OpenForgeRL introduces a framework to train AI agents directly within complex, stateful inference harnesses by decoupling training and inference through a lightweight proxy and Kubernetes-based orchestration. This approach allows developers to optimize agents end-to-end in the same environments where they are deployed, overcoming limitations in current SFT and RL stacks.
The Structured Dynamics Model (SDM) introduces a novel approach to video representation learning by explicitly separating camera motion from object dynamics. By leveraging frozen pretrained vision transformers and weak supervision, SDM provides a robust framework for understanding temporal changes in video without the need for heavy, fully supervised training.
SANA-Video 2.0 introduces a hybrid diffusion transformer architecture that balances linear attention efficiency with softmax-level quality. By utilizing gated-softmax anchors and block attention residuals, the model achieves high-resolution video generation at significantly reduced latency.
WorldWeaver (W^2) introduces cross-agent world state registers to solve the consistency challenges in multi-agent video generation. By decoupling world state modeling from visual frame generation, the architecture enables persistent, logically coherent multi-agent simulations.
ATSplat introduces a novel framework for feed-forward 3D Gaussian Splatting that restores scene-adaptive capacity allocation through sparse 3D anchor tokens. By decoupling primitive placement from input image grids, the method achieves significant reductions in Gaussian density while maintaining state-of-the-art rendering performance.
Current interpretability methods rely on reconstruction scores that are easily gamed by models using private codes. The new RECAP framework introduces decodability supervision, ensuring internal model content is independently verifiable by probes rather than relying on potentially deceptive prose.
Researchers have introduced Self Gradient Forcing (SGF), a two-pass training strategy designed to optimize how autoregressive video diffusion models encode historical context into memory. By allowing future losses to supervise the writing of past latents, SGF enables significantly more stable and consistent long-form video generation.
Researchers have introduced Isospectral Optimization (ISO), a framework that optimizes language models for reinforcement learning by isolating and modifying singular frames while keeping weight spectra fixed. This approach significantly accelerates training and enables data-free model merging, providing a more efficient path for reward-driven adaptation.
Masked Visual Actions (MVA) introduces a novel pixel-space control interface that allows video models to function as unified world models for robotics. By treating action as a partially revealed trajectory, the framework enables forward dynamics prediction and inverse modeling across diverse physical embodiments.
Researchers have introduced GAMUT, a two-level meta-rubric framework designed to evaluate the factual completeness of long-form AI generations. By moving beyond simple precision metrics, GAMUT provides a structured approach to assessing whether responses contain all necessary information, addressing a critical gap in current LLM evaluation pipelines.
A new research initiative introduces the 3D-Fit benchmark to evaluate the spatial reasoning capabilities of Large Language Models in structure-based drug design. The study compares LLM performance against established diffusion models, highlighting the potential for LLMs to handle complex, multi-constrained molecular generation tasks.
HOMIE introduces a novel framework for human-object centric video personalization, addressing the critical trade-off between subject fidelity and interaction accuracy. By leveraging MLLM integration and specialized embedding strategies, it provides a unified approach to both inter- and intra-subject video generation tasks.
Researchers have introduced Appearance Pointers, a novel mechanism for Diffusion Transformers that enables precise, region-specific control over generative image synthesis. By leveraging a modality-agnostic interface, this approach allows developers to guide image generation using text or image inputs without the need for extensive base model retraining.
FlowMimic introduces a novel framework for mask-free video editing by leveraging pixel-pair temporal warped flow fields to generate training data from image-based samples. By aligning image and video modalities through mutual imitation, the system internalizes editing capabilities, removing the need for external masks or auxiliary models.
JoyNexus introduces a service-oriented architecture for Vision-Language-Action (VLA) model post-training, moving away from exclusive resource allocation. By decoupling training, inference, and environment services, it enables efficient multi-tenancy and resource sharing for complex robotic workloads.
New research challenges the industry's reliance on peak success rates for AI security agents, proposing a cost-aware evaluation framework. The findings highlight that offensive and defensive agents exhibit fundamentally different scaling behaviors, requiring developers to prioritize operational efficiency over raw reasoning budgets.
Vision-language-action models often struggle with the discrepancy between camera-frame visual input and robot-frame action output. The introduction of robot-centric pointmaps offers a solution by encoding 3D scene data directly in a robot-relative coordinate system, enhancing generalization across diverse camera setups.
Recent research explores the application of the Muon optimizer in sparse-reward agentic reinforcement learning, demonstrating significant performance gains over traditional AdamW. By optimizing hidden weight matrices within specific policy frameworks, Muon accelerates convergence and improves success rates in complex task environments like ALFWorld.