preprint · 2026-03-09

PostTrainBench: Can LLM Agents Automate LLM Post-Training?

PostTrainBench team · Ben Rank, Hardik Bhatnagar, Ameya Prabhu +4 more

On this page

Why it matters here

PostTrainBench tests agents improving small AI models with ten hours and one H100 GPU per target run. The September 30 v1.2 release removes BFCL, fixes HumanEval scoring and recomputes results across four base models and six tests. Its pinned leaderboard reports Fable 5.1 at 44.65%, Opus 5.5 at 43.80% and GPT-6 Astra at 41.88%; Fable 5.1 includes five Opus 5 GPQA fallback cells. Earlier v1.1 results remain available as a separate historical snapshot.

Read original source ↗

AI-R&D capabilityAI-system improvementEvaluation integrity

What to keep in mind

The original paper’s March v1 results, the historical v1.1 leaderboard and September 30 v1.2 rescoring are distinct evidence versions.
Changes from v1.1 include measurement and judging changes; lower scores do not by themselves imply weaker models.
Fable 5.1 uses some Opus 5 GPQA cells; historical Fable 5 uses Opus 4.8 Max GPQA fallback. Neither is a pure single-model result.
Individual target performance need not imply broad transfer or sustained improvement of the research agent.
The v1.1 Epoch review is dated evidence; its verdict is not independent verification of v1.2.