digg/ai
2d ago

Study Shows How Architecture and Data Scaling Boost Diffusion Planner Generalization

——0——
Original post
florian_shkurti
Florian Shkurti#1881@FLORIAN_SHKURTI

We have a paper on understanding the design decisions that can increase compositional generalization in diffusion planners. We show that you can do this by careful architecture design, scaling data, or frequent replanning. One of my favourite projects from my lab ⬇️

7:50 AM · May 14, 2026 View on X
Reposted by
Abhishek Gupta#541@ABHISHEKUNIQUE7
QUOTE POSTFlorian Shkurti#1881Florian Shkurti@FLORIAN_SHKURTI

We have a paper on understanding the design decisions that can increase compositional generalization in diffusion planners. We show that you can do this by careful architecture design, scaling data, or frequent replanning. One of my favourite projects from my lab ⬇️

Quentin ClarkQuentin Clark@clark_quen44242

Why do some diffusion planners memorize training trajectories, while others generalize and are able to produce new trajectories? Is it the data, the model, the inference method? Our paper sheds some light on this! 👇🧵

3:12 PM · May 12, 2026 · 10K Views
2:50 PM · May 14, 2026 · 7.8K Views
Study Shows How Architecture and Data Scaling Boost Diffusion Planner Generalization · Digg