• Onboarding A New Object Drops From "Prep A CAD Model" To "Snap A Few Photos." PANY swaps single-anchor matching for a multi-view geometric backbone, lifting pose accuracy by 12% on YCB-V and over 20% on LM-O. For teams doing embodied grasping, that cost curve matters more than any single accuracy number.

• Can't Touch The Foundation Model? Bolt A Patch Behind It. PEPA freezes the encoder and adds a 0.26M-parameter plug-in that fixes lost fine structure and bad thresholds in curvilinear segmentation. clDice gains outrun IoU gains — it repairs "connected or broken," not "accurate or not."

• The Model Says It's Looking At The Image. It's Reciting Its Corpus. CFPO adds a causal constraint to VLM reasoning with a counterfactual signal: erase the image, see if the answer changes. It drops into GRPO or DAPO directly. Gains are single-digit percent — the diagnosis is worth more than the numbers.

Also Notable

• Physics-Driven 3DGS Editing That Adds The Overlooked Dimension: Temperature — MeGAS does thermodynamic dynamic Gaussian splatting, giving rendering and simulation that need realistic physical interaction one more dimension to work with.
• An Adaptive Knob Between Capturing Detail And Suppressing Noise For Implicit Neural Representations — this one uses damped oscillation for spectral gating, resolving the old bind where periodic activations memorize noise while compact ones over-smooth.

Read the full edition →