Multimodal World Models & Agents
Key Questions
What multimodal world model projects are highlighted?
Key projects include DiLA Mamba world models, ESI-Bench, Stable Audio 3, and new additions like NEO-ov encoder-free vision-language models and Pantheon360 for 3D digital twins.
What is the MRT model and its speedup benefit?
MRT is a 20B masked region diffusion model for layered image generation that delivers 10-100x speedup compared to prior approaches.
Which benchmark shows no single model dominating video world models?
WBench, a multi-turn benchmark for interactive video world model evaluation, demonstrates that no single model leads across all tasks.
What does SpatialBench evaluate?
SpatialBench provides a comprehensive evaluation of 41 spatial foundation models, with full-context models showing strong performance across spatial reasoning tasks.
How does LocateAnything improve VLM grounding?
LocateAnything uses parallel box decoding to achieve fast and high-quality vision-language grounding trained on 138M samples.
What is the focus of EvalVerse?
EvalVerse introduces a pipeline-aware and expert-calibrated benchmark specifically designed for professional cinematic video generation evaluation.
What new benchmark addresses spatio-temporal reasoning?
VGenST-Bench evaluates spatio-temporal reasoning through active video synthesis tasks.
What roadmap is discussed for native multimodal modeling?
The highlight includes a roadmap toward native multimodal modeling along with V24M any-to-any video capabilities and On-Policy Adversarial Flow Distillation for autoregressive video generation.
AnyGoal probes whether learned dynamics generalize to goal-shifted planning by recombining familiar state and action factors, reporting weaknesses across environments and model families. It complements γ-World and other evaluation leads; PPMem’s planar scene memory and depth feedback claims improved long-video viewpoint consistency, with comparisons and code still worth inspecting.