research
A Learned Action Scores Whole Scientific Trajectories
LEAP turns observed paths into one differentiable objective reusable for ranking, steering, constrained optimization and sampling.
Summary
LEAP turns observed paths into one differentiable objective reusable for ranking, steering, constrained optimization and sampling.
Rather than learn only local transitions, the method compares observed trajectories with a tractable reference process and learns a correction on path space. Tests span stochastic physical systems, orbital dynamics and molecular simulations. The evidence is computational and does not establish every system’s governing action.
Why it matters
LEAP turns observed paths into one differentiable objective reusable for ranking, steering, constrained optimization and sampling.
Limits and context
- Rather than learn only local transitions, the method compares observed trajectories with a tractable reference process and learns a correction on path space.
- The evidence is computational and does not establish every system’s governing action.
Key claims
LEAP turns observed paths into one differentiable objective reusable for ranking, steering, constrained optimization and sampling.
Qualification: Rather than learn only local transitions, the method compares observed trajectories with a tractable reference process and learns a correction on path space.
Evidence: source-2026-10-24-019
Sources
- arXiv preprint 2610.31585, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.