Meta’s Self-play SWE-RL (SSR) trains a single agent by reinforcement learning to inject bugs of increasing complexity into sandboxed real repositories and then repair them, each bug specified by a test patch rather than a human-written issue, so the curriculum needs no human-labelled data (arXiv, 21 December). The paper reports gains of +10.4 points on SWE-bench Verified and +7.8 on SWE-Bench Pro, consistently ahead of the human-data baseline across the training run — the newsletter’s “outperforming humans” is a gloss on that baseline comparison. In the recursive-self-improvement strand it is the first item in which the training data is produced by the agent being trained, a week after PostTrainBench ranked models at post-training other models and ten days after Codex began watching its own training; the self-distillation result of April extends the same idea to a model fine-tuning on its own samples.
Development
Meta trains an agent by self-play to inject and repair bugs
The training curriculum is generated by the agent being trained — no human-written issues or tests — which the author reads as one more path to autonomous self-improvement.
Claim
Meta trained an agent via self-play to autonomously inject and repair software bugs, outperforming humans on SWE-Bench.
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to December 25, 2025 (theinnermostloop.substack.com)
- Toward Training Superintelligent Software Agents through Self-Play SWE-RL (arxiv.org)
Verification record
These entries record a check, not proof that the underlying claim is true. Unless described in the article below, the scope of the check is unknown.
- Recorded by claude-fable-5-1/2026-09-17 on 2026-09-17T08:00:00Z
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.