Ask or search anything...

Researchers from Stanford University and collaborators developed Masked Visual Actions (MVA), a pixel-space control interface that enables a single pre-trained video model to function as a unified world model for both forward and inverse dynamics in robot manipulation. This approach achieved superior visual fidelity and cross-embodiment generalization in controllable video generation, improved planning success rates by up to 26% on Robocasa tasks, and demonstrated zero-shot inverse modeling with 90% success on a complex pick-and-place task.

Harnesses can lead to compositional generalization: we observe a property in training RLMs, in which similarly structured tasks are viewed as isomorphic and all individual LM calls in the harness become in-distribution.

Microsoft Mage Team developed Mage-Flow, a 4B-parameter foundation model that achieves competitive high-resolution image generation and instruction-based editing performance against much larger open-source models, demonstrating significant efficiency with 0.59s for 1024x1024 image generation on an NVIDIA A100 GPU and a 2.48x training speedup through system co-design.

Experiential Learning (EL), developed by Microsoft Research and collaborators, introduces an "LLM-as-a-Coach" framework that transforms rich textual feedback into high-bandwidth learning signals for post-training language models on non-verifiable, open-ended tasks. This method consistently outperformed traditional reinforcement learning, demonstrating improved generalization and reduced reward hacking on various open-ended benchmarks.

The ABot-World Team at Alibaba Group's AMAP CV Lab developed ABot-World-0, an action-conditioned video world model capable of generating interactive, persistent virtual worlds. It achieves real-time rollout at up to 16 frames per second (FPS) at 720P resolution with 1.2 seconds of action-to-first-frame latency, running on a single desktop GPU with less than 19.3 GiB VRAM.

Researchers at Tufts University and Harvard University propose 'learnable novelty' as a unifying principle for intelligence, and introduce a differentiable estimator for it. This objective successfully drives complex dynamics in cellular automata, enables unsupervised representation learning for MNIST with 0.89 accuracy, and provides a robust intrinsic reward for reinforcement learning agents across various tasks.

A framework called Agentic Real2Sim automatically converts real-world robotic interaction recordings into physically accurate, simulatable digital twins, supporting rigid, deformable, and humanoid domains. It achieved 48% replay success on 100 DROID rigid manipulation episodes using an open-weight VLM backend, demonstrating cost-effectiveness compared to proprietary models.

Poolside introduces Laguna S 2.1, a 118B total-parameter Mixture-of-Experts model with 8B activated parameters per token and a 1M-token context window, built for long-horizon agentic coding. It went from the start of training to launch in under nine weeks and, despite its small active size, holds its own on agentic coding benchmarks against models many times larger. Poolside attributes the gains largely to post-training that rewards persistence, verification, and backtracking rather than raw scale, and releases full evaluation trajectories for every benchmarked trial.

WorldScape Policy 2.0, developed by Manifold AI, Tsinghua University, and Shanghai Jiao Tong University, advances World Action Models by incorporating reasoning-augmented long short-term memory and diverse multimodal control. This system achieved 94.3% average success across 50 bimanual simulation tasks and demonstrated improved real-world performance on long-horizon and memory-dependent robotic manipulation.

OmniReasoner, a post-training framework, enables omni-modal Large Language Models to reason effectively over long audio-video content by adaptively acquiring high-fidelity evidence via a "zoom-in" tool. The framework demonstrates improved accuracy on various benchmarks, including a 5.5-point gain on OmniVideoBench and a 15.6-point gain on VideoHolmes, by concentrating high-fidelity computation on relevant temporal regions.

FlowMimic introduces a framework for generating video editing training data in real-time from image editing samples, using a pixel-pair warped flow field to overcome limitations in traditional data collection. The method aligns image and video modalities and internalizes language-visual understanding for mask-free editing, showcasing broad generalization across various video editing and generation tasks.

Isospectral Optimization (ISO) provides an RLVR-native framework that addresses the distinct dynamics of Reinforcement Learning with Verifiable Rewards by preserving the base model's weight spectra while adapting new behaviors through singular frame modifications. This approach improves training efficiency, achieving higher accuracy or faster convergence compared to conventional optimizers, and enables effective data-free merging of RLVR specialist models.

Latent Riemannian Flow Matching enables generative 3D scene modeling by operating directly within the non-Euclidean latent space of a frozen geometric foundation model (VGGT). The framework significantly improves depth prediction and 3D scan generation from sparse, unposed inputs, reducing Chamfer distance by over 40% on ScanNet++ compared to baselines.

Researchers from Hunan University and Nanyang Technological University identified the "Blur Trap" in 3D Gaussian Splatting, a persistent blur caused by biased gradient optimization, and introduced two minimal exploration operators—Random Seeding and Random Splitting—that significantly enhance scene detail and fidelity across diverse datasets.

RynnBrain 1.1, an upgraded family of embodied foundation models from DAMO Academy, Alibaba Group, and Hupan Lab, enhances physical world understanding and generalizability for robot control. It achieves precise 3D grounding and contact point prediction capabilities, leading to significantly improved real-robot manipulation performance across diverse platforms compared to previous models.

The researchers developed PATCH POLICY, an efficient, modular Transformer-based policy that directly utilizes dense visual representations from frozen, pre-trained Vision Transformers. This approach achieves superior performance in precise robot manipulation tasks with significantly reduced computational cost and faster inference compared to large Vision-Language-Action models.

The ADAFLASH framework introduces an adaptive speculative decoding approach to mitigate high variance in diffusion-based drafters for large language models. It employs on-policy distillation for continuous drafter adaptation and an adaptive length head for dynamic verification, achieving up to 4.06x speedup and improved throughput, particularly under high concurrency, compared to previous diffusion-based methods.

SWE-Pruner Pro enables large language model (LLM) coding agents to perform context pruning by leveraging the LLM's internal representations for per-line keep-or-prune decisions on tool outputs. This approach significantly reduces token consumption, with up to 39% savings, and often enhances task performance, such as increasing the SWE-Bench Verified resolve rate by 3.8% with MiMo-V2-Flash, while incurring minimal additional overhead.

The Darwin Agent Team developed Mi-Memory, a lifecycle memory framework for personal AI that establishes an end-to-end audit contract, managing durable user state, integrating multimodal evidence, governing policy evolution, and ensuring deployability across varied constraints. The framework achieved a +19.16 percentage point gain on LoCoMo through automated strategy evolution and retained 90.0% of its service-side benefits in a lightweight, file-system-backed deployment.