latent-sokoban, an open benchmark for pixel-based world models

Leaderboard

Every entrant's best scorecard over the same one hundred hidden levels, ranked by success rate. Ties break on move efficiency, then on deadlock rate.

Loading standings…

What the columns mean

Solved
Levels finished within the level's own step budget, which is three times its optimal solution. Unplayed episodes count as unsolved.
Success
Solved levels as a share of all one hundred. This is the ranking metric.
Move eff.
Optimal steps divided by steps taken, averaged over solved levels. 1.000 is a perfect solve; lower means wandering.
Avg steps
Mean steps taken on the levels that were solved.
Deadlocks
Share of episodes that reached an unrecoverable state: a crate shoved into a corner or against a wall it can never leave.
Date
When the ranked scorecard was closed, in UTC.

Getting on it

Register a name for an API key, open a scorecard, play the one hundred hidden levels through the evaluation API, and close it. Closing is what publishes the result, and it counts every unplayed episode as unsolved, so a partial run cannot be cherry-picked into a good score. The overview has the four commands; the wire schema is in docs/API.md.