💡 The paper proposes a new framework called Latent-to-Pixel transfer paradigm, or L2P, which allows for efficient creation of pixel-space models using pre-trained latent diffusion models. The problem addressed is that training advanced pixel-space models from scratch requires significant computational and data resources. To solve this, L2P harnesses the knowledge of pre-trained latent diffusion models to build powerful pixel-space models with minimal training overhead.

The method involves discarding the Variational Autoencoder in favor of large-patch tokenization and freezing the intermediate layers of the pre-trained latent diffusion model. Only the shallow layers are trained to learn the latent-to-pixel transformation, using synthetic images generated by the pre-trained model as the training data. This approach enables rapid convergence without the need for real data collection.

The results show that L2P achieves negligible training overhead while performing on par with the source latent diffusion model. The framework is able to migrate massive latent priors to the pixel space using only 8 GPUs, and it unlocks native 4K ultra-high resolution generation by eliminating the Variational Autoencoder memory bottleneck. Extensive experiments demonstrate that L2P reaches 93 percent performance on GenEval and performs well on DPG-Bench, making it a promising approach for efficient pixel-space model creation.

💡 The paper introduces a new approach to language models called Multi-Stream LLMs, which aims to overcome the limitations of traditional language models that process information in a single stream of computation. The current models function on a message exchange format, where they successively exchange messages with users, systems, and tools, leading to limitations such as the inability to act while reading, react to new information while writing, think while reading or acting, and act while thinking.

To address these limitations, the authors propose a method that involves transitioning from sequential message-based instruction-tuning to parallel stream processing, enabling simultaneous reading and generation across multiple concurrent data flows. This is achieved by splitting each role into a separate stream, allowing the language model to simultaneously read from multiple input streams and generate tokens in multiple output streams, all of which causally depend on earlier timesteps.

The results of this approach show that it can improve model efficiency through parallelization, enhance model security through better separation of concerns, and increase model monitorability. The authors argue that this data-driven change can remedy the usability limitations of traditional language models, making them more efficient, secure, and transparent. Overall, the paper contributes to the development of more advanced language models that can process information in a more parallel and efficient manner, unlocking their potential for widespread use in various applications.

💡 The paper addresses the challenges of multimodal deep search, which requires an agent to solve open-world problems by combining search, tool use, and visual reasoning. Current systems have two major limitations: they treat images as transient outputs and cannot reuse intermediate visual evidence, and they rely on fixed curation recipes for training data that do not adapt to the agent's evolving capabilities.

To overcome these limitations, the authors introduce a visual-native agent harness with an image bank reference protocol, which allows images to be registered as addressable references and reused by later tools. They also propose On-policy Data Evolution, a closed-loop data generator that refines itself across rounds based on the policy being trained. This approach enables the generation of targeted training data that adapts to the agent's current needs.

The authors evaluate their approach on eight multimodal deep search benchmarks and demonstrate significant improvements in performance. With the On-policy Data Evolution method, the Qwen3-VL-8B agent achieves an average score of 39.0%, surpassing the Gemini-2.5 Pro agent. Further analysis shows that the image bank reuse is particularly effective for complex tasks that require iterative visual refinement, and that the rollout-feedback evolution yields more grounded and policy-matched reinforcement learning tasks.

The contributions of the paper are twofold: first, the introduction of a visual-native agent harness that enables reusable intermediate visual evidence, and second, the development of On-policy Data Evolution, a method for generating targeted training data that adapts to the agent's evolving capabilities. The results demonstrate the effectiveness of these contributions in improving multimodal deep search performance.

💡 The paper introduces AnyFlow, a novel framework for any-step video diffusion distillation that improves upon existing consistency distillation methods. The problem with consistency distillation is that its performance degrades as more sampling steps are used at test time, limiting its effectiveness for any-step video diffusion. This is because consistency distillation replaces the original probability-flow ODE trajectory with a consistency-sampling trajectory, which weakens the desirable test-time scaling behavior of ODE sampling.

To address this limitation, AnyFlow optimizes the full ODE sampling trajectory instead of distilling a model for only a few fixed sampling steps. The method involves shifting the distillation target from endpoint consistency mapping to flow-map transition learning over arbitrary time intervals. Additionally, the authors propose Flow Map Backward Simulation, which decomposes a full Euler rollout into shortcut flow-map transitions, enabling efficient on-policy distillation that reduces test-time errors.

The results of the paper show that AnyFlow achieves performance that matches or surpasses consistency-based counterparts in the few-step regime, while also scaling with sampling step budgets. The experiments were conducted across both bidirectional and causal architectures, at scales ranging from 1.3B to 14B parameters. Overall, the paper contributes a new framework for any-step video diffusion distillation that improves upon existing methods and achieves state-of-the-art results.

💡 The paper introduces TrackCraft3R, a method for dense 3D tracking from monocular video that adapts video diffusion transformers to follow physical points across frames. The problem of dense 3D tracking is fundamental to dynamic scene understanding, but existing methods either lack real-world motion priors or are inefficient. Pre-trained video diffusion transformers offer rich spatio-temporal priors, but their frame-anchored formulation is mismatched with reference-anchored dense 3D tracking.

To address this, TrackCraft3R uses a dual-latent representation that combines per-frame geometry latents and reference-anchored track latents as dense queries. It also employs temporal RoPE alignment, which specifies the target timestamp of each track latent. These designs convert the per-frame generative paradigm of video diffusion transformers into a reference-anchored tracking formulation.

The method achieves state-of-the-art performance on standard sparse and dense 3D tracking benchmarks, while being 1.3 times faster and using 4.6 times less peak memory than the strongest prior method. Additionally, TrackCraft3R demonstrates robustness to large motions and long videos. The key contribution of the paper is the repurposing of video diffusion transformers as a feed-forward dense 3D tracker, enabling efficient and accurate tracking of physical points across frames in a single forward pass.

💡 The Qwen Image VAE 2.0 technical report presents a high compression Variational Autoencoder suite that improves reconstruction fidelity and diffusability. The problem addressed in this paper is the reconstruction bottleneck of high compression in Variational Autoencoders. To solve this problem, the authors propose an improved architecture featuring Global Skip Connections and expanded latent channels. They also scale training to billions of images and incorporate a synthetic rendering engine to improve performance in text rich scenarios.

The method used in this paper involves implementing an enhanced semantic alignment strategy to make the latent space highly amenable to diffusion modeling. The authors also leverage an asymmetric and attention free encoder decoder backbone to minimize encoding overhead. The performance of Qwen Image VAE 2.0 is evaluated on public reconstruction benchmarks and a new benchmark called OmniDoc TokenBench, which is a collection of real world documents with specialized OCR based evaluation metrics.

The results show that Qwen Image VAE 2.0 achieves state of the art reconstruction performance, demonstrating exceptional capabilities in both general domains and text rich scenarios at high compression ratio. Downstream DiT experiments reveal that the models possess superior diffusability, significantly accelerating convergence compared to existing high compression baselines. Overall, Qwen Image VAE 2.0 establishes itself as a leading model with high compression, superior reconstruction, and exceptional diffusability.

💡 The paper introduces Asymmetric Flow Modeling, a method for efficient high-dimensional flow-based generation. The problem with existing flow-based generation methods is that they require modeling high-dimensional noise, which is difficult even when the data has a strong low-rank structure. To address this, the authors propose a rank-asymmetric velocity parameterization that restricts noise prediction to a low-rank subspace while keeping data prediction full-dimensional. This approach allows for the analytical recovery of the full-dimensional velocity without changing the network architecture or training procedures.

The method, called AsymFlow, enables effective fine-tuning from latent models to pixel-space models by aligning the low-rank pixel subspace to the latent space. This provides a seamless initialization that preserves the latent model's high-level semantics and structure, allowing fine-tuning to mainly improve low-level mismatches rather than relearning pixel generation.

The results show that AsymFlow achieves a leading performance on ImageNet 256x256, outperforming prior pixel diffusion models by a large margin. Additionally, the authors demonstrate that AsymFlow provides a route for fine-tuning pretrained latent flow models into pixel-space models, establishing a new state of the art for pixel-space text-to-image generation. The pixel AsymFlow model fine-tuned from a latent base model achieves better performance on several benchmarks, including HPSv3, DPG-Bench, and GenEval, and shows substantially improved visual realism. Overall, the paper presents a significant contribution to the field of flow-based generation, enabling efficient and effective high-dimensional generation and fine-tuning of latent models.

💡 The paper introduces Self Distilled Agentic Reinforcement Learning, a method that improves reinforcement learning for multi turn agent training. The problem with traditional reinforcement learning is that it provides only coarse supervision for long horizon interaction, which can lead to instability in multi turn agents. On Policy Self Distillation is a technique that complements reinforcement learning by providing dense token level guidance from a teacher branch, but it has limitations when applied to multi turn agents, such as compounding instability and negative teacher rejections.

The proposed method, Self Distilled Agentic Reinforcement Learning, addresses these limitations by treating On Policy Self Distillation as a gated auxiliary objective, while keeping reinforcement learning as the primary optimization backbone. It uses a sigmoid gate to selectively strengthen positive token level guidance and mitigate negative teacher rejections. This allows the method to stabilize supervision and improve the performance of multi turn agents.

The results show that Self Distilled Agentic Reinforcement Learning substantially improves over existing methods, such as GRPO, and avoids the instability of naive combinations of GRPO and On Policy Self Distillation. The method consistently outperforms hybrid reinforcement learning and On Policy Self Distillation baselines across different model scales and datasets, including Qwen2.5 and Qwen3 families on ALFWorld, WebShop, and Search-QA. Overall, the paper contributes a new method that improves the performance and stability of multi turn agents in reinforcement learning.

💡 The paper proposes a novel approach called Warp-as-History for camera-controlled video generation. Existing methods for this task typically require large-scale camera-annotated videos for post-training or rely on test-time optimization, which can be time-consuming and costly. The proposed method addresses this problem by transforming camera-induced warps into pseudo-history representations, which enables a frozen video generation model to follow camera trajectories without any training or test-time optimization.

The Warp-as-History method works by constructing camera-warped pseudo-history from past observations and feeding it through the model's visual-history pathway. The positional encoding is aligned with the target frames being denoised, and warped-history tokens without valid source observations are removed. This simple interface reveals a non-trivial zero-shot capability of the model to follow camera trajectories.

The results show that the proposed method can achieve good camera adherence, visual quality, and motion dynamics without requiring large-scale camera-annotated videos or test-time optimization. Furthermore, lightweight offline finetuning on only one camera-annotated video can further improve the model's capability and generalize to unseen videos. Extensive experiments on diverse datasets confirm the effectiveness of the Warp-as-History method, making it a promising approach for camera-controlled video generation.

Overall, the paper's contributions include a novel method for camera-controlled video generation that requires minimal training data and no test-time optimization, and demonstrates the potential for zero-shot capability in video generation models. The proposed approach has the potential to simplify the process of camera-controlled video generation and make it more accessible to a wider range of applications.

💡 The paper presents a systematic approach to transform post-trained reasoning models into rigorous olympiad-level solvers. The problem addressed is achieving gold-medal-level performance on mathematical and physics competitions. The method involves a simple and unified recipe that includes three main components: a reverse-perplexity curriculum, a two-stage reinforcement learning pipeline, and test-time scaling. The reverse-perplexity curriculum is used to instill rigorous proof-search and self-checking behaviors in the model. The two-stage reinforcement learning pipeline progresses from reinforcement learning with verifiable rewards to more delicate proof-level reinforcement learning, allowing the model to scale its behaviors. Finally, test-time scaling is used to boost the solving performance of the model.

The authors applied this recipe to a 30B-A3B backbone with sequence-to-function transformer training on around 340K sub-8K-token trajectories, followed by 200 reinforcement learning steps. The resulting model, SU-01, demonstrates stable reasoning on difficult problems with trajectories exceeding 100K tokens. The results show that the model achieves gold-medal-level performance on mathematical and physical olympiad competitions, including the International Mathematical Olympiad and the International Physics Olympiad. Additionally, the model demonstrates strong generalization of scientific reasoning to domains beyond mathematics and physics. Overall, the paper contributes a simple and unified approach to achieving gold-medal-level olympiad reasoning, with significant implications for advancing long-horizon mathematical and scientific problem solving.

💡 The paper introduces RAVEN, a real-time autoregressive video extrapolation network, and CM-GRPO, a consistency model-based reinforcement learning approach. The problem addressed is the gap between the history distributions encountered during training and those arising at inference in causal autoregressive video diffusion models, which constrains generation quality over long horizons.

To solve this problem, RAVEN repacks each self rollout into an interleaved sequence of clean historical endpoints and noisy denoising states, aligning training attention with inference-time extrapolation. This formulation allows downstream chunk losses to supervise the history representations on which future predictions depend.

Additionally, CM-GRPO reformulates a consistency sampling step as a conditional Gaussian transition and applies online reinforcement learning directly to this kernel, avoiding the Euler-Maruyama auxiliary process adopted in prior flow-model RL formulations.

The results demonstrate that RAVEN surpasses recent causal video distillation baselines across quality, semantic, and dynamic degree evaluations. Furthermore, CM-GRPO provides further gains when combined with RAVEN, indicating the effectiveness of the proposed methods in improving real-time video generation.

Overall, the paper presents a novel approach to real-time video generation through causal autoregressive extrapolation with improved training alignment and consistency model-based reinforcement learning, achieving state-of-the-art results in video generation quality and performance.