Jun Xiao

Kexin Li, Zongxin Yang, Lei Chen et al.

Audio-visual video segmentation~(AVVS) aims to generate pixel-level maps of sound-producing objects within image frames and ensure the maps faithfully adhere to the given audio, such as identifying and segmenting a singing person in a video. However, existing methods exhibit two limitations: 1) they address video temporal features and audio-visual interactive features separately, disregarding the inherent spatial-temporal dependence of combined audio and video, and 2) they inadequately introduce audio constraints and object-level information during the decoding stage, resulting in segmentation outcomes that fail to comply with audio directives. To tackle these issues, we propose a decoupled audio-video transformer that combines audio and video features from their respective temporal and spatial dimensions, capturing their combined dependence. To optimize memory consumption, we design a block, which, when stacked, enables capturing audio-visual fine-grained combinatorial-dependence in a memory-efficient manner. Additionally, we introduce audio-constrained queries during the decoding phase. These queries contain rich object-level information, ensuring the decoded mask adheres to the sounds. Experimental results confirm our approach's effectiveness, with our framework achieving a new SOTA performance on all three datasets using two backbones. The code is available at \url{https://github.com/aspirinone/CATR.github.io}

Yuxi Wang, Jian Liang, Jun Xiao et al.

Contemporary domain adaptation offers a practical solution for achieving cross-domain transfer of semantic segmentation between labeled source data and unlabeled target data. These solutions have gained significant popularity; however, they require the model to be retrained when the test environment changes. This can result in unbearable costs in certain applications due to the time-consuming training process and concerns regarding data privacy. One-shot domain adaptation methods attempt to overcome these challenges by transferring the pre-trained source model to the target domain using only one target data. Despite this, the referring style transfer module still faces issues with computation cost and over-fitting problems. To address this problem, we propose a novel framework called Informative Data Mining (IDM) that enables efficient one-shot domain adaptation for semantic segmentation. Specifically, IDM provides an uncertainty-based selection criterion to identify the most informative samples, which facilitates quick adaptation and reduces redundant training. We then perform a model adaptation method using these selected samples, which includes patch-wise mixing and prototype-based information maximization to update the model. This approach effectively enhances adaptation and mitigates the overfitting problem. In general, we provide empirical evidence of the effectiveness and efficiency of IDM. Our approach outperforms existing methods and achieves a new state-of-the-art one-shot performance of 56.7\%/55.4\% on the GTA5/SYNTHIA to Cityscapes adaptation tasks, respectively. The code will be released at \url{https://github.com/yxiwang/IDM}.

Kaifeng Gao, Long Chen, Hanwang Zhang et al.

Prompt tuning with large-scale pretrained vision-language models empowers open-vocabulary predictions trained on limited base categories, e.g., object classification and detection. In this paper, we propose compositional prompt tuning with motion cues: an extended prompt tuning paradigm for compositional predictions of video data. In particular, we present Relation Prompt (RePro) for Open-vocabulary Video Visual Relation Detection (Open-VidVRD), where conventional prompt tuning is easily biased to certain subject-object combinations and motion patterns. To this end, RePro addresses the two technical challenges of Open-VidVRD: 1) the prompt tokens should respect the two different semantic roles of subject and object, and 2) the tuning should account for the diverse spatio-temporal motion patterns of the subject-object compositions. Without bells and whistles, our RePro achieves a new state-of-the-art performance on two VidVRD benchmarks of not only the base training object and predicate categories, but also the unseen ones. Extensive ablations also demonstrate the effectiveness of the proposed compositional and multi-mode design of prompts. Code is available at https://github.com/Dawn-LX/OpenVoc-VidVRD.

Qiming Yang, Kai Zhang, Chaoxiang Lan et al.

Solid results from Transformers have made them prevailing architectures in various natural language and vision tasks. As a default component in Transformers, Layer Normalization (LN) normalizes activations within each token to boost the robustness. However, LN requires on-the-fly statistics calculation in inference as well as division and square root operations, leading to inefficiency on hardware. What is more, replacing LN with other hardware-efficient normalization schemes (e.g., Batch Normalization) results in inferior performance, even collapse in training. We find that this dilemma is caused by abnormal behaviors of activation statistics, including large fluctuations over iterations and extreme outliers across layers. To tackle these issues, we propose Unified Normalization (UN), which can speed up the inference by being fused with other linear operations and achieve comparable performance on par with LN. UN strives to boost performance by calibrating the activation and gradient statistics with a tailored fluctuation smoothing strategy. Meanwhile, an adaptive outlier filtration strategy is applied to avoid collapse in training whose effectiveness is theoretically proved and experimentally verified in this paper. We demonstrate that UN can be an efficient drop-in alternative to LN by conducting extensive experiments on language and vision tasks. Besides, we evaluate the efficiency of our method on GPU. Transformers equipped with UN enjoy about 31% inference speedup and nearly 18% memory reduction. Code will be released at https://github.com/hikvision-research/Unified-Normalization.

Lin Li, Jun Xiao, Hanrong Shi et al.

Nearly all existing scene graph generation (SGG) models have overlooked the ground-truth annotation qualities of mainstream SGG datasets, i.e., they assume: 1) all the manually annotated positive samples are equally correct; 2) all the un-annotated negative samples are absolutely background. In this paper, we argue that neither of the assumptions applies to SGG: there are numerous noisy ground-truth predicate labels that break these two assumptions and harm the training of unbiased SGG models. To this end, we propose a novel NoIsy label CorrEction and Sample Training strategy for SGG: NICEST. Specifically, it consists of two parts: NICE and NIST, which rule out these noisy label issues by generating high-quality samples and the effective training strategy, respectively. NICE first detects noisy samples and then reassigns them more high-quality soft predicate labels. NIST is a multi-teacher knowledge distillation based training strategy, which enables the model to learn unbiased fusion knowledge. And a dynamic trade-off weighting strategy in NIST is designed to penalize the bias of different teachers. Due to the model-agnostic nature of both NICE and NIST, our NICEST can be seamlessly incorporated into any SGG architecture to boost its performance on different predicate categories. In addition, to better evaluate the generalization of SGG models, we further propose a new benchmark VG-OOD, by re-organizing the prevalent VG dataset and deliberately making the predicate distributions of the training and test sets as different as possible for each subject-object category pair. This new benchmark helps disentangle the influence of subject-object category based frequency biases. Extensive ablations and results on different backbones and tasks have attested to the effectiveness and generalization ability of each component of NICEST.

Yihan Xie, Sijing Li, Tianwei Lin et al.

Although electrocardiograms (ECG) play a dominant role in cardiovascular diagnosis and treatment, their intrinsic data forms and representational patterns pose significant challenges for medical multimodal large language models (Med-MLLMs) in achieving cross-modal semantic alignment. To address this gap, we propose Heartcare Suite, a unified ECG suite designed for dual signal-image modeling and understanding: (i) Heartcare-400K. A fine-grained ECG instruction dataset on top of our data pipeline engine--HeartAgent--by integrating high quality clinical ECG reports from top hospitals with open-source data. (ii) Heartcare-Bench. A systematic benchmark assessing performance of models in multi-perspective ECG understanding and cross-modal generalization, providing guidance for optimizing ECG comprehension models. (iii) HeartcareGPT. Built upon a structure-aware discrete tokenizer Beat, we propose Dual Stream Projection Alignment (DSPA) paradigm--a dual encoder projection alignment mechanism enabling joint optimizing and modeling native ECG signal-image within a shared feature space. HeartcareGPT achieves consistent improvements across diverse ECG understanding tasks, validating both the effectiveness of the unified modeling paradigm and the necessity of a high-quality data pipeline, and establishing a methodological foundation for extending Med-MLLMs towards physiological signal domains. Our project is available at https://github.com/ZJU4HealthCare/HeartcareGPT .

Wenqing Wang, Yawei Luo, Zhiqing Chen et al.

Current video-based scene graph generation (VidSGG) methods have been found to perform poorly on predicting predicates that are less represented due to the inherent biased distribution in the training data. In this paper, we take a closer look at the predicates and identify that most visual relations (e.g. sit_above) involve both actional pattern (sit) and spatial pattern (above), while the distribution bias is much less severe at the pattern level. Based on this insight, we propose a decoupled label learning (DLL) paradigm to address the intractable visual relation prediction from the pattern-level perspective. Specifically, DLL decouples the predicate labels and adopts separate classifiers to learn actional and spatial patterns respectively. The patterns are then combined and mapped back to the predicate. Moreover, we propose a knowledge-level label decoupling method to transfer non-target knowledge from head predicates to tail predicates within the same pattern to calibrate the distribution of tail classes. We validate the effectiveness of DLL on the commonly used VidSGG benchmark, i.e. VidVRD. Extensive experiments demonstrate that the DLL offers a remarkably simple but highly effective solution to the long-tailed problem, achieving the state-of-the-art VidSGG performance.

Michal Nazarczuk, Sibi Catley-Chandar, Thomas Tanay et al.

This paper reviews the challenge on Sparse Neural Rendering that was part of the Advances in Image Manipulation (AIM) workshop, held in conjunction with ECCV 2024. This manuscript focuses on the competition set-up, the proposed methods and their respective results. The challenge aims at producing novel camera view synthesis of diverse scenes from sparse image observations. It is composed of two tracks, with differing levels of sparsity; 3 views in Track 1 (very sparse) and 9 views in Track 2 (sparse). Participants are asked to optimise objective fidelity to the ground-truth images as measured via the Peak Signal-to-Noise Ratio (PSNR) metric. For both tracks, we use the newly introduced Sparse Rendering (SpaRe) dataset and the popular DTU MVS dataset. In this challenge, 5 teams submitted final results to Track 1 and 4 teams submitted final results to Track 2. The submitted models are varied and push the boundaries of the current state-of-the-art in sparse neural rendering. A detailed description of all models developed in the challenge is provided in this paper.

Lin Li, Long Chen, Yifeng Huang et al.

Unbiased SGG has achieved significant progress over recent years. However, almost all existing SGG models have overlooked the ground-truth annotation qualities of prevailing SGG datasets, i.e., they always assume: 1) all the manually annotated positive samples are equally correct; 2) all the un-annotated negative samples are absolutely background. In this paper, we argue that both assumptions are inapplicable to SGG: there are numerous "noisy" groundtruth predicate labels that break these two assumptions, and these noisy samples actually harm the training of unbiased SGG models. To this end, we propose a novel model-agnostic NoIsy label CorrEction strategy for SGG: NICE. NICE can not only detect noisy samples but also reassign more high-quality predicate labels to them. After the NICE training, we can obtain a cleaner version of SGG dataset for model training. Specifically, NICE consists of three components: negative Noisy Sample Detection (Neg-NSD), positive NSD (Pos-NSD), and Noisy Sample Correction (NSC). Firstly, in Neg-NSD, we formulate this task as an out-of-distribution detection problem, and assign pseudo labels to all detected noisy negative samples. Then, in Pos-NSD, we use a clustering-based algorithm to divide all positive samples into multiple sets, and treat the samples in the noisiest set as noisy positive samples. Lastly, in NSC, we use a simple but effective weighted KNN to reassign new predicate labels to noisy positive samples. Extensive results on different backbones and tasks have attested to the effectiveness and generalization abilities of each component of NICE.

Jiacong Wang, Bohong Wu, Haiyong Jiang et al.

Recent advances in Vision-Language Models (VLMs) and the scarcity of high-quality multi-modal alignment data have inspired numerous researches on synthetic VLM data generation. The conventional norm in VLM data construction uses a mixture of specialists in caption and OCR, or stronger VLM APIs and expensive human annotation. In this paper, we present World to Code (W2C), a meticulously curated multi-modal data construction pipeline that organizes the final generation output into a Python code format. The pipeline leverages the VLM itself to extract cross-modal information via different prompts and filter the generated outputs again via a consistency filtering strategy. Experiments have demonstrated the high quality of W2C by improving various existing visual question answering and visual grounding benchmarks across different VLMs. Further analysis also demonstrates that the new code parsing ability of VLMs presents better cross-modal equivalence than the commonly used detail caption ability. Our code is available at https://github.com/foundation-multimodal-models/World2Code.

Wenqing Wang, Kaifeng Gao, Yawei Luo et al.

Video-based scene graph generation (VidSGG) is an approach that aims to represent video content in a dynamic graph by identifying visual entities and their relationships. Due to the inherently biased distribution and missing annotations in the training data, current VidSGG methods have been found to perform poorly on less-represented predicates. In this paper, we propose an explicit solution to address this under-explored issue by supplementing missing predicates that should be appear in the ground-truth annotations. Dubbed Trico, our method seeks to supplement the missing predicates by exploring three complementary spatio-temporal correlations. Guided by these correlations, the missing labels can be effectively supplemented thus achieving an unbiased predicate predictions. We validate the effectiveness of Trico on the most widely used VidSGG datasets, i.e., VidVRD and VidOR. Extensive experiments demonstrate the state-of-the-art performance achieved by Trico, particularly on those tail predicates.

Feifei Shao, Yawei Luo, Fei Gao et al.

Previous weakly-supervised object localization (WSOL) methods aim to expand activation map discriminative areas to cover the whole objects, yet neglect two inherent challenges when relying solely on image-level labels. First, the ``entangled context'' issue arises from object-context co-occurrence (\eg, fish and water), making the model inspection hard to distinguish object boundaries clearly. Second, the ``C-L dilemma'' issue results from the information decay caused by the pooling layers, which struggle to retain both the semantic information for precise classification and those essential details for accurate localization, leading to a trade-off in performance. In this paper, we propose a knowledge-guided causal intervention method, dubbed KG-CI-CAM, to address these two under-explored issues in one go. More specifically, we tackle the co-occurrence context confounder problem via causal intervention, which explores the causalities among image features, contexts, and categories to eliminate the biased object-context entanglement in the class activation maps. Based on the disentangled object feature, we introduce a multi-source knowledge guidance framework to strike a balance between absorbing classification knowledge and localization knowledge during model training. Extensive experiments conducted on several benchmark datasets demonstrate the effectiveness of KG-CI-CAM in learning distinct object boundaries amidst confounding contexts and mitigating the dilemma between classification and localization performance.

Fei Tang, Zhiqiong Lu, Boxuan Zhang et al.

GUI agents drive applications through their visual interfaces instead of programmatic APIs, interacting with arbitrary software via taps, swipes, and keystrokes, reaching a long tail of applications that CLI-based agents cannot. Yet progress in this area is bottlenecked less by modeling capacity than by the absence of a coherent full-stack infrastructure: online RL training suffers from environment instability and closed pipelines, evaluation protocols drift silently across works, and trained agents rarely reach real users on real devices. We present \textbf{ClawGUI}, an open-source framework addressing these three gaps within a single harness. \textbf{ClawGUI-RL} provides the first open-source GUI agent RL infrastructure with validated support for both parallel virtual environments and real physical devices, integrating GiGPO with a Process Reward Model for dense step-level supervision. \textbf{ClawGUI-Eval} enforces a fully standardized evaluation pipeline across 6 benchmarks and 11+ models, achieving 95.8\% reproduction against official baselines. \textbf{ClawGUI-Agent} brings trained agents to Android, HarmonyOS, and iOS through 12+ chat platforms with hybrid CLI-GUI control and persistent personalized memory. Trained end to end within this pipeline, \textbf{ClawGUI-2B} achieves 17.1\% Success Rate on MobileWorld GUI-Only, outperforming the same-scale MAI-UI-2B baseline by 6.0\%.

Long Chen, Yuhang Zheng, Jun Xiao

Data Augmentation (DA) -- generating extra training samples beyond original training set -- has been widely-used in today's unbiased VQA models to mitigate the language biases. Current mainstream DA strategies are synthetic-based methods, which synthesize new samples by either editing some visual regions/words, or re-generating them from scratch. However, these synthetic samples are always unnatural and error-prone. To avoid this issue, a recent DA work composes new augmented samples by randomly pairing pristine images and other human-written questions. Unfortunately, to guarantee augmented samples have reasonable ground-truth answers, they manually design a set of heuristic rules for several question types, which extremely limits its generalization abilities. To this end, we propose a new Knowledge Distillation based Data Augmentation for VQA, dubbed KDDAug. Specifically, we first relax the requirements of reasonable image-question pairs, which can be easily applied to any question types. Then, we design a knowledge distillation (KD) based answer assignment to generate pseudo answers for all composed image-question pairs, which are robust to both in-domain and out-of-distribution settings. Since KDDAug is a model-agnostic DA strategy, it can be seamlessly incorporated into any VQA architectures. Extensive ablation studies on multiple backbones and benchmarks have demonstrated the effectiveness and generalization abilities of KDDAug.

Yulei Lu, Yawei Luo, Li Zhang et al.

A thriving trend for domain adaptive segmentation endeavors to generate the high-quality pseudo labels for target domain and retrain the segmentor on them. Under this self-training paradigm, some competitive methods have sought to the latent-space information, which establishes the feature centroids (a.k.a prototypes) of the semantic classes and determines the pseudo label candidates by their distances from these centroids. In this paper, we argue that the latent space contains more information to be exploited thus taking one step further to capitalize on it. Firstly, instead of merely using the source-domain prototypes to determine the target pseudo labels as most of the traditional methods do, we bidirectionally produce the target-domain prototypes to degrade those source features which might be too hard or disturbed for the adaptation. Secondly, existing attempts simply model each category as a single and isotropic prototype while ignoring the variance of the feature distribution, which could lead to the confusion of similar categories. To cope with this issue, we propose to represent each category with multiple and anisotropic prototypes via Gaussian Mixture Model, in order to fit the de facto distribution of source domain and estimate the likelihood of target samples based on the probability density. We apply our method on GTA5->Cityscapes and Synthia->Cityscapes tasks and achieve 61.2 and 62.8 respectively in terms of mean IoU, substantially outperforming other competitive self-training methods. Noticeably, in some categories which severely suffer from the categorical confusion such as "truck" and "bus", our method achieves 56.4 and 68.8 respectively, which further demonstrates the effectiveness of our design.

Lin Li, Long Chen, Hanrong Shi et al.

The Scene Graph Generation (SGG) task aims to detect all the objects and their pairwise visual relationships in a given image. Although SGG has achieved remarkable progress over the last few years, almost all existing SGG models follow the same training paradigm: they treat both object and predicate classification in SGG as a single-label classification problem, and the ground-truths are one-hot target labels. However, this prevalent training paradigm has overlooked two characteristics of current SGG datasets: 1) For positive samples, some specific subject-object instances may have multiple reasonable predicates. 2) For negative samples, there are numerous missing annotations. Regardless of the two characteristics, SGG models are easy to be confused and make wrong predictions. To this end, we propose a novel model-agnostic Label Semantic Knowledge Distillation (LS-KD) for unbiased SGG. Specifically, LS-KD dynamically generates a soft label for each subject-object instance by fusing a predicted Label Semantic Distribution (LSD) with its original one-hot target label. LSD reflects the correlations between this instance and multiple predicate categories. Meanwhile, we propose two different strategies to predict LSD: iterative self-KD and synchronous self-KD. Extensive ablations and results on three SGG tasks have attested to the superiority and generality of our proposed LS-KD, which can consistently achieve decent trade-off performance between different predicate categories.

Lin Li, Guikun Chen, Jun Xiao et al.

Scene Graph Generation (SGG) aims to detect all the visual relation triplets $<$\texttt{sub}, \texttt{pred}, \texttt{obj}$>$ in a given image. With the emergence of various advanced techniques for better utilizing both the intrinsic and extrinsic information in each relation triplet, SGG has achieved great progress over the recent years. However, due to the ubiquitous long-tailed predicate distributions, today's SGG models are still easily biased to the head predicates. Currently, the most prevalent debiasing solutions for SGG are re-balancing methods, \eg, changing the distributions of original training samples. In this paper, we argue that all existing re-balancing strategies fail to increase the diversity of the relation triplet features of each predicate, which is critical for robust SGG. To this end, we propose a novel Compositional Feature Augmentation (\textbf{CFA}) strategy, which is the first unbiased SGG work to mitigate the bias issue from the perspective of increasing the diversity of triplet features. Specifically, we first decompose each relation triplet feature into two components: intrinsic feature and extrinsic feature, which correspond to the intrinsic characteristics and extrinsic contexts of a relation triplet, respectively. Then, we design two different feature augmentation modules to enrich the feature diversity of original relation triplets by replacing or mixing up either their intrinsic or extrinsic features from other samples. Due to its model-agnostic nature, CFA can be seamlessly incorporated into various SGG frameworks. Extensive ablations have shown that CFA achieves a new state-of-the-art performance on the trade-off between different metrics.

Xingchen Li, Long Chen, Wenbo Ma et al.

Recently, increasing efforts have been focused on Weakly Supervised Scene Graph Generation (WSSGG). The mainstream solution for WSSGG typically follows the same pipeline: they first align text entities in the weak image-level supervisions (e.g., unlocalized relation triplets or captions) with image regions, and then train SGG models in a fully-supervised manner with aligned instance-level "pseudo" labels. However, we argue that most existing WSSGG works only focus on object-consistency, which means the grounded regions should have the same object category label as text entities. While they neglect another basic requirement for an ideal alignment: interaction-consistency, which means the grounded region pairs should have the same interactions (i.e., visual relations) as text entity pairs. Hence, in this paper, we propose to enhance a simple grounding module with both object-aware and interaction-aware knowledge to acquire more reliable pseudo labels. To better leverage these two types of knowledge, we regard them as two teachers and fuse their generated targets to guide the training process of our grounding module. Specifically, we design two different strategies to adaptively assign weights to different teachers by assessing their reliability on each training sample. Extensive experiments have demonstrated that our method consistently improves WSSGG performance on various kinds of weak supervision.

Yangjun Mao, Long Chen, Zhihong Jiang et al.

Distinctive Image Captioning (DIC) -- generating distinctive captions that describe the unique details of a target image -- has received considerable attention over the last few years. A recent DIC work proposes to generate distinctive captions by comparing the target image with a set of semantic-similar reference images, i.e., reference-based DIC (Ref-DIC). It aims to make the generated captions can tell apart the target and reference images. Unfortunately, reference images used by existing Ref-DIC works are easy to distinguish: these reference images only resemble the target image at scene-level and have few common objects, such that a Ref-DIC model can trivially generate distinctive captions even without considering the reference images. To ensure Ref-DIC models really perceive the unique objects (or attributes) in target images, we first propose two new Ref-DIC benchmarks. Specifically, we design a two-stage matching mechanism, which strictly controls the similarity between the target and reference images at object-/attribute- level (vs. scene-level). Secondly, to generate distinctive captions, we develop a strong Transformer-based Ref-DIC baseline, dubbed as TransDIC. It not only extracts visual features from the target image, but also encodes the differences between objects in the target and reference images. Finally, for more trustworthy benchmarking, we propose a new evaluation metric named DisCIDEr for Ref-DIC, which evaluates both the accuracy and distinctiveness of the generated captions. Experimental results demonstrate that our TransDIC can generate distinctive captions. Besides, it outperforms several state-of-the-art models on the two new benchmarks over different metrics.

Miaobo Hu, Shuhao Hu, Bokun Wang et al.

Reinforcement learning improves LLM reasoning, but PPO/GRPO typically use fixed clipping and decoding temperature, which makes training brittle and tuning-heavy. We propose Adaptive Group Policy Optimization (AGPO), a critic-free refinement of GRPO that uses group-level statistics to control both update magnitude and exploration. AGPO uses a shared probe-derived statistical state to drive two controllers: (i) adaptive clipping, which sets the trust-region size from reward dispersion and skewness, probe vote entropy, policy entropy, and step-wise KL drift; and (ii) bidirectional adaptive temperature sampling, which heats or cools decoding around a base temperature according to centered uncertainty relative to a running baseline. On nine English and Chinese math/STEM benchmarks, Qwen2.5-14B trained with AGPO outperforms PPO/GRPO under the same generated-token budget, reaching 67.3% on GSM8K and 40.5% on MATH. Gains transfer to Llama-3-8B and Gemma-2-9B, and ablations confirm both modules are complementary. Our implementation is publicly available at https://github.com/wandugu/paper_agpo.

Wei Wang, Yuqian Yuan, Tianwei Lin et al.

Spatial intelligence requires multimodal large language models (MLLMs) to move beyond single-view perception and reason consistently about objects, visibility, geometry, and interactions across multiple viewpoints. However, progress in cross-view reasoning remains limited by three major gaps: the scarcity of large-scale well-annotated training data, the lack of comprehensive benchmarks for systematic evaluation, and the absence of explicit alignment mechanisms that establish object-level consistency across views. To address these gaps, we thoroughly develop CrossView Suite across three coordinated components: CrossViewSet, CrossViewBench, and CrossViewer. Firstly, we introduce a multi-agent data engine to meticulously curate a large-scale, high-quality cross-view instruction dataset, termed CrossViewSet, covering 17 fine-grained task types with 1.6M samples. Second, we meticulously create a scene-disjoint CrossViewBench to comprehensively assess the cross-view spatial understanding capability of an MLLM, evaluating it across various aspects. Finally, we propose CrossViewer, a progressive three-stage framework for cross-view spatial reasoning in MLLMs, following a Perception -> Alignment -> Reasoning paradigm. Our method equips an adaptive spatial region tokenizer to capture fine-grained object representations, and then aligns the multi-view objects explicitly, and thus fuses aligned features for boosting the cross-view inference capacity for MLLMs. Extensive experiments and analyses show that large-scale training data, systematic evaluation, and explicit cross-view alignment are all critical for advancing MLLMs from single-view perception toward real-world spatial intelligence. The project page is available at https://github.com/Thinkirin/Crossview-Suite.

Xingchen Li, Long Chen, Jian Shao et al.

Current Scene Graph Generation (SGG) methods tend to predict frequent predicate categories and fail to recognize rare ones due to the severe imbalanced distribution of predicates. To improve the robustness of SGG models on different predicate categories, recent research has focused on unbiased SGG and adopted mean Recall@K (mR@K) as the main evaluation metric. However, we discovered two overlooked issues about this de facto standard metric, which makes current unbiased SGG evaluation vulnerable and unfair: 1) mR@K neglects the correlations among predicates and unintentionally breaks category independence when ranking all the triplet predictions together regardless of the predicate categories. 2) mR@K neglects the compositional diversity of different predicates and assigns excessively high weights to some oversimple category samples with limited composable relation triplet types. In addition, we investigate the under-explored correlation between objects and predicates, which can serve as a simple but strong baseline for unbiased SGG. In this paper, we refine mR@K and propose two complementary evaluation metrics for unbiased SGG: Independent Mean Recall (MR) and weighted IMR (wIMR). These two metrics are designed by considering the category independence and diversity of composable relation triplets, respectively. We compare the proposed metrics with the de facto standard metrics through extensive experiments and discuss the solutions to evaluate unbiased SGG in a more trustworthy way.

Zhen Wang, Long Chen, Wenbo Ma et al.

Given an image and a reference caption, the image caption editing task aims to correct the misalignment errors and generate a refined caption. However, all existing caption editing works are implicit models, ie, they directly produce the refined captions without explicit connections to the reference captions. In this paper, we introduce a new task: Explicit Caption Editing (ECE). ECE models explicitly generate a sequence of edit operations, and this edit operation sequence can translate the reference caption into a refined one. Compared to the implicit editing, ECE has multiple advantages: 1) Explainable: it can trace the whole editing path. 2) Editing Efficient: it only needs to modify a few words. 3) Human-like: it resembles the way that humans perform caption editing, and tries to keep original sentence structures. To solve this new task, we propose the first ECE model: TIger. TIger is a non-autoregressive transformer-based model, consisting of three modules: Tagger_del, Tagger_add, and Inserter. Specifically, Tagger_del decides whether each word should be preserved or not, Tagger_add decides where to add new words, and Inserter predicts the specific word for adding. To further facilitate ECE research, we propose two new ECE benchmarks by re-organizing two existing datasets, dubbed COCO-EE and Flickr30K-EE, respectively. Extensive ablations on both two benchmarks have demonstrated the effectiveness of TIger.

Lin Li, Guikun Chen, Zhen Wang et al.

Compositional zero-shot learning aims to recognize unseen state-object compositions by leveraging known primitives (state and object) during training. However, effectively modeling interactions between primitives and generalizing knowledge to novel compositions remains a perennial challenge. There are two key factors: object-conditioned and state-conditioned variance, i.e., the appearance of states (or objects) can vary significantly when combined with different objects (or states). For instance, the state "old" can signify a vintage design for a "car" or an advanced age for a "cat". In this paper, we argue that these variances can be mitigated by predicting composition categories based on pre-observed primitive. To this end, we propose Progressive Language-based Observations (PLO), which can dynamically determine a better observation order of primitives. These observations comprise a series of concepts or languages that allow the model to understand image content in a step-by-step manner. Specifically, PLO adopts pre-trained vision-language models (VLMs) to empower the model with observation capabilities. We further devise two variants: 1) PLO-VLM: a two-step method, where a pre-observing classifier dynamically determines the observation order of two primitives. 2) PLO-LLM: a multi-step scheme, which utilizes large language models (LLMs) to craft composition-specific prompts for step-by-step observing. Extensive ablations on three challenging datasets demonstrate the superiority of PLO compared with state-of-the-art methods, affirming its abilities in compositional recognition.

Jun Xiao, Xinyang Jiang, Ningxin Zheng et al.

Deep learning-based models have achieved remarkable performance in video super-resolution (VSR) in recent years, but most of these models are less applicable to online video applications. These methods solely consider the distortion quality and ignore crucial requirements for online applications, e.g., low latency and low model complexity. In this paper, we focus on online video transmission, in which VSR algorithms are required to generate high-resolution video sequences frame by frame in real time. To address such challenges, we propose an extremely low-latency VSR algorithm based on a novel kernel knowledge transfer method, named convolutional kernel bypass graft (CKBG). First, we design a lightweight network structure that does not require future frames as inputs and saves extra time costs for caching these frames. Then, our proposed CKBG method enhances this lightweight base model by bypassing the original network with ``kernel grafts'', which are extra convolutional kernels containing the prior knowledge of external pretrained image SR models. In the testing phase, we further accelerate the grafted multi-branch network by converting it into a simple single-path structure. Experiment results show that our proposed method can process online video sequences up to 110 FPS, with very low model complexity and competitive SR performance.

Yangjun Mao, Jun Xiao, Dong Zhang et al.

Distinctive Image Captioning (DIC) -- generating distinctive captions that describe the unique details of a target image -- has received considerable attention over the last few years. A recent DIC method proposes to generate distinctive captions by comparing the target image with a set of semantic-similar reference images, i.e., reference-based DIC (Ref-DIC). It aims to force the generated captions to distinguish between the target image and the reference image. To ensure Ref-DIC models really perceive the unique objects (or attributes) in target images, we propose two new Ref-DIC benchmarks and develop a Transformer-based Ref-DIC baseline TransDIC. The model only extracts visual features from the target image, but also encodes the differences between objects in the target and reference images. Taking one step further, we propose a stronger TransDIC++, which consists of an extra contrastive learning module to make full use of the reference images. This new module is model-agnostic, which can be easily incorporated into various Ref-DIC architectures. Finally, for more trustworthy benchmarking, we propose a new evaluation metric named DisCIDEr for Ref-DIC, which evaluates both the accuracy and distinctiveness of the generated captions. Experimental results demonstrate that our TransDIC++ can generate distinctive captions. Besides, it outperforms several state-of-the-art models on the two new benchmarks over different metrics.

Xin Lu, Zhengda Lu, Yiqun Wang et al.

In this paper, we proposed a generative model that learns to synthesize the 4D facial expression with the neutral landmark. Existing works mainly focus on the generation of sequences guided by expression labels, speech, etc, while they are not robust to the change of different identities. Our LM-4DGAN utilizes neutral landmarks to guide the facial expression generation while adding an identity discriminator and a landmark autoencoder to the basic WGAN for achieving better identity robustness. Furthermore, we add a cross-attention mechanism to the existing displacement decoder which is suitable for the given identity.

Zhen Wang, Jun Xiao, Yueting Zhuang et al.

Controllable Image Captioning (CIC) -- generating natural language descriptions about images under the guidance of given control signals -- is one of the most promising directions towards next-generation captioning systems. Till now, various kinds of control signals for CIC have been proposed, ranging from content-related control to structure-related control. However, due to the format and target gaps of different control signals, all existing CIC works (or architectures) only focus on one certain control signal, and overlook the human-like combinatorial ability. By ``combinatorial", we mean that our humans can easily meet multiple needs (or constraints) simultaneously when generating descriptions. To this end, we propose a novel prompt-based framework for CIC by learning Combinatorial Prompts, dubbed as ComPro. Specifically, we directly utilize a pretrained language model GPT-2 as our language model, which can help to bridge the gap between different signal-specific CIC architectures. Then, we reformulate the CIC as a prompt-guide sentence generation problem, and propose a new lightweight prompt generation network to generate the combinatorial prompts for different kinds of control signals. For different control signals, we further design a new mask attention mechanism to realize the prompt-based CIC. Due to its simplicity, our ComPro can be further extended to more kinds of combined control signals by concatenating these prompts. Extensive experiments on two prevalent CIC benchmarks have verified the effectiveness and efficiency of our ComPro on both single and combined control signals.

Xingchen Li, Jun Xiao, Guikun Chen et al.

Today's scene graph generation (SGG) models typically require abundant manual annotations to learn new predicate types. Therefore, it is difficult to apply them to real-world applications with massive uncommon predicate categories whose annotations are hard to collect. In this paper, we focus on Few-Shot SGG (FSSGG), which encourages SGG models to be able to quickly transfer previous knowledge and recognize unseen predicates well with only a few examples. However, current methods for FSSGG are hindered by the high intra-class variance of predicate categories in SGG: On one hand, each predicate category commonly has multiple semantic meanings under different contexts. On the other hand, the visual appearance of relation triplets with the same predicate differs greatly under different subject-object compositions. Such great variance of inputs makes it hard to learn generalizable representation for each predicate category with current few-shot learning (FSL) methods. However, we found that this intra-class variance of predicates is highly related to the composed subjects and objects. To model the intra-class variance of predicates with subject-object context, we propose a novel Decomposed Prototype Learning (DPL) model for FSSGG. Specifically, we first construct a decomposable prototype space to capture diverse semantics and visual patterns of subjects and objects for predicates by decomposing them into multiple prototypes. Afterwards, we integrate these prototypes with different weights to generate query-adaptive predicate representation with more reliable semantics for each query sample. We conduct extensive experiments and compare with various baseline methods to show the effectiveness of our method.

Zhen Wang, Xinyun Jiang, Jun Xiao et al.

Explicit Caption Editing (ECE) -- refining reference image captions through a sequence of explicit edit operations (e.g., KEEP, DETELE) -- has raised significant attention due to its explainable and human-like nature. After training with carefully designed reference and ground-truth caption pairs, state-of-the-art ECE models exhibit limited generalization ability beyond the original training data distribution, i.e., they are tailored to refine content details only in in-domain samples but fail to correct errors in out-of-domain samples. To this end, we propose a new Diffusion-based Explicit Caption editing method: DECap. Specifically, we reformulate the ECE task as a denoising process under the diffusion mechanism, and introduce innovative edit-based noising and denoising processes. Thanks to this design, the noising process can help to eliminate the need for meticulous paired data selection by directly introducing word-level noises for training, learning diverse distribution over input reference caption. The denoising process involves the explicit predictions of edit operations and corresponding content words, refining reference captions through iterative step-wise editing. To further efficiently implement our diffusion process and improve the inference speed, DECap discards the prevalent multi-stage design and directly generates edit operations and content words simultaneously. Extensive ablations have demonstrated the strong generalization ability of DECap in various scenarios. More interestingly, it even shows great potential in improving the quality and controllability of caption generation.

Shaoning Xiao, Long Chen, Kaifeng Gao et al.

Reasoning about causal and temporal event relations in videos is a new destination of Video Question Answering (VideoQA).The major stumbling block to achieve this purpose is the semantic gap between language and video since they are at different levels of abstraction. Existing efforts mainly focus on designing sophisticated architectures while utilizing frame- or object-level visual representations. In this paper, we reconsider the multi-modal alignment problem in VideoQA from feature and sample perspectives to achieve better performance. From the view of feature,we break down the video into trajectories and first leverage trajectory feature in VideoQA to enhance the alignment between two modalities. Moreover, we adopt a heterogeneous graph architecture and design a hierarchical framework to align both trajectory-level and frame-level visual feature with language feature. In addition, we found that VideoQA models are largely dependent on language priors and always neglect visual-language interactions. Thus, two effective yet portable training augmentation strategies are designed to strengthen the cross-modal correspondence ability of our model from the view of sample. Extensive results show that our method outperforms all the state-of-the-art models on the challenging NExT-QA benchmark, which demonstrates the effectiveness of the proposed method.

Jingliang Li, Qiang Zhou, Chaohui Yu et al.

Since the advent of Neural Radiance Fields, novel view synthesis has received tremendous attention. The existing approach for the generalization of radiance field reconstruction primarily constructs an encoding volume from nearby source images as additional inputs. However, these approaches cannot efficiently encode the geometric information of real scenes with various scale objects/structures. In this work, we propose constructing multi-scale encoding volumes and providing multi-scale geometry information to NeRF models. To make the constructed volumes as close as possible to the surfaces of objects in the scene and the rendered depth more accurate, we propose to perform depth prediction and radiance field reconstruction simultaneously. The predicted depth map will be used to supervise the rendered depth, narrow the depth range, and guide points sampling. Finally, the geometric information contained in point volume features may be inaccurate due to occlusion, lighting, etc. To this end, we propose enhancing the point volume feature from depth-guided neighbor feature fusion. Experiments demonstrate the superior performance of our method in both novel view synthesis and dense geometry modeling without per-scene optimization.

Sijing Li, Zhongwei Qiu, Jiang Liu et al.

Accurate tumor analysis is central to clinical radiology and precision oncology, where early detection, reliable lesion characterization, and pathology-level risk assessment guide diagnosis and treatment planning. Chain-of-Thought (CoT) reasoning is particularly important in this setting because it enables step-by-step interpretation from imaging findings to clinical impressions and pathology conclusions, improving traceability and reducing diagnostic errors. Here, we target the clinical tumor analysis task and build a large-scale benchmark that operationalizes a multimodal reasoning pipeline, spanning findings, impressions, and pathology predictions. We curate TumorCoT, a large-scale dataset of 1.5M CoT-labeled VQA instructions paired with 3D CT scans, with step-aligned rationales and cross-modal alignments along the trajectory from findings to impression to pathology, enabling evaluation of both answer accuracy and reasoning consistency. We further propose TumorChain, a multimodal interleaved reasoning framework that tightly couples 3D imaging encoders, clinical text understanding, and organ-level vision-language alignment. Through cross-modal alignment and iterative interleaved causal reasoning, TumorChain grounds visual evidence, aggregates conclusions, and issues pathology predictions after multiple rounds of self-refinement, improving traceability and reducing hallucination risk. Experiments show consistent improvements over strong baselines in lesion detection, impression generation, and pathology classification, and demonstrate strong generalization on the DeepTumorVQA benchmark. These results highlight the potential of multimodal reasoning for reliable and interpretable tumor analysis in clinical practice. Detailed information about our project can be found on our project homepage at https://github.com/ZJU4HealthCare/TumorChain.

Chuanqing Zhuang, Xin Lu, Zehui Deng et al.

Omnidirectional 3D Gaussian Splatting with panoramas is a key technique for 3D scene representation, and existing methods typically rely on slow SfM to provide camera poses and sparse points priors. In this work, we propose a pose-free omnidirectional 3DGS method, named PFGS360, that reconstructs 3D Gaussians from unposed omnidirectional videos. To achieve accurate camera pose estimation, we first construct a spherical consistency-aware pose estimation module, which recovers poses by establishing consistent 2D-3D correspondences between the reconstructed Gaussians and the unposed images using Gaussians' internal depth priors. Besides, to enhance the fidelity of novel view synthesis, we introduce a depth-inlier-aware densification module to extract depth inliers and Gaussian outliers with consistent monocular depth priors, enabling efficient Gaussian densification and achieving photorealistic novel view synthesis. The experiments show significant outperformance over existing pose-free and pose-aware 3DGS methods on both real-world and synthetic 360-degree videos. Code is available at https://github.com/zcq15/PFGS360.

Qian Ye, Masanori Suganuma, Jun Xiao et al.

Reconstructing ghosting-free high dynamic range (HDR) images of dynamic scenes from a set of multi-exposure images is a challenging task, especially with large object motion and occlusions, leading to visible artifacts using existing methods. To address this problem, we propose a deep network that tries to learn multi-scale feature flow guided by the regularized loss. It first extracts multi-scale features and then aligns features from non-reference images. After alignment, we use residual channel attention blocks to merge the features from different images. Extensive qualitative and quantitative comparisons show that our approach achieves state-of-the-art performance and produces excellent results where color artifacts and geometric distortions are significantly reduced.

Haoyu Zheng, Yun Zhu, Shu Yuan et al.

Large language models often solve tasks from a fully specified prompt but degrade when the same requirements unfold over multiple turns, known as the lost-in-conversation (LiC) gap. We trace part of this degradation to self-contamination: intermediate assistant replies enter later context and carry early deviations forward. Motivated by this mechanism, we propose MAIGO, an on-policy self-distillation method that reduces this contamination using history-cleaned references from the model's own policy. For middle turns, MAIGO removes prior assistant replies while preserving the user-visible sharded prefix; for answer turns, it distills from paired full-view references conditioned on the completed user-side dialogue. A reliability weight downweights middle-turn samples that disagree with the clean reference. MAIGO requires no verifier rewards, state labels, or inference-time scaffolding. Under the LiC paired-view protocol with deterministic verifiers, MAIGO improves Qwen2.5-7B-Instruct SHARDED accuracy from 52.8 to 66.1 and the SHARDED/FULL ratio from 66.5% to 84.1%, while keeping FULL accuracy within 2.3 points. These results show that self-contamination is a trainable component of the LiC gap.

Jingliang Li, Zhengda Lu, Yiqun Wang et al.

In recent years, supervised or unsupervised learning-based MVS methods achieved excellent performance compared with traditional methods. However, these methods only use the probability volume computed by cost volume regularization to predict reference depths and this manner cannot mine enough information from the probability volume. Furthermore, the unsupervised methods usually try to use two-step or additional inputs for training which make the procedure more complicated. In this paper, we propose the DS-MVSNet, an end-to-end unsupervised MVS structure with the source depths synthesis. To mine the information in probability volume, we creatively synthesize the source depths by splattering the probability volume and depth hypotheses to source views. Meanwhile, we propose the adaptive Gaussian sampling and improved adaptive bins sampling approach that improve the depths hypotheses accuracy. On the other hand, we utilize the source depths to render the reference images and propose depth consistency loss and depth smoothness loss. These can provide additional guidance according to photometric and geometric consistency in different views without additional inputs. Finally, we conduct a series of experiments on the DTU dataset and Tanks & Temples dataset that demonstrate the efficiency and robustness of our DS-MVSNet compared with the state-of-the-art methods.

Lin Li, Guikun Chen, Hanrong Shi et al.

The rapid evolution of Multimodal Large Language Models (MLLMs) has brought substantial advancements in artificial intelligence, significantly enhancing the capability to understand and generate multimodal content. While prior studies have largely concentrated on model architectures and training methodologies, a thorough analysis of the benchmarks used for evaluating these models remains underexplored. This survey addresses this gap by systematically reviewing 211 benchmarks that assess MLLMs across four core domains: understanding, reasoning, generation, and application. We provide a detailed analysis of task designs, evaluation metrics, and dataset constructions, across diverse modalities. We hope that this survey will contribute to the ongoing advancement of MLLM research by offering a comprehensive overview of benchmarking practices and identifying promising directions for future work. An associated GitHub repository collecting the latest papers is available.

Yuqian Yuan, Wentong Li, Zhaocheng Li et al.

In this paper, we introduce InstructSAM, a unified and streamlined framework designed for multi-instance segmentation under arbitrary instructions. We formulates instruction-driven instance segmentation as a set-structured query prediction problem and propose an explicit reasoning-to-instance query interface that elegantly bridges a vision-language model (VLM) and SAM3. Specifically, a bank of learnable instance queries is injected into the VLM and contextualized with instruction and visual information, enabling each query to serve as an instance-aware slot. A hybrid-attention mechanism further promotes interaction among these queries, visual tokens, and instruction tokens, improving instance enumeration and reducing duplicate predictions. The resulting LLM-conditioned queries are projected into SAM3's detector query space to drive accurate multi-instance segmentation in a single forward pass. This design equips SAM3 with high-level instruction understanding, compositional reasoning, and instance-level set prediction without modifying its core architecture. To support training and evaluation, we further construct Inst2Seg, a high-quality and large-scale instruction-based instance segmentation dataset and benchmark that couples free-form instructions with instance-level masks. Extensive experiments show that only 2B-scale InstructSAM achieves strong results across complex instruction-driven and phrase-level referring segmentation benchmarks, outperforming prior end-to-end methods and SAM3's agentic pipeline while enabling efficient single-pass multi-instance prediction.

Zhengxi Lu, Fei Tang, Guangyi Liu et al.

MLLM-based GUI agents have demonstrated strong capabilities in complex user interface interaction tasks. However, long-horizon scenarios remain challenging, as these agents are burdened with tasks beyond their intrinsic capabilities, suffering from memory degradation, progress confusion, and math hallucination. To address these challenges, we present UI-Copilot, a collaborative framework where the GUI agent focuses on task execution while a lightweight copilot provides on-demand assistance for memory retrieval and numerical computation. We introduce memory decoupling to separate persistent observations from transient execution context, and train the policy agent to selectively invoke the copilot as Retriever or Calculator based on task demands. To enable effective tool invocation learning, we propose Tool-Integrated Policy Optimization (TIPO), which separately optimizes tool selection through single-turn prediction and task execution through on-policy multi-turn rollouts. Experimental results show that UI-Copilot-7B achieves state-of-the-art performance on challenging MemGUI-Bench, outperforming strong 7B-scale GUI agents such as GUI-Owl-7B and UI-TARS-1.5-7B. Moreover, UI-Copilot-7B delivers a 17.1% absolute improvement on AndroidWorld over the base Qwen model, highlighting UI-Copilot's strong generalization to real-world GUI tasks.

Dinging Li, Yingxiu Zhao, Xinrui Cheng et al.

Spatial reasoning over three-dimensional scenes is a core capability for embodied intelligence, yet continuous model improvement remains bottlenecked by the cost of geometric annotation. The self-evolving paradigm offers a promising path, but its reliance on model consensus to construct pseudo-labels causes training to reinforce rather than correct the model's own geometric errors. We identify a property unique to 3D spatial reasoning that circumvents this limitation: ground truth is a deterministic consequence of the underlying geometry, computable exactly from point clouds and camera poses without any model involvement. Building on this insight, we present SpatialEvo, a self-evolving framework for 3D spatial reasoning, centered on the Deterministic Geometric Environment (DGE). The DGE formalizes 16 spatial reasoning task categories under explicit geometric validation rules and converts unannotated 3D scenes into zero-noise interactive oracles, replacing model consensus with objective physical feedback. A single shared-parameter policy co-evolves across questioner and solver roles under DGE constraints: the questioner generates physically valid spatial questions grounded in scene observations, while the solver derives precise answers against DGE-verified ground truth. A task-adaptive scheduler endogenously concentrates training on the model's weakest categories, producing a dynamic curriculum without manual design. Experiments across nine benchmarks demonstrate that SpatialEvo achieves the highest average score at both 3B and 7B scales, with consistent gains on spatial reasoning benchmarks and no degradation on general visual understanding.

Zhengxi Lu, Zhiyuan Yao, Jinyang Wu et al.

Agent skills, structured packages of procedural knowledge and executable resources that agents dynamically load at inference time, have become a reliable mechanism for augmenting LLM agents. Yet inference-time skill augmentation is fundamentally limited: retrieval noise introduces irrelevant guidance, injected skill content imposes substantial token overhead, and the model never truly acquires the knowledge it merely follows. We ask whether skills can instead be internalized into model parameters, enabling zero-shot autonomous behavior without any runtime skill retrieval. We introduce SKILL0, an in-context reinforcement learning framework designed for skill internalization. SKILL0 introduces a training-time curriculum that begins with full skill context and progressively withdraws it. Skills are grouped offline by category and rendered with interaction history into a compact visual context, teaching he model tool invocation and multi-turn task completion. A Dynamic Curriculum then evaluates each skill file's on-policy helpfulness, retaining only those from which the current policy still benefits within a linearly decaying budget, until the agent operates in a fully zero-shot setting. Extensive agentic experiments demonstrate that SKILL0 achieves substantial improvements over the standard RL baseline (+9.7\% for ALFWorld and +6.6\% for Search-QA), while maintaining a highly efficient context of fewer than 0.5k tokens per step. Our code is available at https://github.com/ZJU-REAL/SkillZero.

Zixuan Wang, Yuchen Yan, Hongxing Li et al.

While long-horizon agentic tasks require language agents to perform dozens of sequential decisions, training such agents with reinforcement learning remains challenging. We identify two root causes: credit misattribution, where correct early actions are penalized due to terminal failures, and sample inefficiency, where scarce successful trajectories result in near-total loss of learning signal. We introduce a milestone-guided policy learning framework, BEACON, that leverages the compositional structure of long-horizon tasks to ensure precise credit assignment. BEACON partitions trajectories at milestone boundaries, applies temporal reward shaping within segments to credit partial progress, and estimates advantages at dual scales to prevent distant failures from corrupting the evaluation of local actions. On ALFWorld, WebShop, and ScienceWorld, BEACON consistently outperforms GRPO and GiGPO. Notably, on long-horizon ALFWorld tasks, BEACON achieves 92.9% success rate, nearly doubling GRPO's 53.5%, while improving effective sample utilization from 23.7% to 82.0%. These results establish milestone-anchored credit assignment as an effective paradigm for training long-horizon language agents. Code is available at https://github.com/ZJU-REAL/BEACON.

Tianwei Lin, Wenqiao Zhang, Sijing Li et al.

We present HealthGPT, a powerful Medical Large Vision-Language Model (Med-LVLM) that integrates medical visual comprehension and generation capabilities within a unified autoregressive paradigm. Our bootstrapping philosophy is to progressively adapt heterogeneous comprehension and generation knowledge to pre-trained large language models (LLMs). This is achieved through a novel heterogeneous low-rank adaptation (H-LoRA) technique, which is complemented by a tailored hierarchical visual perception approach and a three-stage learning strategy. To effectively learn the HealthGPT, we devise a comprehensive medical domain-specific comprehension and generation dataset called VL-Health. Experimental results demonstrate exceptional performance and scalability of HealthGPT in medical visual unified tasks. Our project can be accessed at https://github.com/DCDmllm/HealthGPT.

Qi Zhang, Jie Zhou, Qin Chen et al.

Structured sentiment analysis, which aims to extract the complex semantic structures such as holders, expressions, targets, and polarities, has obtained widespread attention from both industry and academia. Unfortunately, the existing structured sentiment analysis datasets refer to a few languages and are relatively small, limiting neural network models' performance. In this paper, we focus on the cross-lingual structured sentiment analysis task, which aims to transfer the knowledge from the source language to the target one. Notably, we propose a Knowledge-Enhanced Adversarial Model (\texttt{KEAM}) with both implicit distributed and explicit structural knowledge to enhance the cross-lingual transfer. First, we design an adversarial embedding adapter for learning an informative and robust representation by capturing implicit semantic information from diverse multi-lingual embeddings adaptively. Then, we propose a syntax GCN encoder to transfer the explicit semantic information (e.g., universal dependency tree) among multiple languages. We conduct experiments on five datasets and compare \texttt{KEAM} with both the supervised and unsupervised methods. The extensive experimental results show that our \texttt{KEAM} model outperforms all the unsupervised baselines in various metrics.

Ruizhong Qiu, Hanqing Zeng, Yinglong Xia et al.

Low-rank adapters (LoRAs) are a parameter-efficient finetuning technique that injects trainable low-rank matrices into pretrained models to adapt them to new tasks. Mixture-of-LoRAs models expand neural networks efficiently by routing each layer input to a small subset of specialized LoRAs of the layer. Existing Mixture-of-LoRAs routers assign a learned routing weight to each LoRA to enable end-to-end training of the router. Despite their empirical promise, we observe that the routing weights are typically extremely imbalanced across LoRAs in practice, where only one or two LoRAs often dominate the routing weights. This essentially limits the number of effective LoRAs and thus severely hinders the expressive power of existing Mixture-of-LoRAs models. In this work, we attribute this weakness to the nature of learnable routing weights and rethink the fundamental design of the router. To address this critical issue, we propose a new router designed that we call Reinforcement Routing for Mixture-of-LoRAs (ReMix). Our key idea is using non-learnable routing weights to ensure all active LoRAs to be equally effective, with no LoRA dominating the routing weights. However, our routers cannot be trained directly via gradient descent due to our non-learnable routing weights. Hence, we further propose an unbiased gradient estimator for the router by employing the reinforce leave-one-out (RLOO) technique, where we regard the supervision loss as the reward and the router as the policy in reinforcement learning. Our gradient estimator also enables to scale up training compute to boost the predictive performance of our ReMix. Extensive experiments demonstrate that our proposed ReMix significantly outperform state-of-the-art parameter-efficient finetuning methods under a comparable number of activated parameters.

Wendong Bu, Kaihang Pan, Yuze Lin et al.

Large language models (LLMs) have unified diverse linguistic tasks within a single framework, yet such unification remains unexplored in human motion generation. Existing methods are confined to isolated tasks, limiting flexibility for free-form and omni-objective generation. To address this, we propose OmniMoGen, a unified framework that enables versatile motion generation through interleaved text-motion instructions. Built upon a concise RVQ-VAE and transformer architecture, OmniMoGen supports end-to-end instruction-driven motion generation. We construct X2Mo, a large-scale dataset of over 137K interleaved text-motion instructions, and introduce AnyContext, a benchmark for evaluating interleaved motion generation. Experiments show that OmniMoGen achieves state-of-the-art performance on text-to-motion, motion editing, and AnyContext, exhibiting emerging capabilities such as compositional editing, self-reflective generation, and knowledge-informed generation. These results mark a step toward the next intelligent motion generation. Project Page: https://OmniMoGen.github.io/.

Keyu Wang, Bingchen Miao, Wendong Bu et al.

The development of Multimodal Virtual Agents has made significant progress through the integration of Multimodal Large Language Models. However, mainstream training paradigms face key challenges: Behavior Cloning is simple and effective through imitation but suffers from low behavioral diversity, while Reinforcement Learning is capable of discovering novel strategies through exploration but heavily relies on manually designed reward functions. To address the conflict between these two methods, we present CORE, a Code-based Inverse Self-Training Framework with Graph Expansion that bridges imitation and exploration, offering a novel training framework that promotes behavioral diversity while eliminating the reliance on manually reward design. Specifically, we introduce Semantic Code Abstraction to automatically infers reward functions from expert demonstrations without manual design. The inferred reward function, referred to as the Label Function, is executable code that verifies one key step within a task. Building on this, we propose Strategy Graph Expansion to enhance in-domain behavioral diversity, which constructs a multi-path graph called Strategy Graph that captures diverse valid solutions beyond expert demonstrations. Furthermore, we introduce Trajectory-Guided Extrapolation, which enriches out-of-domain behavioral diversity by utilizing both successful and failed trajectories to expand the task space. Experiments on Web and Android platforms demonstrate that CORE significantly improves both overall performance and generalization, highlighting its potential as a robust and generalizable training paradigm for building powerful virtual agents.

Fei Tang, Bofan Chen, Zhengxi Lu et al.

GUI grounding, which localizes interface elements from screenshots given natural language queries, remains challenging for small icons and dense layouts. Test-time zoom-in methods improve localization by cropping and re-running inference at higher resolution, but apply cropping uniformly across all instances with fixed crop sizes, ignoring whether the model is actually uncertain on each case. We propose \textbf{UI-Zoomer}, a training-free adaptive zoom-in framework that treats both the trigger and scale of zoom-in as a prediction uncertainty quantification problem. A confidence-aware gate fuses spatial consensus among stochastic candidates with token-level generation confidence to selectively trigger zoom-in only when localization is uncertain. When triggered, an uncertainty-driven crop sizing module decomposes prediction variance into inter-sample positional spread and intra-sample box extent, deriving a per-instance crop radius via the law of total variance. Extensive experiments on ScreenSpot-Pro, UI-Vision, and ScreenSpot-v2 demonstrate consistent improvements over strong baselines across multiple model architectures, achieving gains of up to +13.4\%, +10.3\%, and +4.2\% respectively, with no additional training required.

Lin Li, Ziqi Jiang, Gefan Ye et al.

Recent advances in cross-modal few-shot adaptation treat visual-semantic alignment as a continuous feature transport problem via Flow Matching (FM). However, we argue that Euclidean-based FM overlooks fundamental limitations of flat geometry, where polynomial volume growth fails to accommodate diverse feature distributions, leading to severe path entanglement. To this end, we propose path-decoupled Hyperbolic Flow Matching (HFM), leveraging the Lorentz manifold's exponential expansion for trajectory decoupling. HFM structures the transport via two key designs: 1) Centripetal hyperbolic alignment: It constructs a centripetal hierarchy by anchoring textual roots, which pushes visual leaves to the boundary to initialize orderly flows. 2) Path-decoupled objective: It acts as a ``semantic guardrail'' rigidly confining trajectories within isolated class-specific geodesic corridors via step-wise supervision. Furthermore, we devise an adaptive diameter-based stopping to prevent over-transportation into the crowded origin based on the intrinsic semantic scale. Extensive ablations on 11 benchmarks have shown that HFM establishes a new state-of-the-art, consistently outperforming its Euclidean counterparts. Our codes and models will be released.