PostTrainBench v1.0 asks whether an LLM agent can carry out the post-training of another model unaided, and in its first versioned release Claude Opus 4.6 driving Claude Code came out as the most capable such agent (announcement). The leaderboard first appeared in December with GPT 5.1 Codex Max on top; the version number and the change of leader are what the newsletter marks. Four days later Anthropic’s alignment lead would call recursive self-improvement a present phenomenon, and in July a model post-trains a model at 50.3% on this benchmark.
Development
A benchmark measures whether agents can post-train themselves
The December leaderboard for models post-training models returns with a version number and a new leader, giving the recursion the newsletter is named for a score to track.
Claim
PostTrainBench v1.0 evaluated whether language model agents can automate their own post-training for recursive self-improvement, finding Claude Opus 4.6 with Claude Code the most capable such agent.
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to March 12, 2026 (theinnermostloop.substack.com)
- PostTrainBench v1.0 announcement (x.com)
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
Knowledge graphopen in browser →
Related to
Reported in
Actor
Referenced by
- Nvidia sees no wall in post-training and opens a 120B model
- An alignment lead says recursive self-improvement is already happening
- A model improves by fine-tuning on its own samples
- A benchmark scores models on writing benchmarks
- An open model beats the frontier on ML engineering under cost limits
- A model autonomously post-trains another model
- Self-improving agents rebuild the inference stack serving two open models
- Welcome to March 12, 2026
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.