin-T backbone model achieves a notable 47.0 zero-shot fixed AP on the LVIS minival benchmark, outperforming GLIPv2, GroundingDINO, and DetCLIPv2 by 18.0/19.6/6.6 AP, respectively. DetCLIPv3 also achieves a state-of-the-art 19.7 AP in dense captioning task on VG dataset, showcasing its strong generative capability.

Chunwei Wang, Guansong Lu, Junwei Yang et al.

In this paper, we introduce ILLUME, a unified multimodal large language model (MLLM) that seamlessly integrates multimodal understanding and generation capabilities within a single large language model through a unified next-token prediction formulation. To address the large dataset size typically required for image-text alignment, we propose to enhance data efficiency through the design of a vision tokenizer that incorporates semantic information and a progressive multi-stage training procedure. This approach reduces the dataset size to just 15M for pretraining -- over four times fewer than what is typically needed -- while achieving competitive or even superior performance with existing unified MLLMs, such as Janus. Additionally, to promote synergistic enhancement between understanding and generation capabilities, which is under-explored in previous works, we introduce a novel self-enhancing multimodal alignment scheme. This scheme supervises the MLLM to self-assess the consistency between text descriptions and self-generated images, facilitating the model to interpret images more accurately and avoid unrealistic and incorrect predictions caused by misalignment in image generation. Based on extensive experiments, our proposed ILLUME stands out and competes with state-of-the-art unified MLLMs and specialized models across various benchmarks for multimodal understanding, generation, and editing.

Zhili Liu, Kai Chen, Jianhua Han et al.

Masked Autoencoder~(MAE) is a prevailing self-supervised learning method that achieves promising results in model pre-training. However, when the various downstream tasks have data distributions different from the pre-training data, the semantically irrelevant pre-training information might result in negative transfer, impeding MAE's scalability. To address this issue, we propose a novel MAE-based pre-training paradigm, Mixture of Cluster-conditional Experts (MoCE), which can be trained once but provides customized pre-training models for diverse downstream tasks. Different from the mixture of experts (MoE), our MoCE trains each expert only with semantically relevant images by using cluster-conditional gates. Thus, each downstream task can be allocated to its customized model pre-trained with data most similar to the downstream data. Experiments on a collection of 11 downstream tasks show that MoCE outperforms the vanilla MAE by 2.45\% on average. It also obtains new state-of-the-art self-supervised learning results on detection and segmentation.

Youpeng Wen, Junfan Lin, Yi Zhu et al.

Recent advancements utilizing large-scale video data for learning video generation models demonstrate significant potential in understanding complex physical dynamics. It suggests the feasibility of leveraging diverse robot trajectory data to develop a unified, dynamics-aware model to enhance robot manipulation. However, given the relatively small amount of available robot data, directly fitting data without considering the relationship between visual observations and actions could lead to suboptimal data utilization. To this end, we propose VidMan (Video Diffusion for Robot Manipulation), a novel framework that employs a two-stage training mechanism inspired by dual-process theory from neuroscience to enhance stability and improve data utilization efficiency. Specifically, in the first stage, VidMan is pre-trained on the Open X-Embodiment dataset (OXE) for predicting future visual trajectories in a video denoising diffusion manner, enabling the model to develop a long horizontal awareness of the environment's dynamics. In the second stage, a flexible yet effective layer-wise self-attention adapter is introduced to transform VidMan into an efficient inverse dynamics model that predicts action modulated by the implicit dynamics knowledge via parameter sharing. Our VidMan framework outperforms state-of-the-art baseline model GR-1 on the CALVIN benchmark, achieving a 11.7% relative improvement, and demonstrates over 9% precision gains on the OXE small-scale dataset. These results provide compelling evidence that world models can significantly enhance the precision of robot action prediction. Codes and models will be public.

Runhui Huang, Chunwei Wang, Junwei Yang et al.

We present ILLUME+ that leverages dual visual tokenization and a diffusion decoder to improve both deep semantic understanding and high-fidelity image generation. Existing unified models have struggled to simultaneously handle the three fundamental capabilities in a unified model: understanding, generation, and editing. Models like Chameleon and EMU3 utilize VQGAN for image discretization, due to the lack of deep semantic interaction, they lag behind specialist models like LLaVA in visual understanding tasks. To mitigate this, LaViT and ILLUME employ semantic encoders for tokenization, but they struggle with image editing due to poor texture preservation. Meanwhile, Janus series decouples the input and output image representation, limiting their abilities to seamlessly handle interleaved image-text understanding and generation. In contrast, ILLUME+ introduces a unified dual visual tokenizer, DualViTok, which preserves both fine-grained textures and text-aligned semantics while enabling a coarse-to-fine image representation strategy for multimodal understanding and generation. Additionally, we employ a diffusion model as the image detokenizer for enhanced generation quality and efficient super-resolution. ILLUME+ follows a continuous-input, discrete-output scheme within the unified MLLM and adopts a progressive training procedure that supports dynamic resolution across the vision tokenizer, MLLM, and diffusion decoder. This design allows for flexible and efficient context-aware image editing and generation across diverse tasks. ILLUME+ (3B) exhibits competitive performance against existing unified MLLMs and specialized models across multimodal understanding, generation, and editing benchmarks. With its strong performance, ILLUME+ provides a scalable and versatile foundation for future multimodal applications. Project Page: https://illume-unified-mllm.github.io/.

Zisheng Chen, Chunwei Wang, Xiuwei Chen et al.

In this paper, we introduce SemHiTok, a unified image Tokenizer via Semantic-Guided Hierarchical codebook that provides consistent discrete representations for multimodal understanding and generation. Recently, unified image tokenizers have sparked exploration within research community, which is designed to capture high-level semantic features for understanding and retaining low-level pixel features for generation. Previous works attempt to train a unified image tokenizer by combining loss for semantic distillation and pixel reconstruction. However, due to the differing levels of features prioritized by multimodal understanding and generation, joint training methods face significant challenges in achieving a good trade-off. SemHiTok addresses this challenge through a novel semantic-guided hierarchical codebook, which builds pixel sub-codebooks on a pretrained semantic codebook. This design decouples semantic and pixel both in terms of structure and training strategy, enabling the tokenizer to capture pixel features while retaining its ability to comprehend high-level semantic information. Our experiments demonstrate that SemHiTok achieves SOTA performance in image reconstruction and multimodal understanding under LLaVA-v1.5 setting. Further, we develop a unified MLLM with SemHiTok, which exhibits superior performance across multimodal understanding and generation tasks. For understanding, SemHiTok achieves impressive performance on most benchmarks. For generation, our model achieves SOTA performance on MJHQ30K in unified MLLMs.

Kun Xiang, Heng Li, Terry Jingchen Zhang et al.

We present SeePhys, a large-scale multimodal benchmark for LLM reasoning grounded in physics questions ranging from middle school to PhD qualifying exams. The benchmark covers 7 fundamental domains spanning the physics discipline, incorporating 21 categories of highly heterogeneous diagrams. In contrast to prior works where visual elements mainly serve auxiliary purposes, our benchmark features a substantial proportion of vision-essential problems (75%) that mandate visual information extraction for correct solutions. Through extensive evaluation, we observe that even the most advanced visual reasoning models (e.g., Gemini-2.5-pro and o4-mini) achieve sub-60% accuracy on our benchmark. These results reveal fundamental challenges in current large language models' visual understanding capabilities, particularly in: (i) establishing rigorous coupling between diagram interpretation and physics reasoning, and (ii) overcoming their persistent reliance on textual cues as cognitive shortcuts.

Runhui Huang, Kaixin Cai, Jianhua Han et al.

Despite the success of generating high-quality images given any text prompts by diffusion-based generative models, prior works directly generate the entire images, but cannot provide object-wise manipulation capability. To support wider real applications like professional graphic design and digital artistry, images are frequently created and manipulated in multiple layers to offer greater flexibility and control. Therefore in this paper, we propose a layer-collaborative diffusion model, named LayerDiff, specifically designed for text-guided, multi-layered, composable image synthesis. The composable image consists of a background layer, a set of foreground layers, and associated mask layers for each foreground element. To enable this, LayerDiff introduces a layer-based generation paradigm incorporating multiple layer-collaborative attention modules to capture inter-layer patterns. Specifically, an inter-layer attention module is designed to encourage information exchange and learning between layers, while a text-guided intra-layer attention module incorporates layer-specific prompts to direct the specific-content generation for each layer. A layer-specific prompt-enhanced module better captures detailed textual cues from the global prompt. Additionally, a self-mask guidance sampling strategy further unleashes the model's ability to generate multi-layered images. We also present a pipeline that integrates existing perceptual and generative models to produce a large dataset of high-quality, text-prompted, multi-layered images. Extensive experiments demonstrate that our LayerDiff model can generate high-quality multi-layered images with performance comparable to conventional whole-image generation methods. Moreover, LayerDiff enables a broader range of controllable generative applications, including layer-specific image editing and style transfer.

Yulong Liu, Yunlong Yuan, Chunwei Wang et al.

The distinction between humans and animals lies in the unique ability of humans to use and create tools. Tools empower humans to overcome physiological limitations, fostering the creation of magnificent civilizations. Similarly, enabling foundational models like Large Language Models (LLMs) with the capacity to learn external tool usage may serve as a pivotal step toward realizing artificial general intelligence. Previous studies in this field have predominantly pursued two distinct approaches to augment the tool invocation capabilities of LLMs. The first approach emphasizes the construction of relevant datasets for model fine-tuning. The second approach, in contrast, aims to fully exploit the inherent reasoning abilities of LLMs through in-context learning strategies. In this work, we introduce a novel tool invocation pipeline designed to control massive real-world APIs. This pipeline mirrors the human task-solving process, addressing complicated real-life user queries. At each step, we guide LLMs to summarize the achieved results and determine the next course of action. We term this pipeline `from Summary to action', Sum2Act for short. Empirical evaluations of our Sum2Act pipeline on the ToolBench benchmark show significant performance improvements, outperforming established methods like ReAct and DFSDT. This highlights Sum2Act's effectiveness in enhancing LLMs for complex real-world tasks.