Definition
Plain language
Realizing you've gone down a wrong path, returning to an earlier choice, and trying a different one.
As stated in the literature
The reasoning skill of retreating from a failed branch to an earlier decision point and exploring an alternative; argued to be load-bearing for general reasoning and the specific capability RL learns from its own failed trajectories that imitation learning on clean solutions structurally cannot teach.
Why it matters: Without the ability to abandon a wrong path and retry, a reasoning system gets stuck committing to early mistakes, which is why learning from its own failures matters more than copying clean solutions.
For example, while solving a maze an agent that hits a dead end returns to the last fork and takes the other path instead of pushing uselessly forward.
Heard on the show
“… computational experiments to build intuition, drafting a partial argument, finding a hole in it, backtracking, reformulating definitions. …”Episode 029 — Why Forty-Eight Percent on FrontierMath Isn't the Real Story in DeepMind's New Math Paper