Yuqian Fu

Xingyu Qiu, Yuqian Fu, Jiawei Geng et al.

Cross-domain few-shot object detection (CD-FSOD) remains a challenging problem for existing object detectors and few-shot learning approaches, particularly when generalizing across distinct domains. As part of NTIRE 2026, we hosted the second CD-FSOD Challenge to systematically evaluate and promote progress in detecting objects in unseen target domains under limited annotation conditions. The challenge received strong community interest, with 128 registered participants and a total of 696 submissions. Among them, 31 teams actively participated, and 19 teams submitted valid final results. Participants explored a wide range of strategies, introducing innovative methods that push the performance frontier under both open-source and closed-source tracks. This report presents a detailed overview of the NTIRE 2026 CD-FSOD Challenge, including a summary of the submitted approaches and an analysis of the final results across all participating teams. Challenge Codes: https://github.com/ohMargin/NTIRE2026_CDFSOD.

Yuqian Fu, Yu Xie, Yanwei Fu et al.

Cross-Domain Few-Shot Learning (CD-FSL) is a recently emerging task that tackles few-shot learning across different domains. It aims at transferring prior knowledge learned on the source dataset to novel target datasets. The CD-FSL task is especially challenged by the huge domain gap between different datasets. Critically, such a domain gap actually comes from the changes of visual styles, and wave-SAN empirically shows that spanning the style distribution of the source data helps alleviate this issue. However, wave-SAN simply swaps styles of two images. Such a vanilla operation makes the generated styles ``real'' and ``easy'', which still fall into the original set of the source styles. Thus, inspired by vanilla adversarial learning, a novel model-agnostic meta Style Adversarial training (StyleAdv) method together with a novel style adversarial attack method is proposed for CD-FSL. Particularly, our style attack method synthesizes both ``virtual'' and ``hard'' adversarial styles for model training. This is achieved by perturbing the original style with the signed style gradients. By continually attacking styles and forcing the model to recognize these challenging adversarial styles, our model is gradually robust to the visual styles, thus boosting the generalization ability for novel target datasets. Besides the typical CNN-based backbone, we also employ our StyleAdv method on large-scale pretrained vision transformer. Extensive experiments conducted on eight various target datasets show the effectiveness of our method. Whether built upon ResNet or ViT, we achieve the new state of the art for CD-FSL. Code is available at https://github.com/lovelyqian/StyleAdv-CDFSL.

Kailing Li, Tianwen Qian, Lijin Yang et al.

Vision-Language Navigation (VLN) enables embodied agents to reach target locations in unseen environments by following language instructions. Despite recent progress with vision-language models (VLMs), a critical semantic-geometric gap remains: while VLMs excel at language and 2D visual understanding, they struggle with 3D spatial reasoning and fail to capture the causal dynamics between actions and spatial transitions, resulting in unreliable navigation, particularly in zero-shot settings. To bridge this gap, we propose a Hierarchical Semantic-Geometric Map (HSGM) that transforms 3D geometric information into a structured representation compatible with VLMs, effectively linking them to the physical world. Specifically, HSGM is represented as a multi-channel top-down map organized into three levels: (1) geometric level that records navigable regions and obstacles, (2) semantic level that represents objects and their relations, and (3) decision level that supports high-level task reasoning and goal selection. During navigation, the VLM acts as a high-level semantic planner, interpreting the spatial layout encoded in the HSGM to select geometrically valid waypoints, while low-level, collision-free movements between waypoints are executed by a classical path-planning algorithm, fully decoupling semantic reasoning from action execution. Additionally, complex instructions are decomposed into subtasks to alleviate the problem of progress forgetting or hallucinating in long-horizon navigation. Extensive experiments on R2R-CE and RxR-CE benchmarks demonstrate that our zero-shot framework achieves state-of-the-art performance and even outperforms several supervised methods. Code is available at https://github.com/Teacher-Tom/HSGM_public.

Yousef A. Radwan, Xuhui Liu, Kilichbek Haydarov et al.

Large language models (LLMs) have emerged as powerful representation learners whose internal features increasingly align with human cognition. We study whether modern LLMs can serve as a lens for understanding neural representations in the human brain, focusing on emotional valence in EEG. We first build a one-dimensional valence direction, the V-axis, from modern LLMs using only nine emotion-evocative sentences. We validate it through zero-shot transfer to sentiment benchmarks and cross-model consistency across fourteen LLMs. We then show that this LLM-derived direction maps onto human neural activity. On a public EEG cohort of 123 subjects watching affective videos, a single linear projection on EEG features tracks the V-axis position of each stimulus. Moreover, 36 EEG emotion classifiers trained without exposure to the V-axis spontaneously rediscover the same direction in their internal representations, suggesting that the same valence structure emerges in both language models and human electrophysiology. Yet this convergence does not provide an effective training signal. We test twenty-five alignment strategies, including knowledge distillation, representational similarity, contrastive, and topographic losses; none improve decoding, and sixteen significantly reduce accuracy. We formalize this result as the saturation regularity: once task labels alone drive a brain-decoding network onto the target direction, additional supervision mainly distorts an already-saturated basin, while the load-bearing within-class residual receives little useful gradient. This regularity also indicates where improvement should come from: the residual subspace unreachable by supervision. Motivated by this insight, we ensemble across residual diversity rather than supervising the basin, improving balanced accuracy by 10.5% over the prior best on FACED, with the same effect replicated on SEED-V.

Yuqian Fu, Yu Xie, Yanwei Fu et al.

Recently, Cross-Domain Few-Shot Learning (CD-FSL) which aims at addressing the Few-Shot Learning (FSL) problem across different domains has attracted rising attention. The core challenge of CD-FSL lies in the domain gap between the source and novel target datasets. Though many attempts have been made for CD-FSL without any target data during model training, the huge domain gap makes it still hard for existing CD-FSL methods to achieve very satisfactory results. Alternatively, learning CD-FSL models with few labeled target domain data which is more realistic and promising is advocated in previous work~\cite{fu2021meta}. Thus, in this paper, we stick to this setting and technically contribute a novel Multi-Expert Domain Decompositional Network (ME-D2N). Concretely, to solve the data imbalance problem between the source data with sufficient examples and the auxiliary target data with limited examples, we build our model under the umbrella of multi-expert learning. Two teacher models which can be considered to be experts in their corresponding domain are first trained on the source and the auxiliary target sets, respectively. Then, the knowledge distillation technique is introduced to transfer the knowledge from two teachers to a unified student model. Taking a step further, to help our student model learn knowledge from different domain teachers simultaneously, we further present a novel domain decomposition module that learns to decompose the student model into two domain-related sub parts. This is achieved by a novel domain-specific gate that learns to assign each filter to only one specific domain in a learnable way. Extensive experiments demonstrate the effectiveness of our method. Codes and models are available at https://github.com/lovelyqian/ME-D2N_for_CDFSL.

Yebin Yang, Di Wen, Lei Qi et al.

Text-guided 3D motion editing has seen success in single-person scenarios, but its extension to multi-person settings is less explored due to limited paired data and the complexity of inter-person interactions. We introduce the task of multi-person 3D motion editing, where a target motion is generated from a source and a text instruction. To support this, we propose InterEdit3D, a new dataset with manual two-person motion change annotations, and a Text-guided Multi-human Motion Editing (TMME) benchmark. We present InterEdit, a synchronized classifier-free conditional diffusion model for TMME. It introduces Semantic-Aware Plan Token Alignment with learnable tokens to capture high-level interaction cues and an Interaction-Aware Frequency Token Alignment strategy using DCT and energy pooling to model periodic motion dynamics. Experiments show that InterEdit improves text-to-motion consistency and edit fidelity, achieving state-of-the-art TMME performance. The dataset and code will be released at https://github.com/YNG916/InterEdit.

Yuqian Fu, Yu Xie, Yanwei Fu et al.

Previous few-shot learning (FSL) works mostly are limited to natural images of general concepts and categories. These works assume very high visual similarity between the source and target classes. In contrast, the recently proposed cross-domain few-shot learning (CD-FSL) aims at transferring knowledge from general nature images of many labeled examples to novel domain-specific target categories of only a few labeled examples. The key challenge of CD-FSL lies in the huge data shift between source and target domains, which is typically in the form of totally different visual styles. This makes it very nontrivial to directly extend the classical FSL methods to address the CD-FSL task. To this end, this paper studies the problem of CD-FSL by spanning the style distributions of the source dataset. Particularly, wavelet transform is introduced to enable the decomposition of visual representations into low-frequency components such as shape and style and high-frequency components e.g., texture. To make our model robust to visual styles, the source images are augmented by swapping the styles of their low-frequency components with each other. We propose a novel Style Augmentation (StyleAug) module to implement this idea. Furthermore, we present a Self-Supervised Learning (SSL) module to ensure the predictions of style-augmented images are semantically similar to the unchanged ones. This avoids the potential semantic drift problem in exchanging the styles. Extensive experiments on two CD-FSL benchmarks show the effectiveness of our method. Our codes and models will be released.

Yilun Zhang, Yuqian Fu, Xingjun Ma et al.

Deep learning has achieved great success in video recognition, yet still struggles to recognize novel actions when faced with only a few examples. To tackle this challenge, few-shot action recognition methods have been proposed to transfer knowledge from a source dataset to a novel target dataset with only one or a few labeled videos. However, existing methods mainly focus on modeling the temporal relations between the query and support videos while ignoring the spatial relations. In this paper, we find that the spatial misalignment between objects also occurs in videos, notably more common than the temporal inconsistency. We are thus motivated to investigate the importance of spatial relations and propose a more accurate few-shot action recognition method that leverages both spatial and temporal information. Particularly, a novel Spatial Alignment Cross Transformer (SA-CT) which learns to re-adjust the spatial relations and incorporates the temporal information is contributed. Experiments reveal that, even without using any temporal information, the performance of SA-CT is comparable to temporal based methods on 3/4 benchmarks. To further incorporate the temporal information, we propose a simple yet effective Temporal Mixer module. The Temporal Mixer enhances the video representation and improves the performance of the full SA-CT model, achieving very competitive results. In this work, we also exploit large-scale pretrained models for few-shot action recognition, providing useful insights for this research direction.

Jianxiong Gao, Yanwei Fu, Yuqian Fu et al.

Reconstructing 3D visuals from functional Magnetic Resonance Imaging (fMRI) data, introduced as Recon3DMind, is of significant interest to both cognitive neuroscience and computer vision. To advance this task, we present the fMRI-3D dataset, which includes data from 15 participants and showcases a total of 4,768 3D objects. The dataset consists of two components: fMRI-Shape, previously introduced and available at https://huggingface.co/datasets/Fudan-fMRI/fMRI-Shape, and fMRI-Objaverse, proposed in this paper and available at https://huggingface.co/datasets/Fudan-fMRI/fMRI-Objaverse. fMRI-Objaverse includes data from 5 subjects, 4 of whom are also part of the core set in fMRI-Shape. Each subject views 3,142 3D objects across 117 categories, all accompanied by text captions. This significantly enhances the diversity and potential applications of the dataset. Moreover, we propose MinD-3D++, a novel framework for decoding textured 3D visual information from fMRI signals. The framework evaluates the feasibility of not only reconstructing 3D objects from the human mind but also generating, for the first time, 3D textured meshes with detailed textures from fMRI data. We establish new benchmarks by designing metrics at the semantic, structural, and textured levels to evaluate model performance. Furthermore, we assess the model's effectiveness in out-of-distribution settings and analyze the attribution of the proposed 3D pari fMRI dataset in visual regions of interest (ROIs) in fMRI signals. Our experiments demonstrate that MinD-3D++ not only reconstructs 3D objects with high semantic and spatial accuracy but also provides deeper insights into how the human brain processes 3D visual information. Project page: https://jianxgao.github.io/MinD-3D.

Yuqian Fu, Haohuan Huang, Kaiwen Jiang et al.

On-policy distillation (OPD) is appealing for large language model (LLM) post-training because it evaluates teacher feedback on student-generated rollouts rather than fixed teacher traces. In long-horizon settings, however, the common sampled-token variant is fragile: it reduces distribution matching to a one-token signal and becomes increasingly unreliable as rollouts drift away from prefixes the teacher commonly visits. We revisit OPD from the estimator and implementation sides. Theoretically, token-level OPD is biased relative to sequence-level reverse-KL, but it has a much tighter worst-case variance bound; our toy study shows the same tradeoff empirically, with stronger future-reward coupling producing higher gradient variance and less stable learning. Empirically, we identify three failure modes of sampled-token OPD: an imbalanced one-token signal, unreliable teacher guidance on student-generated prefixes, and distortions caused by tokenizer or special-token mismatch. We address these issues with teacher top-K local support matching, implemented as truncated reverse-KL with top-p rollout sampling and special-token masking. Across single-task math reasoning and multi-task agentic-plus-math training, this objective yields more stable optimization and better downstream performance than sampled-token OPD.

Jiancheng Pan, Yanxing Liu, Yuqian Fu et al.

Object detection, particularly open-vocabulary object detection, plays a crucial role in Earth sciences, such as environmental monitoring, natural disaster assessment, and land-use planning. However, existing open-vocabulary detectors, primarily trained on natural-world images, struggle to generalize to remote sensing images due to a significant data domain gap. Thus, this paper aims to advance the development of open-vocabulary object detection in remote sensing community. To achieve this, we first reformulate the task as Locate Anything on Earth (LAE) with the goal of detecting any novel concepts on Earth. We then developed the LAE-Label Engine which collects, auto-annotates, and unifies up to 10 remote sensing datasets creating the LAE-1M - the first large-scale remote sensing object detection dataset with broad category coverage. Using the LAE-1M, we further propose and train the novel LAE-DINO Model, the first open-vocabulary foundation object detector for the LAE task, featuring Dynamic Vocabulary Construction (DVC) and Visual-Guided Text Prompt Learning (VisGT) modules. DVC dynamically constructs vocabulary for each training batch, while VisGT maps visual features to semantic space, enhancing text features. We comprehensively conduct experiments on established remote sensing benchmark DIOR, DOTAv2.0, as well as our newly introduced 80-class LAE-80C benchmark. Results demonstrate the advantages of the LAE-1M dataset and the effectiveness of the LAE-DINO method.

Yaocheng Zhang, Jiajun Chai, Songjun Tu et al.

On-policy distillation (OPD) provides dense teacher feedback along rollouts generated by the student and has emerged as a promising post-training paradigm for long-horizon reasoning. However, standard OPD typically generates full rollouts during training, which is computationally expensive and may expose the student to unreliable teacher feedback at late rollout positions, especially during early training. We identify the rollout horizon as a key bottleneck in OPD that substantially impacts training efficiency. Unlike Reinforcement Learning with Verifiable Rewards (RLVR), OPD does not require a complete trajectory or a final answer reward to provide learning signals. This observation suggests that full rollouts may not always be necessary for effective OPD. Motivated by this insight, we propose two simple horizon-control strategies: Progressive OPD (POPD), which gradually expands the rollout horizon during training, and Truncated OPD (TOPD), which permanently performs distillation on reliable truncated rollouts. Experiments on mathematical reasoning show that POPD improves the training efficiency of OPD by up to 3$\times$, while TOPD matches OPD performance using only 10\% of the rollout horizon, leading to substantial wall-clock and memory reductions. These results demonstrate that controlling the rollout horizon offers a simple and practical path to more efficient OPD.

Linhai Zhuo, Yuqian Fu, Jingjing Chen et al.

Given sufficient training data on the source domain, cross-domain few-shot learning (CD-FSL) aims at recognizing new classes with a small number of labeled examples on the target domain. The key to addressing CD-FSL is to narrow the domain gap and transferring knowledge of a network trained on the source domain to the target domain. To help knowledge transfer, this paper introduces an intermediate domain generated by mixing images in the source and the target domain. Specifically, to generate the optimal intermediate domain for different target data, we propose a novel target guided dynamic mixup (TGDM) framework that leverages the target data to guide the generation of mixed images via dynamic mixup. The proposed TGDM framework contains a Mixup-3T network for learning classifiers and a dynamic ratio generation network (DRGN) for learning the optimal mix ratio. To better transfer the knowledge, the proposed Mixup-3T network contains three branches with shared parameters for classifying classes in the source domain, target domain, and intermediate domain. To generate the optimal intermediate domain, the DRGN learns to generate an optimal mix ratio according to the performance on auxiliary target data. Then, the whole TGDM framework is trained via bi-level meta-learning so that TGDM can rectify itself to achieve optimal performance on target data. Extensive experimental results on several benchmark datasets verify the effectiveness of our method.

Linan ZHU, Zihao Zhai, Xiao Han et al.

Emotion Recognition in Conversation (ERC) is essential for effective human-machine interaction, aiming to identify speakers' emotional states in multi-turn dialogues. Early text-based methods struggle with complex scenarios like sarcasm because they inherently neglect vital non-verbal information. While recent Vision-Language Models (VLMs) address this by analyzing video directly, they are not inherently tailored for ERC and often focus on emotionally irrelevant background regions or passive listeners rather than the active speaker. Furthermore, fine-tuning these large models incurs prohibitive computational costs. Additionally, isolated visual signals are frequently ambiguous or technically compromised without the context of linguistic content and vocal prosody. To address these challenges, we propose VISAFF, a speaker-centered VISual AFFective feature learning framework for ERC. VISAFF consists of two stages: Speaker-Centered Affective Grounding and Reliability-Guided Affective Complementation. VISAFF utilizes a tuning-free approach to unlock the reasoning capabilities of frozen VLMs, efficiently steering them to focus on the active speaker's emotional visual cues without heavy training overheads. In the second stage, we introduce a reliability-guided affective complementation mechanism that dynamically leverages textual and acoustic modalities to compensate for visual uncertainty. Experiments on two real-world datasets demonstrate that VISAFF achieves highly competitive performance compared to state-of-the-art methods in a tuning-free setting, significantly enhancing computational efficiency by eliminating the need for expensive fine-tuning of large VLMs. The source code is available at https://anonymous.4open.science/r/speaker-2365/.

Siteng Huang, Biao Gong, Yutong Feng et al.

This study focuses on a novel task in text-to-image (T2I) generation, namely action customization. The objective of this task is to learn the co-existing action from limited data and generalize it to unseen humans or even animals. Experimental results show that existing subject-driven customization methods fail to learn the representative characteristics of actions and struggle in decoupling actions from context features, including appearance. To overcome the preference for low-level features and the entanglement of high-level features, we propose an inversion-based method Action-Disentangled Identifier (ADI) to learn action-specific identifiers from the exemplar images. ADI first expands the semantic conditioning space by introducing layer-wise identifier tokens, thereby increasing the representational richness while distributing the inversion across different features. Then, to block the inversion of action-agnostic features, ADI extracts the gradient invariance from the constructed sample triples and masks the updates of irrelevant channels. To comprehensively evaluate the task, we present an ActionBench that includes a variety of actions, each accompanied by meticulously selected samples. Both quantitative and qualitative results show that our ADI outperforms existing baselines in action-customized T2I generation. Our project page is at https://adi-t2i.github.io/ADI.

Nedyalko Prisadnikov, Danda Pani Paudel, Yuqian Fu et al.

Vision Transformers (ViTs) dominate self-supervised learning (SSL). While they have proven highly effective for large-scale pretraining, they are computationally inefficient and scale poorly with image size. Consequently, foundational models like DINO are constrained to low-resolution processing. A recent foveal-inspired transformer achieves resolution agnosticism by iteratively processing a fixed-size context of multi-zoom patches. This model demonstrated promising results via supervised learning, utilizing a sequential, recurrent-like process without backpropagation through time. To unlock its potential as a foundational backbone, we introduce a novel sequential-to-global SSL framework based on DINO's self-distillation objective. Supported by an efficient integral-image patch extraction method, our approach enables large-scale pretraining for image-size agnostic vision encoders. We achieve competitive performance on ImageNet-1K and downstream classification tasks, maintaining a constant computational budget regardless of input resolution.

Kuanning Wang, Ke Fan, Yuqian Fu et al.

We present OCRA, an Object-Centric framework for video-based human-to-Robot Action transfer that learns directly from human demonstration videos to enable robust manipulation. Object-centric learning emphasizes task-relevant objects and their interactions while filtering out irrelevant background, providing a natural and scalable way to teach robots. OCRA leverages multi-view RGB videos, the state-of-the-art 3D foundation model VGGT, and advanced detection and segmentation models to reconstruct object-centric 3D point clouds, capturing rich interactions between objects. To handle properties not easily perceived by vision alone, we incorporate tactile priors via a large-scale dataset of over one million tactile images. These 3D and tactile priors are fused through a multimodal module (ResFiLM) and fed into a Diffusion Policy to generate robust manipulation actions. Extensive experiments on both vision-only and visuo-tactile tasks show that OCRA significantly outperforms existing baselines and ablations, demonstrating its effectiveness for learning from human demonstration videos.

Qi'ao Xu, Tianwen Qian, Yuqian Fu et al.

A core capability towards general embodied intelligence lies in localizing task-relevant objects from an egocentric perspective, formulated as Spatio-Temporal Video Grounding (STVG). Despite recent progress, existing STVG studies remain largely confined to object-centric and descriptive instructions, neglecting the task-oriented reasoning that is crucial for embodied agents to accomplish goal-directed interactions. To bridge this gap, we introduce \textbf{ToG-Bench}, the first task-oriented spatio-temporal video grounding benchmark for egocentric videos. ToG-Bench is characterized by three key features: (1) \textbf{Task-oriented Grounding}, which requires identifying and localizing objects based on intended tasks rather than straightforward descriptions; (2) \textbf{Explicit-Implicit Dual Grounding}, where target objects can be either explicitly mentioned or implicitly inferred by contextual reasoning; (3) \textbf{One-to-Many Grounding}, where a single instruction may correspond to multiple objects involved in task execution. Built upon videos sourced from ScanNet, ToG-Bench comprises 100 annotated clips with 2,704 task-oriented grounding instructions, constructed via a semi-automated pipeline that combines foundation model annotation and human refinement. In addition, we introduce a set of task-level evaluation metrics tailored for multi-object and explicit-implicit object grounding, and systematically benchmark seven state-of-the-art MLLMs. Extensive experiments reveal the intrinsic challenges of task-oriented STVG and substantial performance gaps across explicit-implicit and multi-object grounding, highlighting the difficulty of bridging perception and interaction in embodied scenarios. Data and code will be released at: \href{https://github.com/qaxuDev/ToG-Bench}{https://github.com/qaxuDev/ToG-Bench}..

Runze Wang, Yuqian Fu, Yu Li et al.

Vision-language-action (VLA) models have shown strong potential for generalist robot manipulation, yet they remain limited by insufficient spatial reasoning, particularly in determining where to interact in complex visual scenes. While recent efforts introduce various forms of visual planning to address this issue, existing approaches either rely on global geometric cues, symbolic intermediate representations, or externally generated visual signals, which are often weakly coupled with downstream action prediction. In this work, we revisit visual planning in VLA systems and argue that effective planning should be local, visually grounded, internally generated, and directly aligned with action. Based on this insight, we propose Afford-VLA, a unified framework that internalizes task-conditioned affordance as an explicit visual planning interface within VLA models. Concretely, we introduce learnable <AFF> tokens to query task-relevant interaction regions, decode affordance masks from multimodal features, and convert them into compact embeddings that directly condition action generation. This design enables affordance to be both generated and utilized within the VLA, forming a tightly coupled perception-action pathway. To further support this integration, we adopt a training strategy that allows the affordance pathway to be jointly optimized with action prediction, improving its effectiveness for downstream control. We evaluate our method on multiple simulation benchmarks, including LIBERO, LIBERO-Plus, and SimplerEnv, achieving consistent state-of-the-art performance, along with strong real-world results. These findings demonstrate that internalizing affordance as action-aligned visual planning provides a powerful paradigm for improving VLA systems.

Mingkang Dong, Hongyi Cai, Jie Li et al.

The effectiveness of multimodal instruction tuning depends not only on dataset scale, but critically on whether training samples genuinely require visual reasoning. However, existing instruction datasets often contain a substantial portion of visually redundant samples (solvable from text alone), as well as multimodally misaligned supervision that can degrade learning. To address this, we propose VisNec (Visual Necessity Score), a principled data selection framework that measures the marginal contribution of visual input during instruction tuning. By comparing predictive loss with and without visual context, VisNec identifies whether a training instance is vision-critical, redundant, or misaligned. To preserve task diversity, we combine VisNec with semantic clustering and select high-necessity samples within each cluster. Across 10 downstream benchmarks, training on only 15% of the LLaVA-665K dataset selected by VisNec achieves 100.2% of full-data performance. On the smaller Vision-Flan-186K dataset, our selection not only further reduces data size but also surpasses full-data training by 15.8%. These results demonstrate that measuring and leveraging visual necessity provides an effective solution for both efficient and robust multimodal instruction tuning. Codes and selected subsets will be released upon acceptance.

Carmelo Scribano, Mohammad Mahdi, Nedyalko Prisadnikov et al.

Pretrained vision foundation models deliver strong performance across tasks with limited fine-tuning. However, their Vision Transformer (ViT) backbones impose high inference costs, limiting deployment on resource-constrained devices. In this work, we accelerate large-scale pretrained ViTs while preserving their feature extraction capabilities by exploiting the intrinsic convolution-like behavior of some attention heads. Specifically, we introduce an efficient depthwise convolution-based layer that serves as a drop-in replacement for these heads. Additionally, we propose simple strategies to identify which heads can be replaced and introduce a fine-tuning procedure that recovers downstream task performance. Across both image classification and segmentation tasks, our method achieves 17-20\% percent inference speedup with minimal performance degradation. We validate the approach through detailed derivations, extensive experiments, and efficiency benchmarks. The reference implementation is publicly available.

Kuanning Wang, Ke Fan, Chenhao Qiu et al.

Robust robotic manipulation requires not only predicting how the scene evolves over time, but also recognizing task-relevant objects in complex scenes. However, existing VLA models face two limitations. They typically act only on the current frame, while future prediction and object-aware reasoning are often learned in separate latent spaces. We propose OFlow (injecting Object-Aware Temporal Flow Matching into VLAs), a framework that addresses both limitations by unifying temporal foresight and object-aware reasoning in a shared semantic latent space. Our method forecasts future latents with temporal flow matching, factorizes them into object-aware representations that emphasize physically relevant cues while filtering task-irrelevant variation, and conditions continuous action generation on these predictions. By integrating OFlow into VLA pipelines, our method enables more reliable control under distribution shifts. Extensive experiments across LIBERO, LIBERO-Plus, MetaWorld, and SimplerEnv benchmarks and real-world tasks demonstrate that object-aware foresight consistently enhances robustness and success.

Yuqian Fu, Xingyu Qiu, Bin Ren et al.

Cross-Domain Few-Shot Object Detection (CD-FSOD) poses significant challenges to existing object detection and few-shot detection models when applied across domains. In conjunction with NTIRE 2025, we organized the 1st CD-FSOD Challenge, aiming to advance the performance of current object detectors on entirely novel target domains with only limited labeled data. The challenge attracted 152 registered participants, received submissions from 42 teams, and concluded with 13 teams making valid final submissions. Participants approached the task from diverse perspectives, proposing novel models that achieved new state-of-the-art (SOTA) results under both open-source and closed-source settings. In this report, we present an overview of the 1st NTIRE 2025 CD-FSOD Challenge, highlighting the proposed solutions and summarizing the results submitted by the participants.

Tim Broedermann, Christos Sakaridis, Yuqian Fu et al.

Leveraging multiple sensors is crucial for robust semantic perception in autonomous driving, as each sensor type has complementary strengths and weaknesses. However, existing sensor fusion methods often treat sensors uniformly across all conditions, leading to suboptimal performance. By contrast, we propose a novel, condition-aware multimodal fusion approach for robust semantic perception of driving scenes. Our method, CAFuser, uses an RGB camera input to classify environmental conditions and generate a Condition Token that guides the fusion of multiple sensor modalities. We further newly introduce modality-specific feature adapters to align diverse sensor inputs into a shared latent space, enabling efficient integration with a single and shared pre-trained backbone. By dynamically adapting sensor fusion based on the actual condition, our model significantly improves robustness and accuracy, especially in adverse-condition scenarios. CAFuser ranks first on the public MUSES benchmarks, achieving 59.7 PQ for multimodal panoptic and 78.2 mIoU for semantic segmentation, and also sets the new state of the art on DeLiVER. The source code is publicly available at: https://github.com/timbroed/CAFuser.

Yu Li, Haoyu Luo, Yuejin Xie et al.

Evaluating the safety of LLM-based agents is increasingly important because risks in realistic deployments often emerge over multi-step interactions rather than isolated prompts or final responses. Existing trajectory-level benchmarks remain limited by insufficient interaction diversity, coarse observability of safety failures, and weak long-horizon realism. We introduce ATBench, a trajectory-level benchmark for structured, diverse, and realistic evaluation of agent safety. ATBench organizes agentic risk along three dimensions: risk source, failure mode, and real-world harm. Based on this taxonomy, we construct trajectories with heterogeneous tool pools and a long-context delayed-trigger protocol that captures realistic risk emergence across multiple stages. The benchmark contains 1,000 trajectories (503 safe and 497 unsafe), averaging 9.01 turns and 3.95k tokens, with 1,954 invoked tools drawn from pools spanning 2,084 available tools. Data quality is supported by rule-based and LLM-based filtering plus full human audit. Experiments on frontier LLMs, open-source models, and specialized guard systems show that ATBench is challenging even for strong evaluators, while enabling taxonomy-stratified analysis, cross-benchmark comparison, and diagnosis of long-horizon failure patterns.

Yanjun Li, Yuqian Fu, Tianwen Qian et al.

Recent advances in Multimodal Large Language Models (MLLMs) have significantly pushed the frontier of egocentric video question answering (EgocentricQA). However, existing benchmarks and studies are mainly limited to common daily activities such as cooking and cleaning. In contrast, real-world deployment inevitably encounters domain shifts, where target domains differ substantially in both visual style and semantic content. To bridge this gap, we introduce \textbf{EgoCross}, a comprehensive benchmark designed to evaluate the cross-domain generalization of MLLMs in EgocentricQA. EgoCross covers four diverse and challenging domains, including surgery, industry, extreme sports, and animal perspective, representing realistic and high-impact application scenarios. It comprises approximately 1,000 QA pairs across 798 video clips, spanning four key QA tasks: prediction, recognition, localization, and counting. Each QA pair provides both OpenQA and CloseQA formats to support fine-grained evaluation. Extensive experiments show that most existing MLLMs, whether general-purpose or egocentric-specialized, struggle to generalize to domains beyond daily life, highlighting the limitations of current models. Furthermore, we conduct several pilot studies, \eg, fine-tuning and reinforcement learning, to explore potential improvements. We hope EgoCross and our accompanying analysis will serve as a foundation for advancing domain-adaptive, robust egocentric video understanding. Data and codes will be released at: \href{https://github.com/MyUniverse0726/EgoCross}{https://github.com/MyUniverse0726/EgoCross.}

Kailing Li, Qi'ao Xu, Tianwen Qian et al.

Embodied Visual Reasoning (EVR) seeks to follow complex, free-form instructions based on egocentric video, enabling semantic understanding and spatiotemporal reasoning in dynamic environments. Despite its promising potential, EVR encounters significant challenges stemming from the diversity of complex instructions and the intricate spatiotemporal dynamics in long-term egocentric videos. Prior solutions either employ Large Language Models (LLMs) over static video captions, which often omit critical visual details, or rely on end-to-end Vision-Language Models (VLMs) that struggle with stepwise compositional reasoning. Consider the complementary strengths of LLMs in reasoning and VLMs in perception, we propose CLiViS. It is a novel training-free framework that leverages LLMs for high-level task planning and orchestrates VLM-driven open-world visual perception to iteratively update the scene context. Building on this synergy, the core of CLiViS is a dynamic Cognitive Map that evolves throughout the reasoning process. This map constructs a structured representation of the embodied scene, bridging low-level perception and high-level reasoning. Extensive experiments across multiple benchmarks demonstrate the effectiveness and generality of CLiViS, especially in handling long-term visual dependencies. Code is available at https://github.com/Teacher-Tom/CLiViS.

Yuli Zhou, Yawei Li, Yuqian Fu et al.

Video camouflaged object segmentation (VCOS), aiming at segmenting camouflaged objects that seamlessly blend into their environment, is a fundamental vision task with various real-world applications. With the release of SAM2, video segmentation has witnessed significant progress. However, SAM2's capability of segmenting camouflaged videos is suboptimal, especially when given simple prompts such as point and box. To address the problem, we propose Camouflaged SAM2 (CamSAM2), which enhances SAM2's ability to handle camouflaged scenes without modifying SAM2's parameters. Specifically, we introduce a decamouflaged token to provide the flexibility of feature adjustment for VCOS. To make full use of fine-grained and high-resolution features from the current frame and previous frames, we propose implicit object-aware fusion (IOF) and explicit object-aware fusion (EOF) modules, respectively. Object prototype generation (OPG) is introduced to abstract and memorize object prototypes with informative details using high-quality features from previous frames. Extensive experiments are conducted to validate the effectiveness of our approach. While CamSAM2 only adds negligible learnable parameters to SAM2, it substantially outperforms SAM2 on three VCOS datasets, especially achieving 12.2 mDice gains with click prompt on MoCA-Mask and 19.6 mDice gains with mask prompt on SUN-SEG-Hard, with Hiera-T as the backbone. The code is available at https://github.com/zhoustan/CamSAM2.

Tao Lin, Yuxin Du, Jiting Liu et al.

Vision-Language-Action models have emerged as a promising paradigm for robotic manipulation by unifying perception, language grounding, and action generation. However, they often struggle in scenarios requiring precise spatial understanding, as current VLA models primarily rely on 2D visual representations that lack depth information and detailed spatial relationships. While recent approaches incorporate explicit 3D inputs such as depth maps or point clouds to address this issue, they often increase system complexity, require additional sensors, and remain vulnerable to sensing noise and reconstruction errors. Another line of work explores implicit 3D-aware spatial modeling directly from RGB observations without extra sensors, but it often relies on large geometry foundation models, resulting in higher training and deployment costs. To address these challenges, we propose Evo-Depth, a lightweight depth-enhanced VLA framework that enhances spatially grounded manipulation without relying on additional sensing hardware or compromising deployment efficiency. Evo-Depth employs a lightweight Implicit Depth Encoding Module to extract compact depth features from multi-view RGB images. These features are incorporated into vision-language representations through a Spatial Enhancement Module via depth-aware modulation, enabling efficient spatial-semantic enhancement. A Progressive Alignment Training strategy is further introduced to align the resulting depth-enhanced representations with downstream action learning. With only 0.9B parameters, Evo-Depth achieves superior performance across four simulation benchmarks. In real-world experiments, Evo-Depth attains the highest average success rate while also exhibiting the smallest model size, lowest GPU memory usage, and highest inference frequency among compared methods.

Yifei Wang, Zhenkai Li, Tianwen Qian et al.

As embodied intelligence advances toward real-world deployment, the ability to continuously perceive and reason over streaming visual inputs becomes essential. In such settings, an agent must maintain situational awareness of its environment, comprehend the interactions with surrounding entities, and dynamically plan actions informed by past observations, current contexts, and anticipated future events. To facilitate progress in this direction, we introduce StreamEQA, the first benchmark designed for streaming video question answering in embodied scenarios. StreamEQA evaluates existing MLLMs along two orthogonal dimensions: Embodied and Streaming. Along the embodied dimension, we categorize the questions into three levels: perception, interaction, and planning, which progressively assess a model's ability to recognize fine-grained visual details, reason about agent-object interactions, and perform high-level goal-directed reasoning. For the streaming dimension, questions are divided into backward, real-time, and forward reasoning, with each mode relying on a distinct temporal context. Built upon 156 independent long videos, StreamEQA defines 42 tasks and generates approximately 21K question-answer pairs with precise timestamps through a hybrid pipeline combining automated generation and human refinement. Evaluations of 13 state-of-the-art video-LLMs reveal that, despite strong performance on conventional benchmarks, these models still struggle with streaming video understanding in embodied scenarios. We hope StreamEQA will catalyze research on streaming video understanding for embodied applications.

Xu Zheng, Zihao Dongfang, Lutao Jiang et al.

Humans possess spatial reasoning abilities that enable them to understand spaces through multimodal observations, such as vision and sound. Large multimodal reasoning models extend these abilities by learning to perceive and reason, showing promising performance across diverse spatial tasks. However, systematic reviews and publicly available benchmarks for these models remain limited. In this survey, we provide a comprehensive review of multimodal spatial reasoning tasks with large models, categorizing recent progress in multimodal large language models (MLLMs) and introducing open benchmarks for evaluation. We begin by outlining general spatial reasoning, focusing on post-training techniques, explainability, and architecture. Beyond classical 2D tasks, we examine spatial relationship reasoning, scene and layout understanding, as well as visual question answering and grounding in 3D space. We also review advances in embodied AI, including vision-language navigation and action models. Additionally, we consider emerging modalities such as audio and egocentric video, which contribute to novel spatial understanding through new sensors. We believe this survey establishes a solid foundation and offers insights into the growing field of multimodal spatial reasoning. Updated information about this survey, codes and implementation of the open benchmarks can be found at https://github.com/zhengxuJosh/Awesome-Spatial-Reasoning.

Bin Ren, Yawei Li, Xu Zheng et al.

Image Restoration (IR) aims to recover high quality images from degraded inputs affected by various corruptions such as noise, blur, haze, rain, and low light conditions. Despite recent advances, most existing approaches treat IR as a direct mapping problem, relying on shared representations across degradation types without modeling their structural diversity. In this work, we present MIRAGE, a unified and lightweight framework for all in one IR that explicitly decomposes the input feature space into three semantically aligned parallel branches, each processed by a specialized module attention for global context, convolution for local textures, and MLP for channel-wise statistics. This modular decomposition significantly improves generalization and efficiency across diverse degradations. Furthermore, we introduce a cross layer contrastive learning scheme that aligns shallow and latent features to enhance the discriminability of shared representations. To better capture the underlying geometry of feature representations, we perform contrastive learning in a Symmetric Positive Definite (SPD) manifold space rather than the conventional Euclidean space. Extensive experiments show that MIRAGE not only achieves new state of the art performance across a variety of degradation types but also offers a scalable solution for challenging all-in-one IR scenarios. Our code and models will be publicly available at https://amazingren.github.io/MIRAGE/.

Weiyu Ma, Yuqian Fu, Zecheng Zhang et al.

We introduce Attentive VLM Agent (AVA), a multimodal StarCraft II agent that aligns artificial agent perception with the human gameplay experience. Traditional frameworks such as SMAC rely on abstract state representations that diverge significantly from human perception, limiting the ecological validity of agent behavior. Our agent addresses this limitation by incorporating RGB visual inputs and natural language observations that more closely simulate human cognitive processes during gameplay. The AVA architecture consists of three integrated components: (1) a vision-language model enhanced with specialized self-attention mechanisms for strategic unit targeting and battlefield assessment, (2) a retrieval-augmented generation system that leverages domain-specific StarCraft II knowledge to inform tactical decisions, and (3) a dynamic role-based task distribution system that enables coordinated multi-agent behavior. The experimental evaluation in our proposed AVACraft environment, which contains 21 multimodal StarCraft II scenarios, demonstrates that AVA powered by foundation models (specifically Qwen-VL and GPT-4o) can execute complex tactical maneuvers without explicit training, achieving comparable performance to traditional MARL methods that require substantial training iterations. This work establishes a foundation for developing human-aligned StarCraft II agents and advances the broader research agenda of multimodal game AI. Our implementation is available at https://github.com/camel-ai/VLM-Play-StarCraft2.

Yuxin Du, Tao Lin, Zile Zhong et al.

Monocular depth foundation models generalize well across scenes, yet they are typically optimized with uniform pixel-wise objectives that do not distinguish user-specified or task-relevant target regions from the surrounding context. We therefore introduce Focusable Monocular Depth Estimation (FDE), a region-aware depth estimation task in which, given a specified target region, the model is required to prioritize foreground depth accuracy, preserve sharp boundary transitions, and maintain coherent global scene geometry. To prioritize task-critical region modeling, we propose FocusDepth, a prompt-conditioned monocular relative depth estimation framework that guides depth modeling to focus on target regions via box/text prompts. The core Multi-Scale Spatial-Aligned Fusion (MSSA) in FocusDepth spatially aligns multi-scale features from Segment Anything Model 3 to the Depth Anything family and injects them through scale-specific, gated conditional fusion. This enables dense prompt cue injection without disrupting geometric representations, thereby endowing the depth estimation model with focused perception capability. To study FDE, we establish FDE-Bench, a target-centric monocular relative depth benchmark built from image-target-depth triplets across five datasets, containing 252.9K/72.5K train/val triplets and 972 categories spanning real-world and embodied simulation environments. On FDE-Bench, FocusDepth consistently improves over globally fine-tuned DA2/DA3 baselines under both box and text prompts, with the largest gains appearing in target boundary and foreground regions while preserving global scene geometry. Ablations show that MSSA's spatial alignment is the key design factor, as disrupting prompt-geometry correspondence increases AbsRel by up to 13.8%.

Kunyu Peng, Di Wen, Jia Fu et al.

Referring Atomic Video Action Recognition (RAVAR) aims to recognize fine-grained, atomic-level actions of a specific person of interest conditioned on natural language descriptions. Distinct from conventional action recognition and detection tasks, RAVAR emphasizes precise language-guided action understanding, which is particularly critical for interactive human action analysis in complex multi-person scenarios. In this work, we extend our previously introduced RefAVA dataset to RefAVA++, which comprises >2.9 million frames and >75.1k annotated persons in total. We benchmark this dataset using baselines from multiple related domains, including atomic action localization, video question answering, and text-video retrieval, as well as our earlier model, RefAtomNet. Although RefAtomNet surpasses other baselines by incorporating agent attention to highlight salient features, its ability to align and retrieve cross-modal information remains limited, leading to suboptimal performance in localizing the target person and predicting fine-grained actions. To overcome the aforementioned limitations, we introduce RefAtomNet++, a novel framework that advances cross-modal token aggregation through a multi-hierarchical semantic-aligned cross-attention mechanism combined with multi-trajectory Mamba modeling at the partial-keyword, scene-attribute, and holistic-sentence levels. In particular, scanning trajectories are constructed by dynamically selecting the nearest visual spatial tokens at each timestep for both partial-keyword and scene-attribute levels. Moreover, we design a multi-hierarchical semantic-aligned cross-attention strategy, enabling more effective aggregation of spatial and temporal tokens across different semantic hierarchies. Experiments show that RefAtomNet++ establishes new state-of-the-art results. The dataset and code are released at https://github.com/KPeng9510/refAVA2.

Chenfei Liao, Wensong Wang, Zichen Wen et al.

Recent endeavors to accelerate inference in Multimodal Large Language Models (MLLMs) have primarily focused on visual token compression. The effectiveness of these methods is typically assessed by measuring the accuracy drop on established benchmarks, comparing model performance before and after compression. However, these benchmarks are originally designed to assess the perception and reasoning capabilities of MLLMs, rather than to evaluate compression techniques. As a result, directly applying them to visual token compression introduces a task mismatch. Strikingly, our investigation reveals that simple image downsampling consistently outperforms many advanced compression methods across multiple widely used benchmarks. Through extensive experiments, we make the following observations: (i) Current benchmarks are noisy for the visual token compression task. (ii) Down-sampling is able to serve as a data filter to evaluate the difficulty of samples in the visual token compression task. Motivated by these findings, we introduce VTC-Bench, an evaluation framework that incorporates a data filtering mechanism to denoise existing benchmarks, thereby enabling fairer and more accurate assessment of visual token compression methods. All data and code are available at https://github.com/Chenfei-Liao/VTC-Bench.

Hongyi Cai, Yuqian Fu, Hongming Fu et al.

Instruction tuning is crucial for optimizing Large Language Models (LLMs), yet mainstream data selection methods heavily rely on LLMs as instruction quality scorers, leading to high computational costs and reduced data diversity. To address these limitations, we propose MergeIT, a novel LLM-based Merging strategy for better Instruction Tuning that shifts the focus from selection to synthesis. MergeIT operates in two stages: first, topic-aware filtering clusters and refines the dataset, preserving diversity while eliminating redundancy without relying on LLM-based scoring. Second, LLM-based merging synthesizes semantically similar instructions into more informative and compact training data, enhancing data richness while further reducing dataset size. Experimental results demonstrate that MergeIT enables efficient, diverse, and scalable instruction selection and synthesis, establishing LLM-based merging as a promising alternative to conventional scoring-based selection methods for instruction tuning. Our source code and datasets are now available at https://github.com/XcloudFance/MergeIT

Yuqian Fu, Yanwei Fu, Yu-Gang Jiang

A recent study finds that existing few-shot learning methods, trained on the source domain, fail to generalize to the novel target domain when a domain gap is observed. This motivates the task of Cross-Domain Few-Shot Learning (CD-FSL). In this paper, we realize that the labeled target data in CD-FSL has not been leveraged in any way to help the learning process. Thus, we advocate utilizing few labeled target data to guide the model learning. Technically, a novel meta-FDMixup network is proposed. We tackle this problem mainly from two aspects. Firstly, to utilize the source and the newly introduced target data of two different class sets, a mixup module is re-proposed and integrated into the meta-learning mechanism. Secondly, a novel disentangle module together with a domain classifier is proposed to extract the disentangled domain-irrelevant and domain-specific features. These two modules together enable our model to narrow the domain gap thus generalizing well to the target datasets. Additionally, a detailed feasibility and pilot study is conducted to reflect the intuitive understanding of CD-FSL under our new setting. Experimental results show the effectiveness of our new setting and the proposed method. Codes and models are available at https://github.com/lovelyqian/Meta-FDMixup.

Yuqian Fu, Yu Wang, Yixuan Pan et al.

This paper studies the challenging cross-domain few-shot object detection (CD-FSOD), aiming to develop an accurate object detector for novel domains with minimal labeled examples. While transformer-based open-set detectors, such as DE-ViT, show promise in traditional few-shot object detection, their generalization to CD-FSOD remains unclear: 1) can such open-set detection methods easily generalize to CD-FSOD? 2) If not, how can models be enhanced when facing huge domain gaps? To answer the first question, we employ measures including style, inter-class variance (ICV), and indefinable boundaries (IB) to understand the domain gap. Based on these measures, we establish a new benchmark named CD-FSOD to evaluate object detection methods, revealing that most of the current approaches fail to generalize across domains. Technically, we observe that the performance decline is associated with our proposed measures: style, ICV, and IB. Consequently, we propose several novel modules to address these issues. First, the learnable instance features align initial fixed instances with target categories, enhancing feature distinctiveness. Second, the instance reweighting module assigns higher importance to high-quality instances with slight IB. Third, the domain prompter encourages features resilient to different styles by synthesizing imaginary domains without altering semantic contents. These techniques collectively contribute to the development of the Cross-Domain Vision Transformer for CD-FSOD (CD-ViTO), significantly improving upon the base DE-ViT. Experimental results validate the efficacy of our model.

Yuqian Fu, Tinghong Chen, Jiajun Chai et al.

Large language models (LLMs) have achieved remarkable progress in reasoning tasks, yet the optimal integration of Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) remains a fundamental challenge. Through comprehensive analysis of token distributions, learning dynamics, and integration mechanisms from entropy-based perspectives, we reveal key differences between these paradigms: SFT induces coarse-grained global changes to LLM policy distributions, while RL performs fine-grained selective optimizations, with entropy serving as a critical indicator of training effectiveness. Building on these observations, we propose Supervised Reinforcement Fine-Tuning (SRFT), a single-stage method that unifies both fine-tuning paradigms through entropy-aware weighting mechanisms. Our approach simultaneously applies SFT and RL to directly optimize the LLM using demonstrations and self-exploration rollouts rather than through two-stage sequential methods. Extensive experiments show that SRFT achieves 59.1% average accuracy, outperforming zero-RL methods by 9.0% on five mathematical reasoning benchmarks and 10.9% on three out-of-distribution benchmarks.

Songjun Tu, Jiahao Lin, Xiangyu Tian et al.

Recent advancements in post-training methodologies for large language models (LLMs) have highlighted reinforcement learning (RL) as a critical component for enhancing reasoning. However, the substantial computational costs associated with RL-based approaches have led to growing interest in alternative paradigms, such as Direct Preference Optimization (DPO). In this study, we investigate the effectiveness of DPO in facilitating self-improvement for LLMs through iterative preference-based learning. We demonstrate that a single round of DPO with coarse filtering significantly enhances mathematical reasoning performance, particularly for strong base model. Furthermore, we design an iterative enhancement framework for both the generator and the reward model (RM), enabling their mutual improvement through online interaction across multiple rounds of DPO. Finally, with simple verifiable rewards, our model DPO-VP achieves RL-level performance with significantly lower computational overhead. These findings highlight DPO as a scalable and cost-effective alternative to RL, offering a practical solution for enhancing LLM reasoning in resource-constrained situations.

Yuqian Fu, Yuanheng Zhu, Jiajun Chai et al.

Ensembling large language models (LLMs) can effectively combine diverse strengths of different models, offering a promising approach to enhance performance across various tasks. However, existing methods typically rely on fixed weighting strategies that fail to adapt to the dynamic, context-dependent characteristics of LLM capabilities. In this work, we propose Reinforcement Learning-Assisted Ensemble for LLMs (RLAE), a novel framework that reformulates LLM ensemble through the lens of a Markov Decision Process (MDP). Our approach introduces a RL agent that dynamically adjusts ensemble weights by considering both input context and intermediate generation states, with the agent being trained using rewards that directly correspond to the quality of final outputs. We implement RLAE using both single-agent and multi-agent reinforcement learning algorithms ($\text{RLAE}_\text{PPO}$ and $\text{RLAE}_\text{MAPPO}$ ), demonstrating substantial improvements over conventional ensemble methods. Extensive evaluations on a diverse set of tasks show that RLAE outperforms existing approaches by up to $3.3\%$ accuracy points, offering a more effective framework for LLM ensembling. Furthermore, our method exhibits superior generalization capabilities across different tasks without the need for retraining, while simultaneously achieving lower time latency.

Jianxiong Gao, Yuqian Fu, Yun Wang et al.

In this paper, we introduce Recon3DMind, an innovative task aimed at reconstructing 3D visuals from Functional Magnetic Resonance Imaging (fMRI) signals, marking a significant advancement in the fields of cognitive neuroscience and computer vision. To support this pioneering task, we present the fMRI-Shape dataset, which includes data from 14 participants and features 360-degree videos of 3D objects to enable comprehensive fMRI signal capture across various settings, thereby laying a foundation for future research. Furthermore, we propose MinD-3D, a novel and effective three-stage framework specifically designed to decode the brain's 3D visual information from fMRI signals, demonstrating the feasibility of this challenging task. The framework begins by extracting and aggregating features from fMRI frames through a neuro-fusion encoder, subsequently employs a feature bridge diffusion model to generate visual features, and ultimately recovers the 3D object via a generative transformer decoder. We assess the performance of MinD-3D using a suite of semantic and structural metrics and analyze the correlation between the features extracted by our model and the visual regions of interest (ROIs) in fMRI signals. Our findings indicate that MinD-3D not only reconstructs 3D objects with high semantic relevance and spatial similarity but also significantly enhances our understanding of the human brain's capabilities in processing 3D visual information. Project page at: https://jianxgao.github.io/MinD-3D.

Xu Zheng, Chenfei Liao, Yuqian Fu et al.

Recent advances in Multimodal Large Language Models (MLLMs) have shown promising results in integrating diverse modalities such as texts and images. MLLMs are heavily influenced by modality bias, often relying on language while under-utilizing other modalities like visual inputs. This position paper argues that MLLMs are deeply affected by modality bias. Firstly, we diagnose the current state of modality bias, highlighting its manifestations across various tasks. Secondly, we propose a systematic research road-map related to modality bias in MLLMs. Thirdly, we identify key factors of modality bias in MLLMs and offer actionable suggestions for future research to mitigate it. To substantiate these findings, we conduct experiments that demonstrate the influence of each factor: 1. Data Characteristics: Language data is compact and abstract, while visual data is redundant and complex, creating an inherent imbalance in learning dynamics. 2. Imbalanced Backbone Capabilities: The dominance of pretrained language models in MLLMs leads to overreliance on language and neglect of visual information. 3. Training Objectives: Current objectives often fail to promote balanced cross-modal alignment, resulting in shortcut learning biased toward language. These findings highlight the need for balanced training strategies and model architectures to better integrate multiple modalities in MLLMs. We call for interdisciplinary efforts to tackle these challenges and drive innovation in MLLM research. Our work provides a fresh perspective on modality bias in MLLMs and offers insights for developing more robust and generalizable multimodal systems-advancing progress toward Artificial General Intelligence.

Jiawei Wang, Jiacai Liu, Yuqian Fu et al.

In long-horizon tasks, recent agents based on Large Language Models (LLMs) face a significant challenge that sparse, outcome-based rewards make it difficult to assign credit to intermediate steps. Previous methods mainly focus on creating dense reward signals to guide learning, either through traditional reinforcement learning techniques like inverse reinforcement learning or by using Process Reward Models for step-by-step feedback. In this paper, we identify a fundamental problem in the learning dynamics of LLMs: the magnitude of policy gradients is inherently coupled with the entropy, which leads to inefficient small updates for confident correct actions and potentially destabilizes large updates for uncertain ones. To resolve this, we propose Entropy-Modulated Policy Gradients (EMPG), a framework that re-calibrates the learning signal based on step-wise uncertainty and the final task outcome. EMPG amplifies updates for confident correct actions, penalizes confident errors, and attenuates updates from uncertain steps to stabilize exploration. We further introduce a bonus term for future clarity that encourages agents to find more predictable solution paths. Through comprehensive experiments on three challenging agent tasks, WebShop, ALFWorld, and Deep Search, we demonstrate that EMPG achieves substantial performance gains and significantly outperforms strong policy gradient baselines. Project page is at https://empgseed-seed.github.io/

Linhai Zhuo, Zheng Wang, Yuqian Fu et al.

The source-free cross-domain few-shot learning (CD-FSL) task aims to transfer pretrained models to target domains utilizing minimal samples, eliminating the need for source domain data. Addressing this issue requires models to have robust generalization abilities and strong feature representation, aligning with the characteristics of large-scale pretrained models. However, large-scale models tend to lose representational ability in cross-domain scenarios due to limited sample diversity. \zlh{Given the abundant diversity provided by semantic modality, this paper leverages textual modality to enhance training sample diversity with CLP model}, meanwhile improving model transfer efficiency. Specifically, we propose the SeGD-VPT framework, which is divided into two phases. The first step aims to increase feature diversity by adding diversity prompts to each support sample, thereby generating varying input and enhancing sample diversity. Furthermore, we use diversity descriptions of classes to guide semantically meaningful learning of diversity prompts, proposing random combinations and selections of texts to increase textual diversity. Additionally, deep prompt tuning is introduced to enhance the model's transfer capability. After training of the first step, support samples with different diversity prompts are input into the CLIP backbone to generate enhanced features. After generation, the second phase trains classifiers using the generated features. Extensive experimental results across several benchmarks verify our method is comparable to SOTA source-utilized models and attain the best performance under the source-free CD-FSL setting.

Kaixuan Xu, Jiajun Chai, Sicheng Li et al.

Diplomacy is a complex multiplayer game that requires both cooperation and competition, posing significant challenges for AI systems. Traditional methods rely on equilibrium search to generate extensive game data for training, which demands substantial computational resources. Large Language Models (LLMs) offer a promising alternative, leveraging pre-trained knowledge to achieve strong performance with relatively small-scale fine-tuning. However, applying LLMs to Diplomacy remains challenging due to the exponential growth of possible action combinations and the intricate strategic interactions among players. To address this challenge, we propose DipLLM, a fine-tuned LLM-based agent that learns equilibrium policies for Diplomacy. DipLLM employs an autoregressive factorization framework to simplify the complex task of multi-unit action assignment into a sequence of unit-level decisions. By defining an equilibrium policy within this framework as the learning objective, we fine-tune the model using only 1.5% of the data required by the state-of-the-art Cicero model, surpassing its performance. Our results demonstrate the potential of fine-tuned LLMs for tackling complex strategic decision-making in multiplayer games.

Chengxing Xie, Xiaoming Zhang, Linze Li et al.

Image super-resolution (SR) has significantly advanced through the adoption of Transformer architectures. However, conventional techniques aimed at enlarging the self-attention window to capture broader contexts come with inherent drawbacks, especially the significantly increased computational demands. Moreover, the feature perception within a fixed-size window of existing models restricts the effective receptive field (ERF) and the intermediate feature diversity. We demonstrate that a flexible integration of attention across diverse spatial extents can yield significant performance enhancements. In line with this insight, we introduce Multi-Range Attention Transformer (MAT) for SR tasks. MAT leverages the computational advantages inherent in dilation operation, in conjunction with self-attention mechanism, to facilitate both multi-range attention (MA) and sparse multi-range attention (SMA), enabling efficient capture of both regional and sparse global features. Combined with local feature extraction, MAT adeptly capture dependencies across various spatial ranges, improving the diversity and efficacy of its feature representations. We also introduce the MSConvStar module, which augments the model's ability for multi-range representation learning. Comprehensive experiments show that our MAT exhibits superior performance to existing state-of-the-art SR models with remarkable efficiency (~3.3 faster than SRFormer-light).

Jiancheng Pan, Shiye Lei, Yuqian Fu et al.

Remote sensing image (RSI) interpretation typically faces challenges due to the scarcity of labeled data, which limits the performance of RSI interpretation tasks. To tackle this challenge, we propose EarthSynth, a diffusion-based generative foundation model that enables synthesizing multi-category, cross-satellite labeled Earth observation for downstream RSI interpretation tasks. To the best of our knowledge, EarthSynth is the first to explore multi-task generation for remote sensing, tackling the challenge of limited generalization in task-oriented synthesis for RSI interpretation. EarthSynth, trained on the EarthSynth-180K dataset, employs the Counterfactual Composition training strategy with a three-dimensional batch-sample selection mechanism to improve training data diversity and enhance category control. Furthermore, a rule-based method of R-Filter is proposed to filter more informative synthetic data for downstream tasks. We evaluate our EarthSynth on scene classification, object detection, and semantic segmentation in open-world scenarios. There are significant improvements in open-vocabulary understanding tasks, offering a practical solution for advancing RSI interpretation.

Yuqian Fu, Runze Wang, Bin Ren et al.

Bridging the gap between ego-centric and exo-centric views has been a long-standing question in computer vision. In this paper, we focus on the emerging Ego-Exo object correspondence task, which aims to understand object relations across ego-exo perspectives through segmentation. While numerous segmentation models have been proposed, most operate on a single image (view), making them impractical for cross-view scenarios. PSALM, a recently proposed segmentation method, stands out as a notable exception with its demonstrated zero-shot ability on this task. However, due to the drastic viewpoint change between ego and exo, PSALM fails to accurately locate and segment objects, especially in complex backgrounds or when object appearances change significantly. To address these issues, we propose ObjectRelator, a novel approach featuring two key modules: Multimodal Condition Fusion (MCFuse) and SSL-based Cross-View Object Alignment (XObjAlign). MCFuse introduces language as an additional cue, integrating both visual masks and textual descriptions to improve object localization and prevent incorrect associations. XObjAlign enforces cross-view consistency through self-supervised alignment, enhancing robustness to object appearance variations. Extensive experiments demonstrate ObjectRelator's effectiveness on the large-scale Ego-Exo4D benchmark and HANDAL-X (an adapted dataset for cross-view segmentation) with state-of-the-art performance. Code is made available at: http://yuqianfu.com/ObjectRelator.