Mechanistic Shaping raises worst-group accuracy on a biased image-classification benchmark from 51.6% to 97.7%. It also cuts unwanted behaviors in language models up to 32B parameters, though models can sometimes hide concepts internally.
Key points
Worst-group accuracy improves from 51.6% to 97.7% on the benchmark.
Unwanted behaviors in LLMs up to 32B parameters are reduced.
Sources
WorldFoundry launch – 270 models, 74 benchmarks, 62 metrics
WorldFoundry is an open-source infrastructure for reproducible inference and evaluation of world models across video, 3D/4D, embodied AI, and interactive worlds. The catalog lists 270 model entries, 74 benchmarks, and 62 registered metric entries.
Key points
Model entries: 270
Benchmarks: 74
Sources
CMDiff Left Atrial Scar Segmentation Feasibility Study
CMDiff is a conditional mask diffusion model for 3-D LGE MRI left atrial scar segmentation, achieving DSC 0.47 ± 0.08. The study evaluated CMDiff on 60 public LASCARQS subjects with five-fold cross-validation and released the source code publicly.
Key points
DSC 0.47 ± 0.08, HD95 12.89 ± 6.39 mm, ASD 1.91 ± 0.96 mm for CMDiff.
Deterministic baseline DSC 0.51 ± 0.08, HD95 10.86 ± 5.39 mm, ASD 1.58 ± 1.03 mm.
Sources
GR00T 2 released by NVIDIA, predicts video and actions
GR00T 2 is a new robot foundation model from NVIDIA that predicts future video and actions together. It learns what the world will look like, not just the next action.
Key points
GR00T 2 released by NVIDIA
Predicts future video and actions together
Sources
PolicyIK replaces numerical IK with single forward pass
PolicyIK replaces numerical inverse kinematics with a goal-conditioned RL policy that maps end-effector error to joint velocities in one forward pass. The method bypasses singularities that cause lockups and solves dynamic obstacle avoidance in the same pass.
Key points
Goal-conditioned RL policy converts end-effector pose errors into joint velocities.
Single forward pass bypasses singularities that cause lockups and solves dynamic obstacle avoidance.
Sources
Virtual Humans Toolbox – surface demos released
Virtual Humans Toolbox adds browser demos for implicit and explicit surfaces, CSG, level sets, subdivision and Procrustes. The demos run in a web browser and are part of the Virtual Humans lecture series.
Key points
browser/03_1_surface_representations.html demo includes implicit and explicit surfaces, CSG, level sets, meshes, Bézier curves and patches, subdivision.
browser/03_2_procrustes.html demo includes similarity transforms, the SVD, and Procrustes alignment.
Sources
- Original post
- Virtual Humans Toolbox - Linked page
- virtualhumans.mpi-inf.mpg.de - Linked in the post
- youtube.com - Linked in the post
Inkling 2 Small free on opencode for limited time
Inkling 2 Small is free on opencode for a limited time. It was trained for agentic coding and knowledge work.
Key points
Inkling 2 Small is available on @opencode, free for a limited time.
Trained for agentic coding and knowledge work.