Must-read papers of the week: VFlash, GR00T 2, agent guardrails, AI-scientist benchmarks
TheTuringPost · x · 2026-11-10
The Turing Post curates this week's must-read papers spanning embodied AI (NVIDIA's VFlash visual planning with 25x faster 3D trajectory decoding, GR00T 2, La-JEPA, Open-HumanWAM, NEO-VLA, RoboColosseum), agents (ToolSpace, LEANGUARD Lean-verified guardrails, agentic manipulation policies, Route Anything), AI scientists (Scientist World, FrontierPhysics), and multimodal (DreamX). The highlighted VFlash paper (NVIDIA, arXiv 2611.08449) exploits the known structure of visual primitives for parallel decoding, cutting latency while improving accuracy on Jetson Thor, DGX Spark, and H100; a world action model conditioned on VFlash plans hits 70.2% average success vs 62.6% baseline.
More from Research
- Ant Group's SkinMind: agentic RAG with graph probing for dermatology AI — antgroup · 2026-11-10
- New training losses make ViTs develop specialized spatial attention heads — Yuji Nozawa · 2026-11-10
- DeepMind's AlphaProof Code Formally Verifies Signal Protocol in a Month — sebkrier · 2026-11-10
- Shaun Gallagher dissects how "robodiment" differs from human embodiment — AnnaCiaunica · 2026-11-10
- Sillage 3D: open-source browser engine renders Gaussian splats as brush strokes — janusch_patas · 2026-11-10
- Alexander Mattick's masterclass unifies energy-based models, diffusion, flows and flow matching — burny_tech · 2026-11-10