• Semantic Action Interfaces Unlock Existing VLMs for Robot Control. Closed models work zero-shot, while small open models need only hours of GPU fine-tuning. Deterministic adapters support reuse across robot bodies.

• Reverse Distillation Escapes the Teacher's Capability Ceiling. Students explore their own distributions and amplify only verifier-backed teacher guidance. This enables model upgrades and multi-teacher integration with fewer updates.

• Linear Attention Now Estimates the Reliability of Old Memories. Kalman Delta Networks regulate writes with Kalman gain. Tests at 750M and 1.3B parameters improved perplexity and downstream accuracy.

• Program State Gives Interactive Worlds an Inspectable Logic Core. Separating rule execution from video rendering produced 94% counting accuracy and 98% state accuracy on CombatStateBench.

Also Notable

• Parametric Generation Builds the Base, Then an Agent Repairs the Layout. SceneMosaic explores how to balance speed, diversity, and usability in simulated scene generation.
• A Taxonomy Defines Where Visual Generation Becomes Agentic. It covers planning, tool use, intermediate inspection, and rework.
• Video Data Preparation Becomes an Open Five-Stage Pipeline. VidaForge makes the effects of data selection independently reproducible and comparable.
• Normal Estimation Errors May Start With VAE Compression. Geometry teams can audit representation bottlenecks and edge reconstruction before changing the diffusion backbone.
• Spatial Grids Put Multiple Shots Into One Joint Model. This post-training structure targets both within-shot motion and narrative consistency across shots.
• Draft Models Can Train Online During Long-Context RL Rollouts. The system also addresses obstacles created by context and pipeline parallelism.
• Low-Bit Attention Targets Softmax's Remaining Latency. EFQ-Softmax removes high-precision exponentiation and intermediate probability blocks from the critical path.
• Long-Term Assistants Must Use Memories, Not Just Store Them. PRAGMA tests whether systems retrieve cross-conversation evidence and apply it correctly to personalized guidance.
• Outcome Labels Lose Supervision From Multiple Successful Branches. DDO uses state-aligned branch sets and relative objectives to expand successful strategy coverage.
• Cross-Device State Transfer Enters GUI Agent Evaluation. JarvisGUI combines changing tasks to expose coordination failures that single-device benchmarks miss.

Read the full edition →