• Real-Time AR Video's Bottleneck Is Shifting. Causal Forcing++ pushes frame-wise distillation to 1-2 steps. RAVEN attacks long-rollout history distribution mismatch with consistency-model GRPO.

• SANA-WM Holds Minute-Scale World Models at 2.6B Parameters. Hybrid linear attention generates 60 seconds of 720p natively on a single H100. A distilled NVFP4 build runs 34 seconds on an RTX 5090.

• Multimodal Long-Term Memory Selection Has Data Now. MemLens compares long-context vs. memory banks across 789 multi-session questions. Neither path alone clears 30%.

• ATLAS Turns "Tool Use vs. Latent Reasoning" Into a Next-Token Decision. No architecture changes, no extra visual supervision. Standard SFT + RL handles the mode switch.

• Synthetic Data Beats Proprietary on Layered Design. The bottleneck for design tools' "AI-direct editable output" is a synthesis pipeline problem now. Returns saturate around 50K samples.

Also Notable

• PDI-Bench Adds Quantitative Geometric Consistency Evaluation for Video World Models. After length and speed got crowded, geometric fidelity is the next axis. Pairs naturally with today's three video generation papers.
• PaSaMaster: A Self-Improving Agentic Literature Retrieval System. Targets the reliability of keyword search plus LLM-level complex intent understanding. Researcher-facing; worth a scan for academic and consulting retrieval scenarios.
• Sat3DGen Generates 3D Street Scenes From a Single Satellite Image. Engineering value: pulls geometric fidelity and semantic richness into the same framework instead of as a tradeoff.
• VAE Latents Sit on a Thin Spherical Shell. Euclidean Straight-Line Flow Drifts Off. Spherical flow matching corrects it. A hidden geometric bug in latent diffusion gets called out.
• T2I Multi-Step Reasoning + Closed-Loop Verification. Together with today's layered design paper, hints at a direction: image generation is moving from single-step to multi-step pipelines with structured intermediate representations.

Read the full edition →