Explorative Modeling matches diffusion in a single forward pass
A new training axis — explore K samples, keep the best — hits 1.43 FID on ImageNet and cuts robot-policy inference from 100 steps to one. Code is on GitHub.
Copy markdown
The idea: exploration as a third pretraining axis
XM samples K candidate outputs during training and back-props only the best match, sidestepping the mode-averaging that turns one-shot generators to mush. The paper reports 6.2× sample, 4.1× FLOP, and 47% better parameter efficiency over standard baselines.
1.43 FID on ImageNet, no guidance needed
On ImageNet 256×256 it hits a near-SOTA 1.43 FID without classifier-free guidance and converges roughly 300× faster than the baseline. Frontier image quality without the guidance-tuning tax.
Robots and planners: 1 forward pass, not 100
XM matches Diffusion Policy on robot manipulation using a single network forward pass instead of ~100, and solves maze planning in about 80× fewer steps. Real-time control and planning get far cheaper.
Build this weekend: wrap your loss in best-of-K
Adoption is a small diff — wrap your existing loss in a loop that explores K candidates and trains on the winner. "Forward XM" pseudocode plus full code are up at github.com/alexiglad/XM.
Elsewhere: a Lean 4 kernel soundness bug, patched in an hour
A phantom-parameter flaw let metaprogramming forge a proof of False in the Lean 4 kernel — patched within an hour (PR #14577). It was only reachable by feeding declarations straight to the kernel, so normal proofs stayed safe.