• VLA memory shouldn't bolt on — it belongs in the native latent layer. LaMem-VLA rebuilds past experience into memory tokens and feeds them into the reasoning sequence alongside the current observation, relocating the long-horizon bottleneck to "which layer remembers" rather than "how much."

• A world model for robots should pick physics over creativity. LingBot-Video doesn't reuse an off-the-shelf video model. It rebuilds the DiT video pretraining recipe for embodiment and swaps dense for MoE — the first large-scale open-source MoE video foundation model.

• An interactive world model clears two barriers at once: unbounded duration and real-time high frame rate. LingBot-World 2.0 (Infinity) runs unbounded interaction without quality decay, and its distilled real-time variant drives a 720p stream at 60fps. A 1.3B lightweight version deploys on a single GPU.

Also Notable

• Proprioception Aligned Explicitly With Vision Tokens, Not Stitched as an Isolated Vector — GeoProp grounds 3D kinematics onto the 2D feature map, adding the missing visual correspondence to state fusion in general manipulation.
• Creative Work Doesn't Want an AI That Hands You the Final Answer — structural design needs interactive systems that help externalize ideas, explore alternatives, and refine half-formed drafts, not a one-shot generator.

Read the full edition →