PostTrainBench one year on: AI agents now run complex post-training pipelines

maksym_andr · x · 2026-11-17

One year after releasing PostTrainBench, which measures how well agents like Claude Code can post-train base LLMs, maksymandr reports progress outpaced expectations: frontier models went from barely launching simple SFT runs to reliably implementing complex (if not yet novel) post-training pipelines. The leaderboard shows CLI agents powered by various models scoring from 18.1% (Gemini 3.1 Pro) up to 44.6% (Fable 5.2 Max) vs an 8.0% base-model baseline, with official instruct models at 48.4% — a key indicator for AI R&D automation.

Original post →

More from coding & agent

coding & agent channel →