NUWA Risk Evaluation · Benchmark
PostTrainBench: Can LLM Agents Automate LLM Post-Training?
NUWA research note · independent review pending
NUWA Research Note
Research questionDesign and assumptionsSource findingsNUWA interpretationLimits and contrary evidenceResearch and decision relevanceInstitutions and rolesOriginal resourcesModel-assisted reading · Selected sections of the original full text or official article; a complete section-by-section audit remains pending.
At a glance
NUWA interpretation: selected post-training gains do not demonstrate general autonomous AI research.
Research question
Can coding agents autonomously improve a supplied base model through post-training?
Design and assumptions
Each run pairs one of four small base models with a target benchmark, constrained to ten hours on one H100 and held-out checkpoint evaluation.
Source findings
The best reported agent average was 23.2%, against 7.5% base and 51.1% official instruction-tuned averages.
Some targeted model/task pairs exceeded the official counterpart.
NUWA interpretation
NUWA interpretation: selected post-training gains do not demonstrate general autonomous AI research.
Limits and contrary evidence
Single-task optimization, constrained compute and only three frontier-agent runs limit generalization and variance estimates; contamination judgments can err.
Research and decision relevance
Model recommendation: retain target task and checkpoint integrity audits alongside performance gains.
Precise source locations
- PostTrainBench : Can LLM Agents Automate LLM Post-Training? · AbstractSelected original resource https://arxiv.org/html/2603.08640; sha256:09f5be78dc2b8f33ac7a82fc692485521fd9442d75cca2bd6d3aad76d727d70c; archived source capture; exact capture time not independently established
- PostTrainBench : Can LLM Agents Automate LLM Post-Training? · §1 Introduction: Our benchmarkSelected original resource https://arxiv.org/html/2603.08640; sha256:09f5be78dc2b8f33ac7a82fc692485521fd9442d75cca2bd6d3aad76d727d70c; archived source capture; exact capture time not independently established
- PostTrainBench : Can LLM Agents Automate LLM Post-Training? · §3.1 Main ResultsSelected original resource https://arxiv.org/html/2603.08640; sha256:09f5be78dc2b8f33ac7a82fc692485521fd9442d75cca2bd6d3aad76d727d70c; archived source capture; exact capture time not independently established
- PostTrainBench : Can LLM Agents Automate LLM Post-Training? · §7 Discussion: LimitationsSelected original resource https://arxiv.org/html/2603.08640; sha256:09f5be78dc2b8f33ac7a82fc692485521fd9442d75cca2bd6d3aad76d727d70c; archived source capture; exact capture time not independently established
Source checking, extraction, editorial approval and replication are recorded separately. Methodology →
Institutions and roles
ELLIS Institute Tübingen
Max Planck Institute for Intelligent Systems
University of TübingenThoughtful Lab
Authors (7)
Ben Rank · Hardik Bhatnagar · Ameya Prabhu · Shira Eisenberg · Karina Nguyen · Matthias Bethge · Maksym Andriushchenko
Evaluation protocols
Cite and reuse
Cite the original publication for its findings. Cite this NUWA record separately for its explanation and retained snapshot.
Original publication · BibTeX
@misc{posttrainbench,
title = {PostTrainBench: Can LLM Agents Automate LLM Post-Training?},
url = {https://arxiv.org/abs/2603.08640},
author = {Ben Rank and Hardik Bhatnagar and Ameya Prabhu and Shira Eisenberg and Karina Nguyen and Matthias Bethge and Maksym Andriushchenko},
year = {2026},
date = {2026-03-09},
version = {arXiv:2603.08640v2},
note = {arXiv:2603.08640v2}
}
NUWA record · BibTeX
@misc{posttrainbench-nuwa,
title = {PostTrainBench: Can LLM Agents Automate LLM Post-Training? — NUWA research record},
url = {https://whitzard.tech/en/nuwa/whitzard-index/work/posttrainbench/},
version = {2026-10-29.1},
organization = {NUWA Frontier AI Safety Lab},
howpublished = {Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--posttrainbench.json},
note = {NUWA research note. Source-reading scope and independent review status are stated on the record. Snapshot 2026-10-29.1. Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--posttrainbench.json.}
}
Evidence snapshot 2026-10-29.1 · Index updates · RSS
References
- [1] Ben Rank ∗, Hardik Bhatnagar ∗, et al.. PostTrainBench: Can LLM Agents Automate LLM Post-Training?. 2026-03-09. v2 [arXiv]; Author affiliations on the original paper title page; Abstract; §1 Introduction: Our benchmark