Linjiang Huang

Linjiang Huang, Kaixin Lu, Guanglu Song et al.

In this paper, we present a novel training scheme, namely Teach-DETR, to learn better DETR-based detectors from versatile teacher detectors. We show that the predicted boxes from teacher detectors are effective medium to transfer knowledge of teacher detectors, which could be either RCNN-based or DETR-based detectors, to train a more accurate and robust DETR model. This new training scheme can easily incorporate the predicted boxes from multiple teacher detectors, each of which provides parallel supervisions to the student DETR. Our strategy introduces no additional parameters and adds negligible computational cost to the original detector during training. During inference, Teach-DETR brings zero additional overhead and maintains the merit of requiring no non-maximum suppression. Extensive experiments show that our method leads to consistent improvement for various DETR-based detectors. Specifically, we improve the state-of-the-art detector DINO with Swin-Large backbone, 4 scales of feature maps and 36-epoch training schedule, from 57.8% to 58.9% in terms of mean average precision on MSCOCO 2017 validation set. Code will be available at https://github.com/LeonHLJ/Teach-DETR.

Linjiang Huang, Liang Wang, Hongsheng Li

Weakly supervised temporal action localization aims to localize temporal boundaries of actions and simultaneously identify their categories with only video-level category labels. Many existing methods seek to generate pseudo labels for bridging the discrepancy between classification and localization, but usually only make use of limited contextual information for pseudo label generation. To alleviate this problem, we propose a representative snippet summarization and propagation framework. Our method seeks to mine the representative snippets in each video for propagating information between video snippets to generate better pseudo labels. For each video, its own representative snippets and the representative snippets from a memory bank are propagated to update the input features in an intra- and inter-video manner. The pseudo labels are generated from the temporal class activation maps of the updated features to rectify the predictions of the main branch. Our method obtains superior performance in comparison to the existing methods on two benchmarks, THUMOS14 and ActivityNet1.3, achieving gains as high as 1.2% in terms of average mAP on THUMOS14.

Jingqiu Zhou, Linjiang Huang, Liang Wang et al.

The task of weakly supervised temporal action localization targets at generating temporal boundaries for actions of interest, meanwhile the action category should also be classified. Pseudo-label-based methods, which serve as an effective solution, have been widely studied recently. However, existing methods generate pseudo labels during training and make predictions during testing under different pipelines or settings, resulting in a gap between training and testing. In this paper, we propose to generate high-quality pseudo labels from the predicted action boundaries. Nevertheless, we note that existing post-processing, like NMS, would lead to information loss, which is insufficient to generate high-quality action boundaries. More importantly, transforming action boundaries into pseudo labels is quite challenging, since the predicted action instances are generally overlapped and have different confidence scores. Besides, the generated pseudo-labels can be fluctuating and inaccurate at the early stage of training. It might repeatedly strengthen the false predictions if there is no mechanism to conduct self-correction. To tackle these issues, we come up with an effective pipeline for learning better pseudo labels. Firstly, we propose a Gaussian weighted fusion module to preserve information of action instances and obtain high-quality action boundaries. Second, we formulate the pseudo-label generation as an optimization problem under the constraints in terms of the confidence scores of action instances. Finally, we introduce the idea of $Δ$ pseudo labels, which enables the model with the ability of self-correction. Our method achieves superior performance to existing methods on two benchmarks, THUMOS14 and ActivityNet1.3, achieving gains of 1.9\% on THUMOS14 and 3.7\% on ActivityNet1.3 in terms of average mAP.

Runze He, Kai Ma, Linjiang Huang et al.

Introducing user-specified visual concepts in image editing is highly practical as these concepts convey the user's intent more precisely than text-based descriptions. We propose FreeEdit, a novel approach for achieving such reference-based image editing, which can accurately reproduce the visual concept from the reference image based on user-friendly language instructions. Our approach leverages the multi-modal instruction encoder to encode language instructions to guide the editing process. This implicit way of locating the editing area eliminates the need for manual editing masks. To enhance the reconstruction of reference details, we introduce the Decoupled Residual ReferAttention (DRRA) module. This module is designed to integrate fine-grained reference features extracted by a detail extractor into the image editing process in a residual way without interfering with the original self-attention. Given that existing datasets are unsuitable for reference-based image editing tasks, particularly due to the difficulty in constructing image triplets that include a reference image, we curate a high-quality dataset, FreeBench, using a newly developed twice-repainting scheme. FreeBench comprises the images before and after editing, detailed editing instructions, as well as a reference image that maintains the identity of the edited object, encompassing tasks such as object addition, replacement, and deletion. By conducting phased training on FreeBench followed by quality tuning, FreeEdit achieves high-quality zero-shot editing through convenient language instructions. We conduct extensive experiments to evaluate the effectiveness of FreeEdit across multiple task types, demonstrating its superiority over existing methods. The code will be available at: https://freeedit.github.io/.

Zhen Zou, Xiaoxiao Ma, Mingde Yao et al.

Autoregressive (AR)-Diffusion hybrid paradigms combine AR's structured semantic modeling with diffusion's high-fidelity synthesis, yet suffer from a dual speed bottleneck: the sequential AR stage and the iterative multi-step denoising of the diffusion vision decode stage. Existing methods address each in isolation without a unified principle design. We observe that the per-position \emph{prediction entropy} of continuous-space AR models naturally encodes spatially varying generation uncertainty, which simultaneously governing draft prediction quality in the AR stage and reflecting the corrective effort required by vision decoding stage, which is not fully explored before. Since entropy is inherently tied to both bottlenecks, it serves as a natural unifying signal for joint acceleration. In this work, we propose \textbf{Drift-AR}, which leverages entropy signal to accelerate both stages: 1) for AR acceleration, we introduce Entropy-Informed Speculative Decoding that align draft--target entropy distributions via a causal-normalized entropy loss, resolving the entropy mismatch that causes excessive draft rejection; 2) for visual decoder acceleration, we reinterpret entropy as the \emph{physical variance} of the initial state for an anti-symmetric drifting field -- high-entropy positions activate stronger drift toward the data manifold while low-entropy positions yield vanishing drift -- enabling single-step (1-NFE) decoding without iterative denoising or distillation. Moreover, both stages share the same entropy signal, which is computed once with no extra cost. Experiments on MAR, TransDiff, and NextStep-1 demonstrate 3.8--5.5$\times$ speedup with genuine 1-NFE decoding, matching or surpassing original quality. Code will be available at https://github.com/aSleepyTree/Drift-AR.

Rongyao Fang, Chengqi Duan, Kun Wang et al.

Current image generation and editing methods primarily process textual prompts as direct inputs without reasoning about visual composition and explicit operations. We present Generation Chain-of-Thought (GoT), a novel paradigm that enables generation and editing through an explicit language reasoning process before outputting images. This approach transforms conventional text-to-image generation and editing into a reasoning-guided framework that analyzes semantic relationships and spatial arrangements. We define the formulation of GoT and construct large-scale GoT datasets containing over 9M samples with detailed reasoning chains capturing semantic-spatial relationships. To leverage the advantages of GoT, we implement a unified framework that integrates Qwen2.5-VL for reasoning chain generation with an end-to-end diffusion model enhanced by our novel Semantic-Spatial Guidance Module. Experiments show our GoT framework achieves excellent performance on both generation and editing tasks, with significant improvements over baselines. Additionally, our approach enables interactive visual generation, allowing users to explicitly modify reasoning steps for precise image adjustments. GoT pioneers a new direction for reasoning-driven visual generation and editing, producing images that better align with human intent. To facilitate future research, we make our datasets, code, and pretrained models publicly available at https://github.com/rongyaofang/GoT.

Chengqi Duan, Rongyao Fang, Yuqing Wang et al.

Visual generation models have made remarkable progress in creating realistic images from text prompts, yet struggle with complex prompts that specify multiple objects with precise spatial relationships and attributes. Effective handling of such prompts requires explicit reasoning about the semantic content and spatial layout. We present GoT-R1, a framework that applies reinforcement learning to enhance semantic-spatial reasoning in visual generation. Building upon the Generation Chain-of-Thought approach, GoT-R1 enables models to autonomously discover effective reasoning strategies beyond predefined templates through carefully designed reinforcement learning. To achieve this, we propose a dual-stage multi-dimensional reward framework that leverages MLLMs to evaluate both the reasoning process and final output, enabling effective supervision across the entire generation pipeline. The reward system assesses semantic alignment, spatial accuracy, and visual quality in a unified approach. Experimental results demonstrate significant improvements on T2I-CompBench benchmark, particularly in compositional tasks involving precise spatial relationships and attribute binding. GoT-R1 advances the state-of-the-art in image generation by successfully transferring sophisticated reasoning capabilities to the visual generation domain. To facilitate future research, we make our code and pretrained models publicly available at https://github.com/gogoduan/GoT-R1.

Hang Xu, Linjiang Huang, Feng Zhao

Test-time scaling (TTS) aims to achieve better results by increasing random sampling and evaluating samples based on rules and metrics. However, in text-to-image(T2I) diffusion models, most related works focus on search strategies and reward models, yet the impact of the stochastic characteristic of noise in T2I diffusion models on the method's performance remains unexplored. In this work, we analyze the effects of randomness in T2I diffusion models and explore a new format of randomness for TTS: text embedding perturbation, which couples with existing randomness like SDE-injected noise to enhance generative diversity and quality. We start with a frequency-domain analysis of these formats of randomness and their impact on generation, and find that these two randomness exhibit complementary behavior in the frequency domain: spatial noise favors low-frequency components (early steps), while text embedding perturbation enhances high-frequency details (later steps), thereby compensating for the potential limitations of spatial noise randomness in high-frequency manipulation. Concurrently, text embedding demonstrates varying levels of tolerance to perturbation across different dimensions of the generation process. Specifically, our method consists of two key designs: (1) Introducing step-based text embedding perturbation, combining frequency-guided noise schedules with spatial noise perturbation. (2) Adapting the perturbation intensity selectively based on their frequency-specific contributions to generation and tolerance to perturbation. Our approach can be seamlessly integrated into existing TTS methods and demonstrates significant improvements on multiple benchmarks with almost no additional computation. Code is available at \href{https://github.com/xuhang07/TEP-Diffusion}{https://github.com/xuhang07/TEP-Diffusion}.

Rongyao Fang, Aldrich Yu, Chengqi Duan et al.

The advancement of open-source text-to-image (T2I) models has been hindered by the absence of large-scale, reasoning-focused datasets and comprehensive evaluation benchmarks, resulting in a performance gap compared to leading closed-source systems. To address this challenge, We introduce FLUX-Reason-6M and PRISM-Bench (Precise and Robust Image Synthesis Measurement Benchmark). FLUX-Reason-6M is a massive dataset consisting of 6 million high-quality FLUX-generated images and 20 million bilingual (English and Chinese) descriptions specifically designed to teach complex reasoning. The image are organized according to six key characteristics: Imagination, Entity, Text rendering, Style, Affection, and Composition, and design explicit Generation Chain-of-Thought (GCoT) to provide detailed breakdowns of image generation steps. The whole data curation takes 15,000 A100 GPU days, providing the community with a resource previously unattainable outside of large industrial labs. PRISM-Bench offers a novel evaluation standard with seven distinct tracks, including a formidable Long Text challenge using GCoT. Through carefully designed prompts, it utilizes advanced vision-language models for nuanced human-aligned assessment of prompt-image alignment and image aesthetics. Our extensive evaluation of 19 leading models on PRISM-Bench reveals critical performance gaps and highlights specific areas requiring improvement. Our dataset, benchmark, and evaluation code are released to catalyze the next wave of reasoning-oriented T2I generation. Project page: https://flux-reason-6m.github.io/ .

Haojun Xu, Jiaqi Xiang, Wu Wei et al.

A typical human strategy for giving navigation guidance is to sketch route maps based on the environmental layout. Inspired by this, we introduce Sketch map-based visual Navigation (SkeNa), an embodied navigation task in which an agent must reach a goal in an unseen environment using only a hand-drawn sketch map as guidance. To support research for SkeNa, we present a large-scale dataset named SoR, comprising 54k trajectory and sketch map pairs across 71 indoor scenes. In SoR, we introduce two navigation validation sets with varying levels of abstraction in hand-drawn sketches, categorized based on their preservation of spatial scales in the environment, to facilitate future research. To construct SoR, we develop an automated sketch-generation pipeline that efficiently converts floor plans into hand-drawn representations. To solve SkeNa, we propose SkeNavigator, a navigation framework that aligns visual observations with hand-drawn maps to estimate navigation targets. It employs a Ray-based Map Descriptor (RMD) to enhance sketch map valid feature representation using equidistant sampling points and boundary distances. To improve alignment with visual observations, a Dual-Map Aligned Goal Predictor (DAGP) leverages the correspondence between sketch map features and on-site constructed exploration map features to predict goal position and guide navigation. SkeNavigator outperforms prior floor plan navigation methods by a large margin, improving SPL on the high-abstract validation set by 105% relatively. Our code and dataset will be released.

Hang Xu, Jie Huang, Linjiang Huang et al.

Domain Adaptation(DA) for dense prediction tasks is an important topic, which enhances the dense prediction model's performance when tested on its unseen domain. Recently, with the development of Diffusion-based Dense Prediction (DDP) models, the exploration of DA designs tailored to this framework is worth exploring, since the diffusion model is effective in modeling the distribution transformation that comprises domain information. In this work, we propose a training-free mechanism for DDP frameworks, endowing them with DA capabilities. Our motivation arises from the observation that the exposure bias (e.g., noise statistics bias) in diffusion brings domain shift, and different domains in conditions of DDP models can also be effectively captured by the noise prediction statistics. Based on this, we propose a training-free Domain Noise Alignment (DNA) approach, which alleviates the variations of noise statistics to domain changes during the diffusion sampling process, thereby achieving domain adaptation. Specifically, when the source domain is available, we directly adopt the DNA method to achieve domain adaptation by aligning the noise statistics of the target domain with those of the source domain. For the more challenging source-free DA, inspired by the observation that regions closer to the source domain exhibit higher confidence meeting variations of sampling noise, we utilize the statistics from the high-confidence regions progressively to guide the noise statistic adjustment during the sampling process. Notably, our method demonstrates the effectiveness of enhancing the DA capability of DDP models across four common dense prediction tasks. Code is available at \href{https://github.com/xuhang07/FreeDNA}{https://github.com/xuhang07/FreeDNA}.

Linjiang Huang, Rongyao Fang, Aiping Zhang et al.

In this study, we delve into the generation of high-resolution images from pre-trained diffusion models, addressing persistent challenges, such as repetitive patterns and structural distortions, that emerge when models are applied beyond their trained resolutions. To address this issue, we introduce an innovative, training-free approach FouriScale from the perspective of frequency domain analysis. We replace the original convolutional layers in pre-trained diffusion models by incorporating a dilation technique along with a low-pass operation, intending to achieve structural consistency and scale consistency across resolutions, respectively. Further enhanced by a padding-then-crop strategy, our method can flexibly handle text-to-image generation of various aspect ratios. By using the FouriScale as guidance, our method successfully balances the structural integrity and fidelity of generated images, achieving an astonishing capacity of arbitrary-size, high-resolution, and high-quality generation. With its simplicity and compatibility, our method can provide valuable insights for future explorations into the synthesis of ultra-high-resolution images. The code will be released at https://github.com/LeonHLJ/FouriScale.

Linjiang Huang, Liang Wang, Hongsheng Li

As a challenging task of high-level video understanding, weakly supervised temporal action localization has been attracting increasing attention. With only video annotations, most existing methods seek to handle this task with a localization-by-classification framework, which generally adopts a selector to select snippets of high probabilities of actions or namely the foreground. Nevertheless, the existing foreground selection strategies have a major limitation of only considering the unilateral relation from foreground to actions, which cannot guarantee the foreground-action consistency. In this paper, we present a framework named FAC-Net based on the I3D backbone, on which three branches are appended, named class-wise foreground classification branch, class-agnostic attention branch and multiple instance learning branch. First, our class-wise foreground classification branch regularizes the relation between actions and foreground to maximize the foreground-background separation. Besides, the class-agnostic attention branch and multiple instance learning branch are adopted to regularize the foreground-action consistency and help to learn a meaningful foreground classifier. Within each branch, we introduce a hybrid attention mechanism, which calculates multiple attention scores for each snippet, to focus on both discriminative and less-discriminative snippets to capture the full action boundaries. Experimental results on THUMOS14 and ActivityNet1.3 demonstrate the state-of-the-art performance of our method. Our code is available at https://github.com/LeonHLJ/FAC-Net.

Xuesong Chen, Linjiang Huang, Tao Ma et al.

The integration of Vision-Language Models (VLMs) into autonomous driving systems has shown promise in addressing key challenges such as learning complexity, interpretability, and common-sense reasoning. However, existing approaches often struggle with efficient integration and realtime decision-making due to computational demands. In this paper, we introduce SOLVE, an innovative framework that synergizes VLMs with end-to-end (E2E) models to enhance autonomous vehicle planning. Our approach emphasizes knowledge sharing at the feature level through a shared visual encoder, enabling comprehensive interaction between VLM and E2E components. We propose a Trajectory Chain-of-Thought (T-CoT) paradigm, which progressively refines trajectory predictions, reducing uncertainty and improving accuracy. By employing a temporal decoupling strategy, SOLVE achieves efficient cooperation by aligning high-quality VLM outputs with E2E real-time performance. Evaluated on the nuScenes dataset, our method demonstrates significant improvements in trajectory prediction accuracy, paving the way for more robust and reliable autonomous driving systems.

Jingqiu Zhou, Lue Fan, Xuesong Chen et al.

In this paper, we present GaussianPainter, the first method to paint a point cloud into 3D Gaussians given a reference image. GaussianPainter introduces an innovative feed-forward approach to overcome the limitations of time-consuming test-time optimization in 3D Gaussian splatting. Our method addresses a critical challenge in the field: the non-uniqueness problem inherent in the large parameter space of 3D Gaussian splatting. This space, encompassing rotation, anisotropic scales, and spherical harmonic coefficients, introduces the challenge of rendering similar images from substantially different Gaussian fields. As a result, feed-forward networks face instability when attempting to directly predict high-quality Gaussian fields, struggling to converge on consistent parameters for a given output. To address this issue, we propose to estimate a surface normal for each point to determine its Gaussian rotation. This strategy enables the network to effectively predict the remaining Gaussian parameters in the constrained space. We further enhance our approach with an appearance injection module, incorporating reference image appearance into Gaussian fields via a multiscale triplane representation. Our method successfully balances efficiency and fidelity in 3D Gaussian generation, achieving high-quality, diverse, and robust 3D content creation from point clouds in a single forward pass.

Pengyiang Liu, Zhongyue Shi, Hongye Hao et al.

Video understanding requires models to continuously track and update world state during playback. While existing benchmarks have advanced video understanding evaluation across multiple dimensions, the observation of how models maintain world state remains insufficient. We propose VCBench, a streaming counting benchmark that repositions counting as a minimal probe for diagnosing world state maintenance capability. We decompose this capability into object counting (tracking currently visible objects vs.\ tracking cumulative unique identities) and event counting (detecting instantaneous actions vs.\ tracking complete activity cycles), forming 8 fine-grained subcategories. VCBench contains 406 videos with frame-by-frame annotations of 10,071 event occurrence moments and object state change moments, generating 1,000 streaming QA pairs with 4,576 query points along timelines. By observing state maintenance trajectories through streaming multi-point queries, we design three complementary metrics to diagnose numerical precision, trajectory consistency, and temporal awareness. Evaluation on mainstream video-language models shows that current models still exhibit significant deficiencies in spatial-temporal state maintenance, particularly struggling with tasks like periodic event counting. VCBench provides a diagnostic framework for measuring and improving state maintenance in video understanding systems.

Weikang Shi, Aldrich Yu, Rongyao Fang et al.

While Large Language Models (LLMs) have excelled in textual reasoning, they struggle with mathematical domains like geometry that intrinsically rely on visual aids. Existing approaches to Visual Chain-of-Thought (VCoT) are often limited by rigid external tools or fail to generate the high-fidelity, strategically-timed diagrams necessary for complex problem-solving. To bridge this gap, we introduce MathCanvas, a comprehensive framework designed to endow unified Large Multimodal Models (LMMs) with intrinsic VCoT capabilities for mathematics. Our approach consists of two phases. First, a Visual Manipulation stage pre-trains the model on a novel 15.2M-pair corpus, comprising 10M caption-to-diagram pairs (MathCanvas-Imagen) and 5.2M step-by-step editing trajectories (MathCanvas-Edit), to master diagram generation and editing. Second, a Strategic Visual-Aided Reasoning stage fine-tunes the model on MathCanvas-Instruct, a new 219K-example dataset of interleaved visual-textual reasoning paths, teaching it when and how to leverage visual aids. To facilitate rigorous evaluation, we introduce MathCanvas-Bench, a challenging benchmark with 3K problems that require models to produce interleaved visual-textual solutions. Our model, BAGEL-Canvas, trained under this framework, achieves an 86% relative improvement over strong LMM baselines on MathCanvas-Bench, demonstrating excellent generalization to other public math benchmarks. Our work provides a complete toolkit-framework, datasets, and benchmark-to unlock complex, human-like visual-aided reasoning in LMMs. Project Page: https://mathcanvas.github.io/

Ruipu Wu, Yige Zhang, Jinyu Chen et al.

Aerial Vision-and-Language Navigation (VLN) is an emerging task that enables Unmanned Aerial Vehicles (UAVs) to navigate outdoor environments using natural language instructions and visual cues. However, due to the extended trajectories and complex maneuverability of UAVs, achieving reliable UAV-VLN performance is challenging and often requires human intervention or overly detailed instructions. To harness the advantages of UAVs' high mobility, which could provide multi-grained perspectives, while maintaining a manageable motion space for learning, we introduce a novel task called Dual-Altitude UAV Collaborative VLN (DuAl-VLN). In this task, two UAVs operate at distinct altitudes: a high-altitude UAV responsible for broad environmental reasoning, and a low-altitude UAV tasked with precise navigation. To support the training and evaluation of the DuAl-VLN, we construct the HaL-13k, a dataset comprising 13,838 collaborative high-low UAV demonstration trajectories, each paired with target-oriented language instructions. This dataset includes both unseen maps and an unseen object validation set to systematically evaluate the model's generalization capabilities across novel environments and unfamiliar targets. To consolidate their complementary strengths, we propose a dual-UAV collaborative VLN framework, AeroDuo, where the high-altitude UAV integrates a multimodal large language model (Pilot-LLM) for target reasoning, while the low-altitude UAV employs a lightweight multi-stage policy for navigation and target grounding. The two UAVs work collaboratively and only exchange minimal coordinate information to ensure efficiency.

Jingqiu Zhou, Lue Fan, Linjiang Huang et al.

Driving scene reconstruction and rendering have advanced significantly using the 3D Gaussian Splatting. However, most prior research has focused on the rendering quality along a pre-recorded vehicle path and struggles to generalize to out-of-path viewpoints, which is caused by the lack of high-quality supervision in those out-of-path views. To address this issue, we introduce an Inverse View Warping technique to create compact and high-quality images as supervision for the reconstruction of the out-of-path views, enabling high-quality rendering results for those views. For accurate and robust inverse view warping, a depth bootstrap strategy is proposed to obtain on-the-fly dense depth maps during the optimization process, overcoming the sparsity and incompleteness of LiDAR depth data. Our method achieves superior in-path and out-of-path reconstruction and rendering performance on the widely used Waymo Open dataset. In addition, a simulator-based benchmark is proposed to obtain the out-of-path ground truth and quantitatively evaluate the performance of out-of-path rendering, where our method outperforms previous methods by a significant margin.

Hongyu Li, Manyuan Zhang, Dian Zheng et al.

Instruction-based image editing has emerged as a prominent research area, which, benefiting from image generation foundation models, have achieved high aesthetic quality, making instruction-following capability the primary challenge. Existing approaches improve instruction adherence via supervised or reinforcement learning, yet single-turn success rates remain limited due to inherent stochasticity and a lack of deliberation. In this work, we propose a deliberative editing framework to 'think' while they edit, which simulates the human cognitive loop by iteratively executing a Think-while-Edit cycle: Critiquing results and Refining instructions , followed by Repeating the generation until satisfactory. Specifically, we train a single MLLM, EditThinker, to act as the reasoning engine of this framework, which jointly produce the critique score, reasoning process, and refined instructions. We employ reinforcement learning to align the EditThinker's thinking with its editing, thereby generating more targeted instruction improvements. Extensive experiments on four benchmarks demonstrate that our approach significantly improves the instruction-following capability of any image editing model by a large margin. We will release our data construction framework, datasets, and models to benefit the community.

Yuting Gao, Wang Lan, Hengyuan Zhao et al.

Multimodal Mixture-of-Experts (MoE) models offer a promising path toward scalable and efficient large vision-language systems. However, existing approaches rely on rigid routing strategies (typically activating a fixed number of experts per token) ignoring the inherent heterogeneity in semantic importance across modalities. This leads to suboptimal compute allocation, where redundant tokens consume as many resources as critical ones. To address this, we propose AnyExperts, a novel on-demand, budget-aware dynamic routing framework that allocates a variable total number of expert slots per token based on its semantic importance. Crucially, to prevent uncontrolled compute growth, the total slots per token are constrained within a fixed range, and each slot is filled by either a real expert or a virtual expert, with the virtual share capped at a small maximum (e.g., 20%). The model then adaptively balances the real-to-virtual ratio per token, assigning more real experts to semantically rich regions and relying more on virtual experts for redundant content. Evaluated across diverse tasks in visual understanding, audio understanding, and NLP understanding, AnyExperts improves performance under the same compute budget. Notably, on general image/video tasks, it achieves comparable accuracy with 40% fewer real expert activations; on text-dense tasks (OCR and NLP), it maintains performance while reducing real expert usage by 10%. These results demonstrate that fine-grained, importance-driven expert allocation significantly enhances both the efficiency and effectiveness of multimodal MoE models.

Guohui Zhang, Hu Yu, Xiaoxiao Ma et al.

Recent studies have extended Reinforcement Learning with Verifiable Rewards (RLVR) to autoregressive (AR) visual generation and achieved promising progress. However, existing methods typically apply uniform optimization across all image tokens, while the varying contributions of different image tokens for RLVR's training remain unexplored. In fact, the key obstacle lies in how to identify more critical image tokens during AR generation and implement effective token-wise optimization for them. To tackle this challenge, we propose $\textbf{G}$roup $\textbf{C}$ritical-token $\textbf{P}$olicy $\textbf{O}$ptimization ($\textbf{GCPO}$), which facilitates effective policy optimization on critical tokens. We identify the critical tokens in RLVR-based AR generation from three perspectives, specifically: $\textbf{(1)}$ Causal dependency: early tokens fundamentally determine the later tokens and final image effect due to unidirectional dependency; $\textbf{(2)}$ Entropy-induced spatial structure: tokens with high entropy gradients correspond to image structure and bridges distinct visual regions; $\textbf{(3)}$ RLVR-focused token diversity: tokens with low visual similarity across a group of sampled images contribute to richer token-level diversity. For these identified critical tokens, we further introduce a dynamic token-wise advantage weight to encourage exploration, based on confidence divergence between the policy model and reference model. By leveraging 30\% of the image tokens, GCPO achieves better performance than GRPO with full tokens. Extensive experiments on multiple text-to-image benchmarks for both AR models and unified multimodal models demonstrate the effectiveness of GCPO for AR visual generation.

Qixin Sun, Ziqin Wang, Hengyuan Zhao et al.

Retrieval Augmented Generation enhances the response accuracy of Large Language Models (LLMs) by integrating retrieval and generation modules with external knowledge, demonstrating particular strength in real-time queries and Visual Question Answering tasks. However, the effectiveness of RAG is frequently hindered by the precision of the retriever: many retrieved samples fed into the generation phase are irrelevant or misleading, posing a critical bottleneck to LLMs' performance. To address this challenge, we introduce VaccineRAG, a novel Chain-of-Thought-based retrieval-augmented generation dataset. On one hand, VaccineRAG employs a benchmark to evaluate models using data with varying positive/negative sample ratios, systematically exposing inherent weaknesses in current LLMs. On the other hand, it enhances models' sample-discrimination capabilities by prompting LLMs to generate explicit Chain-of-Thought (CoT) analysis for each sample before producing final answers. Furthermore, to enhance the model's ability to learn long-sequence complex CoT content, we propose Partial-GRPO. By modeling the outputs of LLMs as multiple components rather than a single whole, our model can make more informed preference selections for complex sequences, thereby enhancing its capacity to learn complex CoT. Comprehensive evaluations and ablation studies on VaccineRAG validate the effectiveness of the proposed scheme. The code and dataset will be publicly released soon.

Guohui Zhang, Jiangtong Tan, Linjiang Huang et al.

Diffusion models (DMs) have become dominant in visual generation but suffer performance drop when tested on resolutions that differ from the training scale, whether lower or higher. In fact, the key challenge in generating variable-scale images lies in the differing amounts of information across resolutions, which requires information conversion procedures to be varied for generating variable-scaled images. In this paper, we investigate the issues of three critical aspects in DMs for a unified analysis in variable-scaled generation: dilated convolution, attention mechanisms, and initial noise. Specifically, 1) dilated convolution in DMs for the higher-resolution generation loses high-frequency information. 2) Attention for variable-scaled image generation struggles to adjust the information aggregation adaptively. 3) The spatial distribution of information in the initial noise is misaligned with variable-scaled image. To solve the above problems, we propose \textbf{InfoScale}, an information-centric framework for variable-scaled image generation by effectively utilizing information from three aspects correspondingly. For information loss in 1), we introduce Progressive Frequency Compensation module to compensate for high-frequency information lost by dilated convolution in higher-resolution generation. For information aggregation inflexibility in 2), we introduce Adaptive Information Aggregation module to adaptively aggregate information in lower-resolution generation and achieve an effective balance between local and global information in higher-resolution generation. For information distribution misalignment in 3), we design Noise Adaptation module to re-distribute information in initial noise for variable-scaled generation. Our method is plug-and-play for DMs and extensive experiments demonstrate the effectiveness in variable-scaled image generation.

Jianhua Yang, Yan Huang, Kai Niu et al.

Text-based video segmentation aims to segment an actor in video sequences by specifying the actor and its performing action with a textual query. Previous methods fail to explicitly align the video content with the textual query in a fine-grained manner according to the actor and its action, due to the problem of \emph{semantic asymmetry}. The \emph{semantic asymmetry} implies that two modalities contain different amounts of semantic information during the multi-modal fusion process. To alleviate this problem, we propose a novel actor and action modular network that individually localizes the actor and its action in two separate modules. Specifically, we first learn the actor-/action-related content from the video and textual query, and then match them in a symmetrical manner to localize the target tube. The target tube contains the desired actor and action which is then fed into a fully convolutional network to predict segmentation masks of the actor. Our method also establishes the association of objects cross multiple frames with the proposed temporal proposal aggregation mechanism. This enables our method to segment the video effectively and keep the temporal consistency of predictions. The whole model is allowed for joint learning of the actor-action matching and segmentation, as well as achieves the state-of-the-art performance for both single-frame segmentation and full video segmentation on A2D Sentences and J-HMDB Sentences datasets.