Apple's RLTL;DR lifts Qwen 3.5 9B from 0% to 13% Pass@1

Reinforcement learning with verifiable rewards normally works by having an AI try a task many times and optimizing toward the attempts that succeed. That breaks down when tasks are so hard that the model has low or even no chance of success, and when there are no teacher models or example solutions to learn from. Apple's RLTL;DR targets exactly that gap, showing the policy the verifier's output after each failure and letting it write its own single TL;DR insight, with the next attempt conditioned on all previous insights.

The results indicate the internalized insight mapping, not the insights left in context at test time, is what drives the gain. During training, with insights in context, RLTL;DR reaches 14–31% Pass@1; at evaluation, with no insight present, it still reaches 12–13%. A stripped-down variant, SFTL;DR, trains only on (task, insight) tuples without any rollouts, and just 4k of those tuples recovers almost the full performance of the complete method.

The stakes appear to hinge on whether that compacted pattern, described by the authors as "on this sort of task, keep this sort of thing in mind", generalizes beyond the tool-calling and coding datasets used here. If it does, teams working on problems where no successful attempt exists may have a way to generate training signal where none was available before.