• Three Classes of Physical 3D Assets Merge Into One Pipeline. PhysX-Omni puts rigid, deformable, and articulated objects into a single framework. Output assets carry physics properties and drop straight into simulators. Multi-pipeline maintenance cost for sim-to-real teams should fall.

• Image Generation Is Moving From a Model Problem to an Agent Problem. GenEvolve models each generation as a trajectory and distills visual experience across tasks. It sidesteps the retune-everything-per-request pattern.

• Optimizer Is the Ignored Scaling Axis. Same FFN width increment, swap the optimizer, and the effective-capacity scaling exponent jumps from 0.44 to 1.02. Sweep optimizer as a variable before estimating scaling laws.

Also Notable

• Microsoft's Lens Matches 6B+ Models on 19.3% of the Training Compute. 3.8B T2I model. Recipe is distillation plus a redesigned pretraining flow. A training recipe small and mid-budget teams should copy.
• Mining Implicit Safety Signal From Ordinary Crowd Preference Data. ICML paper. Skips dedicated safety annotation. Uses existing preference datasets as a source of safety-related implicit objectives.
• DualOptim+ Adds a Dual-Optimizer Structure for LLM Unlearning. Base state plus delta state keeps forget and retain optimization states separate. Eases retained-capability damage from unlearning.
• Expectation Consistency Loss Brings Calibration Into Covariate Shift. Out-of-distribution deployment gets more reliable confidence. Meaningful for safety-sensitive launches.

Read the full edition →