Large language model (LLM) agents increasingly rely on external tools such as search engines to solve complex, multi-step problems, yet their rollouts are structurally heterogeneous: variations in tool-call number, placement, and outcomes induce distinct behaviors and reward distributions. As a result, policy gradient methods with a single global baseline suffer from , an "apples-to-oranges" comparison that distorts credit assignment and impedes exploration. To address this issue, we propose . Its core component, (SAN), partitions trajectories into homogeneous strata based on structural properties and computes advantages locally within each stratum, ensuring comparisons only among true peers. We show that SAN eliminates cross-stratum bias, yields conditionally unbiased unit-variance estimates within strata, and preserves the global unbiasedness and unit-variance properties of standard normalization, resulting in a more reliable learning signal. To improve robustness in finite-sample regimes, we further linearly blend SAN with the global estimator. Experiments on factual QA and deep-research agent benchmarks demonstrate that Stratified GRPO consistently outperforms GRPO by up to 12.6 points, achieving higher training rewards, improved training stability, and more effective search policies. These results establish stratification as a principled remedy for structural heterogeneity in RL for LLM search agents.

Recent unified models have made remarkable strides in generating high-quality images, yet they consistently fail on reasoning-intensive tasks, i.e., solving mazes, assembling tangrams. Intriguingly, we find that vision-language models (VLMs) and large language models (LLMs) can accurately solve these tasks, but cannot generate the corresponding images because they lack a structured visual output interface. This reveals that the core bottleneck is not reasoning capacity, but the lack of a structured interface to translate high-level reasoning into precise visual output. To bridge this gap, we propose using code-structured visual hints (i.e., SVG/HTML) overlays that explicitly encode reasoning steps directly on the image plane. Accordingly, we develop an automatic data construction pipeline that can generate high-quality code-structured hints for existing datasets and train a unified model called Hint2Gen based on FLUX.1 Kontext to condition its generation on such hints. Furthermore, to comprehensively evaluate the effectiveness of our approach, we introduce Reason2Gen, a benchmark comprising 4,000 samples spanning 20 categories across 7 core dimensions, including path connectivity, spatial assembly, etc. Extensive experiments demonstrate that even simply providing such hints as extra inputs—without any retraining—boosts their performance. And our model significantly outperforms all leading open-source/closed-source methods on reasoning-aware generation and editing across all the dimensions.

Image editing has achieved remarkable progress recently. Modern editing models could already follow complex instructions to manipulate the original content. However, beyond completing the editing instructions, the accompanying physical effects are the key to the generation realism. For example, removing an object should also remove its shadow, reflections, and interactions with nearby objects. Unfortunately, existing models and benchmarks mainly focus on instruction completion but overlook these physical effects. So, at this moment, how far are we from physically realistic image editing? To answer this, we introduce PICABench, which systematically evaluates physical realism across eight sub-dimension (spanning optics, mechanics, and state transitions) for most of the common editing operations (add, remove, attribute change, etc). We further propose the PICAEval, a reliable evaluation protocol that uses VLM-as-a-judge with per-case, region-level human annotations and questions. Beyond benchmarking, we also explore effective solutions by learning physics from videos and construct a training dataset PICA-100K. After evaluating most of the mainstream models, we observe that physical realism remains a challenging problem with large rooms to explore. We hope that our benchmark and proposed solutions can serve as a foundation for future work moving from naive content editing toward physically consistent realism.

Recent video generation models are increasingly framed as world models. Many physical processes can unfold in more than one valid way. Therefore, a world model should reproduce not only a plausible trajectory, but also the distribution of possible behaviors under the same initial observation and action. We call this distribution-level requirement probabilistic alignment. However, existing evaluations largely assess individual-video plausibility and do not test whether repeated generations recover the correct distribution. This raises a central question: how far are current video generators from probabilistically aligned world modeling? To answer it, we formalize probabilistic alignment as a distributional criterion for world models and introduce PAWBench, a benchmark for evaluating video generators as stochastic samplers of world dynamics. We further introduce PAWEval, an outcome-level protocol that converts repeated video rollouts into empirical distributions over possible physical behaviors. Across 50 scenarios and eleven current systems, no model consistently matches the reference probabilities while recovering the range of valid behaviors. Having established this gap, we test whether language prompts, initial noise sampling, or model training can reshape the model's predictive distribution. We believe our work can serve as a foundation for future efforts to move towards probabilistically aligned world modeling.

Personalized text-to-image generation aims to synthesize images that faithfully preserve user-specified subjects under diverse textual prompts. While Low-Rank Adaptation (LoRA)-based customization has shown strong performance in single-subject personalization, multi-subject generation remains challenging because it requires simultaneously preserving subject identity, maintaining prompt consistency, and producing coherent spatial composition. Existing methods often handle these requirements in a partially entangled manner, making it difficult to balance prior preservation, subject disentanglement, and controllability within a unified generation process. To address this issue, we propose DeCoGen, a decoupled concept learning framework for multi-subject personalized image generation. Rather than learning subject-specific adaptation directly inside the main denoising pathway, DeCoGen assigns different functional roles to different branches: the main branch preserves the pretrained generative prior and global scene semantics, while each auxiliary reference branch specializes in one personalized subject. Building on this design, we introduce a mask-guided fusion mechanism that selectively injects subject-specific features into spatially relevant regions, thereby reducing cross-subject interference during composition. We further incorporate an optional image-based layout guidance strategy to improve spatial controllability in layout-sensitive scenarios. Extensive experiments on single-subject and multi-subject personalization benchmarks show that DeCoGen achieves strong performance in subject fidelity, compositional generation, and controllable subject placement. The results suggest that explicitly decoupling prior-preserving generation, subject specialization, and spatially selective fusion provides an effective and extensible framework for multi-subject personalized generation.

Interactive segmentation enables users to extract binary masks of target objects through simple interactions such as clicks, scribbles, and boxes. However, existing methods often support only limited interaction forms and struggle to capture fine details. In this paper, we revisit the classical coarse-to-fine design of FocalClick and introduce significant extensions. Inspired by its multi-stage strategy, we propose a novel pipeline, FocalClick-XL , to address these challenges simultaneously. Following the emerging trend of large-scale pretraining, we decompose interactive segmentation into meta-tasks that capture different levels of information (context, object, and detail), and assign a dedicated subnet to each level. This decomposition allows each subnet to undergo scaled pretraining with independent data and supervision, maximizing its effectiveness. To enhance flexibility, we share context- and detail-level information across different interaction forms as common knowledge while introducing a prompting layer at the object level to encode specific interaction types. As a result, FocalClick-XL achieves state-of-the-art performance on click-based benchmarks and demonstrates remarkable adaptability to diverse interaction formats, including boxes, scribbles, and coarse masks. Beyond binary mask generation, it is also capable of predicting alpha mattes with fine-grained details, making it a versatile and powerful tool for interactive segmentation.

Recent advancements adopt online reinforcement learning (RL) from LLMs to text-to-image rectified flow diffusion models for reward alignment. The use of group-level rewards successfully aligns the model with the targeted reward. However, it faces challenges including low efficiency, dependency on stochastic samplers, and reward hacking. The problem is that rectified flow models are fundamentally different from LLMs: 1) For efficiency, online image sampling takes much more time and dominates the time of training. 2) For stochasticity, rectified flow is deterministic once the initial noise is fixed. Aiming at these problems and inspired by the effects of group-level rewards from LLMs, we design Group-level Direct Reward Optimization (GDRO). GDRO is a new post-training paradigm for group-level reward alignment that combines the characteristics of rectified flow models. Through rigorous theoretical analysis, we point out that GDRO supports full offline training that saves the large time cost for image rollout sampling. Also, it is diffusion-sampler-independent, which eliminates the need for the ODE-to-SDE approximation to obtain stochasticity. We also empirically study the reward hacking trap that may mislead the evaluation, and involve this factor in the evaluation using a corrected score that not only considers the original evaluation reward but also the trend of reward hacking. Extensive experiments demonstrate that GDRO effectively and efficiently improves the reward score of the diffusion model through group-wise offline optimization across the OCR and GenEval tasks, while demonstrating strong stability and robustness in mitigating reward hacking.

Recent unified multi-modal models have made unprecedented progress in understanding and generation, yet they largely support multi-modal inputs with single-modality outputs, struggling to produce complex interleaved text–image content due to data scarcity and the difficulty of modeling long-range cross-modal context. We introduce Weaver, which frames interleaved generation as an autoregressive planning–visualization process within a unified multi-modal architecture. A planner, i.e., understanding expert, digests rich text–image context to produce visualization triggers and their dense textual guidance except for plain text, while a visualizer, i.e., generation expert, produces images conditioned on the planner’s textual guidance and visual references. This design enables decoupled learning: we train the two experts on large collections of textual planning and reference-guided image data in parallel, yielding powerful interleaved multi-modal generation capability at inference. Moreover, training the planner with datasets from diverse understanding and generation tasks equips the model with automatic task inference. To analyze and evaluate the model from multiple dimensions, we further introduce a benchmark that covers a range of everyday use cases. Extensive experiments show that, even without or with only very limited real interleaved data training, Weaver achieves superior performance on interleaved multi-modal generation.

The demand for high-resolution video generation is growing rapidly. However, the generation resolution is severely constrained by slow inference speeds. For instance, Wan2.1 requires over 50 minutes to generate a single 720p video. While previous works explore accelerating video generation from various aspects, most of them compromise the distinctive signatures (e.g., layout, semantic, motion) of the original model. In this work, we propose SURF , an efficient framework for generating high-resolution videos, while maximally keeping the signatures. Specifically, SURF divides video generation into two stages: First, we leverage the pretrained model to infer at optimal resolution and downsample latent to generate low-resolution previews in fast speed; then we design a Refiner to upscale the preview. In the preview stage, we identify that directly inferring a model (trained with higher resolution) on lower resolution causes severe losses in signatures. So we introduce noise reshifting, a training-free technique that mitigates this issue by conducting initial denoising steps on the original resolution and switching to low resolution in later steps. In the refine stage, we establish a mapping relationship between the preview and the high-resolution target, which significantly reduces the denoising steps. We further integrate shifting windows and carefully design the training paradigm to get a powerful and efficient Refiner. In this way, SURF enables generating high-resolution videos efficiently while maximally closer to the signatures of the given pretrained model. SURF is conceptually simple and could serve as a plug-in that is compatible with various base model and acceleration methods. For example, it achieves 12.5× speedup for generating 5-second, 16fps, 720p Wan 2.1 videos and 8.7× speedup for generating 5-second, 24fps, 720p HunyuanVideo.

Video world models simulate future states conditioned on current observations and user actions. Recent systems have demonstrated impressive video consistency and action controllability over long sequences. However, fairly comparing these interactive models remains challenging. In practice, a human player typically evaluates a world model by pursuing long-horizon objectives through interaction. For example, a user may turn around 360 degrees to see whether the environment remains consistent, or walk into the water and inspect whether realistic water ripples are generated. The action sequence required to achieve the same objective may vary substantially between models, making fixed action-conditioned evaluation unsuitable for cross-model comparison. To address this, we employ multi-modal Agent Players to interact with world models toward specified long-horizon objectives. Building on this paradigm, we introduce PlayWorld, a benchmark providing 171 scenarios, each with a specified objective. To evaluate performance thoroughly, we assess models along four core dimensions: geometry consistency, interaction fidelity, out-of-sight evolution, and insight evolution. In addition, we incorporate basic ability metrics for video quality and controllability. Experiments across nine state-of-the-art world models reveal that current models remain unreliable on long-horizon interactive objectives, particularly in maintaining spatial consistency and persistent state evolution. Code and data are available at https://github.com/kxding/PlayWorld.

Instruction-based video editing (IVE) is an emerging field with broad applications, yet evaluating editing models remains challenging. Existing benchmarks suffer from two major limitations: limited task coverage inherited from image editing, which overlooks video-specific dimensions, and inadequate metrics that fail to measure instruction fidelity, allowing incorrect edits to receive high scores due to strong visual priors from the original video. To address these issues, we introduce a comprehensive and structured benchmark for IVE. Our benchmark decomposes editing tasks into multiple video-specific dimensions, including spatial, temporal, audio, and reference-based editing, extending beyond conventional frame-level evaluation. It also distinguishes explicit and implicit instructions and incorporates reasoning-based scenarios to better reflect real-world requirements. Furthermore, we propose an evaluation framework that assesses editing quality from four complementary dimensions: accuracy, preservation, realism, and consistency, using both human judgments and state-of-the-art vision-language models. To emphasize instruction fidelity, we introduce an accuracy-aware penalty mechanism that conditions other scores on accuracy, preventing visually plausible but incorrect edits from receiving inflated evaluations. Extensive experiments on representative open-source and commercial models show that current IVE models remain far from satisfactory. OmniEdit-Bench provides a comprehensive and reliable testbed for evaluating instruction-based video editing and offers insights into future research directions.

Open-vocabulary Video Instance Segmentation addresses the challenging task of detecting, segmenting, and tracking objects in videos, including categories not encountered during training. However, existing approaches often overlook rich temporal cues from preceding frames, limiting their ability to leverage causal context for robust open-world generalization. To bridge this gap, we propose CPOVIS, a novel framework that introduces causal prompts—dynamically propagated visual and taxonomy prompts from historical frames—to enhance temporal reasoning and semantic consistency. Built upon a Mask2Former architecture with a CLIP backbone, CPOVIS integrates three core innovations: (1) PromptCLIP, which aligns cross-modal embeddings while preserving open-vocabulary capabilities; (2) a Visual Prompt Injector that propagates object-level features to maintain spatial-temporal coherence; and (3) a Taxonomy Prompt Infuser that leverages hierarchical semantic relationships to stabilize unseen category recognition. Furthermore, we introduce a contrastive learning strategy to disentangle object representations across frames and adapt the Segment Anything Model (SAM2) to boost open-vocabulary segmentation and tracking capacity in open-vocabulary video scenarios. Extensive experiments on seven challenging open- and closed-vocabulary video segmentation benchmarks demonstrate CPOVIS's state-of-the-art performance, outperforming existing methods by significant margins. Our findings highlight the critical role of causal prompt propagation in advancing video understanding in open-world scenarios.

Recent unified models have made unprecedented progress in both understanding and generation. However, while most of them accept multi-modal inputs, they typically produce only single-modality outputs. This challenge of producing interleaved content is mainly due to training data scarcity and the difficulty of modeling long-range cross-modal context. To address this issue, we decompose interleaved generation into textual planning and visual consistency modeling, and introduce a framework consisting of a planner and a visualizer. The planner produces dense textual descriptions for visual content, while the visualizer synthesizes images accordingly. Under this guidance, we construct large-scale textual-proxy interleaved data (where visual content is represented in text) to train the planner, and curate reference-guided image data to train the visualizer. These designs give rise to Wan-Weaver, which exhibits emergent interleaved generation ability with long-range textual coherence and visual consistency. Meanwhile, the integration of diverse understanding and generation data into planner training enables Wan-Weaver to achieve robust task reasoning and generation proficiency. To assess the model's capability in interleaved generation, we further construct a benchmark that spans a wide range of use cases across multiple dimensions. Extensive experiments demonstrate that, even without access to any real interleaved data, Wan-Weaver achieves superior performance over existing methods.

While modern visual generation models excel at creating aesthetically pleasing natural images, they struggle with producing or editing structured visuals like charts, diagrams, and mathematical figures, which demand composition planning, text rendering, and multimodal reasoning for factual fidelity. To address this, we present the first comprehensive, systematic investigation of this domain, encompassing data construction, model training, and an evaluation benchmark. First, we construct a large-scale dataset of 1.3 million high-quality structured image pairs derived from executable drawing programs and augmented with chain-of-thought reasoning annotations. Leveraging this dataset, we train a unified model that integrates a multimodal language model with FLUX.1-Kontext via a lightweight connector for enhanced multimodal understanding. A three-stage training curriculum enables progressive feature alignment, knowledge infusion, and reasoning-augmented generation, further boosted by an external reasoner at inference time. Finally, we introduce StructBench, a novel benchmark for generation and editing with over 2,000 challenging samples, and an accompanying evaluation metric, StructScore, which employs a multi-round Q&A protocol to assess fine-grained factual accuracy. Evaluations of 15 models reveal that even state-of-the-art systems score below 50\%, while our model achieves the strongest open-source performance, with consistent gains from inference-time reasoning. By releasing dataset, model, and benchmark, we aim to advance unified multimodal foundations for structured visuals.

Remote sensing (RS) image-text retrieval plays a critical role in understanding massive RS imagery. However, the dense multi-object distribution and complex backgrounds in RS imagery make it difficult to simultaneously achieve fine-grained cross-modal alignment and efficient retrieval. Existing methods either rely on complex cross-modal interactions that lead to low retrieval efficiency, or depend on large-scale vision-language model pre-training, which requires massive data and computational resources. To address these issues, we propose a fast-then-fine (FTF) two-stage retrieval framework that decomposes retrieval into a text-agnostic recall stage for efficient candidate selection and a text-guided rerank stage for fine-grained alignment. Specifically, in the recall stage, text-agnostic coarse-grained representations are employed for efficient candidate selection; in the rerank stage, a parameter-free balanced text-guided interaction block enhances fine-grained alignment without introducing additional learnable parameters. Furthermore, an inter- and intra-modal loss is designed to jointly optimize cross-modal alignment across multi-granular representations. Extensive experiments on public benchmarks demonstrate that the FTF achieves competitive retrieval accuracy while significantly improving retrieval efficiency compared with existing methods.

Weight-only post-training quantization (PTQ) is crucial for efficient Large Language Model (LLM) deployment but suffers from accuracy degradation caused by weight and activation outliers. Existing mitigation strategies often face critical limitations: they either yield insufficient outlier suppression or incur significant deployment inefficiencies, such as inference latency, heavy preprocessing, or reliance on complex operator fusion. To resolve these limitations, we leverage a key insight: over-parameterized LLMs often converge to Flat Minima, implying a vast equivalent solution space where weights can be adjusted without compromising accuracy. Building on this, we propose Astro, an Activation-guided Structured Regularization framework designed to suppress the negative effects of outliers in a hardware-friendly and efficient manner. Leveraging the activation-guided regularization objective, Astro actively reconstructs intrinsically robust weights, aggressively suppressing weight outliers corresponding to high-magnitude activations without sacrificing model accuracy. Crucially, Astro introduces zero inference latency and is orthogonal to mainstream quantization methods like GPTQ. Extensive experiments show that Astro achieves highly competitive performance; notably, on LLaMA-2-7B, it achieves better performance than complex learning-based rotation methods with almost 1/3 of the quantization time.

Recent progress in AIGC has significantly advanced text-to-3D generation, yet current models still struggle when processing long textual descriptions. Due to limitations inherited from pretraining, these models often fail to preserve fine-grained details contained in lengthy inputs, leading to incomplete or inaccurate 3D outputs. We present L2P3D, a framework designed to address this limitation by leveraging the reasoning capability of large language models to handle long-text inputs in a structured manner. The framework first analyzes the input text using an LLM and decomposes it into a set of part-specific descriptions. Each part description is then used to generate a corresponding image that captures the intended local details. A semantic-aware segmentation model is applied to extract the relevant region from each generated image, producing a collection of part-specific images and their associated masks. These images are treated as multi-condition guidance and integrated within a multi-image feature fusion pipeline to synthesize the final 3D asset. Experiments demonstrate that L2P3D effectively preserves fine-grained semantics from long textual inputs and produces more accurate and detailed 3D results compared to existing text-to-3D methods.

Human brains image complicated scenes when reading a novel. Replicating this imagination is one of the ultimate goals of AI-Generated Content (AIGC). However, current AIGC methods, such as score-based diffusion, are still deficient in terms of rapidity and efficiency. This deficiency is rooted in the difference between the brain and digital computers. Digital computers have physically separated storage and processing units, resulting in frequent data transfers during iterative calculations, incurring large time and energy overheads. This issue is further intensified by the conversion of inherently continuous and analog generation dynamics, which can be formulated by neural differential equations, into discrete and digital operations. Inspired by the brain, we propose a time-continuous and analog in-memory neural differential equation solver for score-based diffusion, employing emerging resistive memory. The integration of storage and computation within resistive memory synapses surmount the von Neumann bottleneck, benefiting the generative speed and energy efficiency. The closed-loop feedback integrator is time-continuous, analog, and compact, physically implementing an infinite-depth neural network. Moreover, the software-hardware co-design is intrinsically robust to analog noise. We experimentally validate our solution with 180 nm resistive memory in-memory computing macros. Demonstrating equivalent generative quality to the software baseline, our system achieved remarkable enhancements in generative speed for both unconditional and conditional generation tasks, by factors of 64.8 and 156.5, respectively. Moreover, it accomplished reductions in energy consumption by factors of 5.2 and 4.1. Our approach heralds a new horizon for hardware solutions in edge computing for generative AI applications.