Researchers introduced RLTL;DR, a method for model self-improvement. It lets a policy write its own feedback after each failed attempt. The next rollout is conditioned on all previous insights. It achieves Pass@1 of 14-31% with insights during training. It retains 12-13% Pass@1 without insights at evaluation. The key is internalizing a direct task-to-insight mapping. A simplified version, SFTL;DR, trains only on task-insight pairs. It recovers nearly full performance with just 4,000 tuples. The paper targets tool-calling and coding datasets. It was submitted to arXiv on 29 September 2026.
What changed
Standard GRPO training achieved Pass@1 of 0% to 1% on filtered datasets.
What it unlocks
Models can now learn from self-generated feedback to solve previously unsolvable tasks.
- Pass@1 0-1% to 14-31% with insights
- Pass@1 12-13% without insights
- 4,000 (task, insight) tuples
What you need to act on it
- tasks with Pass@128=0
- Qwen 3.5 9B Thinking policy
Sources