💡 The paper proposes a training-free acceleration strategy for text-to-image diffusion models called MrFlow. The problem with existing multi-resolution generation strategies is that they can produce noticeable blurring or artifacts due to upsampling in the latent space and selective modification of partial regions. MrFlow addresses this issue by using a staged low-to-high-resolution pipeline. It first generates the main structure at low resolution, then performs super-resolution in the pixel space using a lightweight pretrained model, injects low-strength noise to enable high-frequency resampling, and finally refines the details at high resolution. The results show that MrFlow achieves a 10x end-to-end acceleration while maintaining a high level of image quality, with only a 1 percent gap in performance compared to the original model. Additionally, MrFlow can be combined with other acceleration strategies, such as timestep distillation, to achieve even higher acceleration of up to 25x. The key advantage of MrFlow is that it does not require any training or runtime modifications, making it a hardware-agnostic and efficient solution for accelerating text-to-image diffusion models.

💡 MetaSpatial is a framework that uses reinforcement learning to improve 3D spatial reasoning in vision-language models, which are used to generate 3D scenes. The problem with current models is that they lack internalized 3D spatial reasoning, which limits their ability to generate realistic layouts. Additionally, traditional supervised fine-tuning methods are not effective for layout generation tasks because perfect ground truth annotations are not available.

To address these challenges, MetaSpatial introduces a multi-turn reinforcement learning-based optimization mechanism that integrates physics-aware constraints and rendered image evaluations. This mechanism allows the model to refine spatial arrangements over multiple turns by analyzing rendered outputs, improving scene coherence progressively.

The method works by having the model analyze rendered outputs and refine the spatial arrangements in an iterative process. This process ensures that the generated 3D layouts are coherent, physically plausible, and aesthetically consistent.

The results of the empirical evaluations demonstrate that MetaSpatial significantly enhances the spatial consistency and formatting stability of various scale models. After training, object placements are more realistic, aligned, and functionally coherent, which validates the effectiveness of reinforcement learning for 3D spatial reasoning in applications such as metaverse, AR/VR, digital twins, and game development.

Overall, the contributions of MetaSpatial are the introduction of a reinforcement learning-based framework that enhances 3D spatial reasoning in vision-language models, and the demonstration of its effectiveness in generating realistic and coherent 3D scenes. The code, data, and training pipeline are publicly available, which can facilitate further research and development in this area.

💡 The paper introduces Embodied.cpp, a portable C++ runtime that enables efficient deployment of embodied AI models across heterogeneous edge devices. The problem addressed is the fragmentation of embodied AI model deployment, which is currently limited by model-specific Python stacks, backend assumptions, and robot-side glue code. This makes it difficult to deploy these models on various devices, especially on heterogeneous edge devices.

The authors propose Embodied.cpp as a solution, which is based on an architectural analysis of representative vision-language-action and world-action models. The runtime is organized into five layers: input adapters, sequence builders, backbone execution, head plugins, and deployment adapters. This modular design provides multi-rate execution, latency-first fused inference, and extensible operator and I/O support, allowing for deployment across diverse devices, robots, and simulators through a single backend abstraction.

The results show that Embodied.cpp achieves successful closed-loop execution with high task success rates on two vision-language-action models, and reduces block memory usage on a preliminary world-action model benchmark. Specifically, the VLA deployments achieve 100.0% and 91.0% task success rates, while the WAM benchmark reduces block memory from 312.2 MiB to 88.1 MiB. These results demonstrate that Embodied.cpp improves deployment efficiency while preserving high accuracy across diverse embodied model architectures. Overall, the paper contributes a portable and efficient runtime for embodied AI models, enabling their deployment on a wide range of devices and robots.

💡 The paper introduces VLA-Corrector, a lightweight framework that improves the robustness of vision-language-action models in contact-rich manipulation tasks. The problem addressed is the limitation of action chunking, where a fixed action horizon can lead to compounding errors due to small local perturbations. To solve this, VLA-Corrector proposes a detect-and-correct mechanism that continuously monitors the visual feature evolution and detects deviations from the predicted trajectory. When a deviation is detected, the system triggers a corrective replanning event, discarding the remaining stale actions and invoking online gradient guidance to replan the actions. This approach induces an adaptive action horizon, preserving long-horizon execution when the current chunk is reliable and invoking short-horizon corrective replanning when execution begins to drift. The method can be integrated into different vision-language-action models without retraining the backbone, and it mitigates the trade-off between execution robustness and policy-call frequency. The results show that VLA-Corrector substantially improves robustness in long-horizon, contact-rich robotic manipulation tasks while preserving much of the efficiency benefit of action chunking.

💡 The paper addresses the problem of dense spatial perception in visual systems, which is crucial for physical intelligence and embodied artificial intelligence applications. Current visual foundation models prioritize semantic invariance over detailed spatial understanding, resulting in limited ability to recover structured and metric representations from pixel observations. To overcome this, the authors propose a new approach to vision pretraining that focuses on boundary modeling, which is motivated by the idea that boundaries and shape discontinuities provide essential cues for perceiving geometric properties.

The proposed method, called masked boundary modeling, is a self-supervised paradigm that dynamically learns sub-pixel boundary representations. This is achieved by discovering boundary-bearing tokens and using them as masked targets to facilitate dense visual token learning. The authors scale this framework and develop a model called LingBot-Vision, which is evaluated on a diverse set of downstream vision tasks.

The results show that LingBot-Vision outperforms a strong baseline, DINOv3, and drives significant improvements in depth completion and estimation. Specifically, it enables the progression from LingBot-Depth 1.0 to LingBot-Depth 2.0, resulting in enhanced depth estimation, a key component of embodied artificial intelligence. The findings demonstrate that boundary modeling is a scalable pretraining principle for learning spatially structured visual representations, going beyond simple line segments and offering a new approach to vision pretraining. Overall, the paper contributes a novel approach to vision pretraining that prioritizes dense spatial perception and boundary modeling, with significant implications for embodied artificial intelligence applications.

💡 The paper presents ResearchStudio-Idea, a skill suite designed to support effective research ideation by combining literature search, novelty checking, and pattern-guided generation. The goal is to help researchers develop well-grounded research proposals by identifying meaningful bottlenecks, differentiating from existing solutions, and evaluating risks. The suite consists of three main components: Paper-Search, a multi-source literature search skill, Scoop-Check, a prior-art collision checker, and IdeaSpark, an end-to-end skill that composes evidence grounding, pattern-guided generation, and idea-card rendering into one workflow.

IdeaSpark is constructed from a corpus of 1947 machine learning conference papers collected from ICLR, ICML, and NeurIPS between 2021 and 2025. Analysis of these papers reveals 31 recurring ideation sub-patterns, which are consolidated into 15 reusable ideation patterns. Each pattern is operationalized as a structured card containing research context, bottleneck types, differentiation strategies, supporting precedents, and common failure modes.

Given a research problem and an evidence bundle, IdeaSpark evaluates evidence readiness, reconstructs the surrounding research context, identifies unresolved bottlenecks, selects relevant patterns, instantiates one candidate direction, retrieves potentially conflicting prior work, and performs outcome-informed auditing. This workflow transforms reusable ideation patterns into traceable research proposals.

The results show that IdeaSpark consistently produces stronger research proposals than no-skill and generic-skill baselines while maintaining competitive novelty. The evaluation is based on blind automated-judge assessments, which demonstrate the effectiveness of ResearchStudio-Idea in supporting research ideation. Overall, the paper contributes a reusable skill suite that can help researchers develop well-grounded research proposals by leveraging evidence-grounded research ideation and pattern-guided generation.

💡 The paper addresses the challenge of evaluating robotic policies, which is a critical bottleneck in the development of embodied robot foundation models. Unlike large language models, robotic policies require real-world rollouts that are slow and costly, limited by hardware and human supervision. To overcome this, the authors propose the use of world models as surrogate policy evaluators. However, the key properties that make a world model reliable for policy assessment are not well understood.

To study this, the authors introduce WMBench, a benchmark constructed from real-robot teleoperation data and matched policy rollouts covering diverse manipulation tasks. They use WMBench to analyze seven video world models, four action representation schemes, and over 324,000 simulated policy rollouts paired with real robot executions. The analysis is further enriched with large-scale community submissions, curated synthetic trajectories, and a large dataset of training videos.

The experiments reveal three core insights: first, evaluator quality is dominated by long-horizon, action-faithful rollout consistency rather than short-term visual realism; second, pretraining gains come not only from data scale but also from balancing general world knowledge with robot-specific controllability; and third, architectural choices such as action encoding, memory design, and evaluator-focused post-training strongly determine alignment with real-world robot behavior.

Based on these results, the authors derive a practical design roadmap and realize it in GigaWorld-1, a world model specially optimized for policy evaluation. The authors fully release their code, models, datasets, and toolkits to advance scalable evaluation research for embodied foundation models. The contributions of the paper include a systematic study of world models for robotic policy evaluation, the introduction of WMBench as a benchmark, and the development of GigaWorld-1 as a reliable world model for policy assessment. Overall, the paper provides a comprehensive framework for evaluating robotic policies and advancing the development of embodied robot foundation models.

💡 The paper introduces a multiplayer world model that can simulate complex physics-based environments with multiple agents. The model is trained on a large dataset of gameplay from the game Rocket League, which involves fast and tightly coupled dynamics between players. The key contribution of the paper is that the model conditions on the action streams of multiple agents, allowing it to attribute changes in the scene to the correct player and stay coherent under different combinations of actions.

The model uses a latent diffusion approach with a generative objective and is trained on short clips of gameplay. Despite this, the model is able to generate stable long-horizon rollouts, with distributional quality holding steady for up to five minutes and in practice continuing for hours without collapsing. The model is also able to produce four-player matches in real time, generating 20 frames per second on a single GPU.

The paper systematically investigates the design choices behind the model, including the video codec, generative objective, and multiplayer conditioning scheme. It also characterizes how the model's behavior changes with scale, including the capabilities that emerge and the failure modes that persist. The authors develop targeted evaluations that probe the model's physical understanding, rather than just its visual appearance.

The paper's contributions include the introduction of the first multiplayer world model for highly dynamic environments, the development of a latent diffusion model that can generate stable long-horizon rollouts, and the creation of a large-scale dataset and codebase for continued research on multiplayer world models. The authors release their dataset, codebase, and a live demo to support further research in this area. Overall, the paper presents a significant advance in the development of world models that can simulate complex multiplayer environments.

💡 The paper introduces PixWorld, a unified approach for 3D scene generation and reconstruction in pixel space. The problem with current methods is that they use separate paradigms for reconstruction and generation, with reconstruction using pixel-based regression and generation using latent diffusion. Recent attempts to unify these tasks in latent space have limitations, including information loss and the need for a pretrained autoencoder.

The method presented in the paper reformulates these tasks under a unified pixel-space diffusion paradigm. PixWorld is a single model that jointly addresses 3D reconstruction and generation by supervising diffusion directly on rendered images. This approach removes the limitations of latent space methods and aligns optimization with 3D scene fidelity. The model also introduces a geometry perception loss that aligns rendered views with their ground truth in the geometry-aware feature space of a pretrained 3D foundation model, providing 3D structural supervision.

The results show that PixWorld consistently outperforms prior latent-space generation methods and matches state-of-the-art reconstruction methods. This demonstrates the superiority of a unified pixel-space approach for 3D scene generation and reconstruction. Overall, the paper presents a novel approach that overcomes the limitations of current methods and achieves superior results, making it a significant contribution to the field of 3D scene generation and reconstruction.