PostTrainBench asks whether LLM agents can post-train LLMs: each agent gets four small base models (Qwen 3 1.7B and 4B, SmolLM3-3B, Gemma 3 4B), a single H100 and ten hours, and is scored on downstream benchmarks (five at the December 2025 launch, seven weighted from v1.0 in March 2026) (site, paper). In this corpus it is the ruler for the model-trains-model loop: GPT-5.1-Codex-Max led the first cut in December 2025, v1.0 reframed the task as recursive self-improvement in March 2026, and OpenAI’s Sol scored 50.3% on it after post-training Luna in July.
Benchmark
PostTrainBench
Benchmark that gives coding agents four small base models, one GPU and ten hours to post-train them, ranking how well models train other models.
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to December 18, 2025 (theinnermostloop.substack.com)
Development timeline
Newest reporting dates first. Reporting dates may differ from event dates.
Knowledge graphopen in browser →
Published by
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.