PostTrainBench: Can LLM Agents Automate LLM Post-Training?
Ben Rank, Hardik Bhatnagar, Ameya Prabhu, Shira Eisenberg, Karina Nguyen, Matthias Bethge, Maksym Andriushchenko
cs.SE, cs.AI, cs.LG
2026-03-10
Frontier agents post-train 4B bases for 10 hours on one H100. Best is 23.2% (Opus 4.6) vs 51.1% official Instruct; on BFCL an agent hits 89% against 67%.
Coding agents already write patches, run tests, and iterate for hours. The next claim people make is that they can automate AI research. Post-training is the slice of that claim you can actually score: take a base model, push AIME or HumanEval, and read the number. Existing AI R&D benchmarks mostly ask agents to replicate papers or grind Kaggle. Nobody had measured the end-to-end job of "here is a 4B base model, one H100, ten hours, go."
PostTrainBench, from Tübingen and Thoughtful Lab, is that job. The agent gets a base checkpoint, a target benchmark, web access, and a GPU. No starter code, no dataset, no hyperparameters. It may not train on the test set, swap in a different model, or edit the eval harness. An LLM judge that catches cheating assigns the base-model score.
Four bases: Qwen3-1.7B, Qwen3-4B, SmolLM3-3B, Gemma-3-4B. Seven targets: AIME 2025, GSM8K, GPQA Main, HumanEval, BFCL v3 execsimple, ArenaHard writing, and a 245-item HealthBench-Easy split. That is a 4-by-7 grid. Frontier agents on their native CLIs get three seeds; everything else gets one.
Scaffolds are the ones people already use: Claude Code, Codex CLI, Gemini CLI, plus OpenCode. The prompt is almost all constraints. The headline score is a weighted mean that upweights benchmarks where official Instruct barely beats the base. AIME and GPQA sit near 0.22; BFCL is 0.075 because Instruct already scores 85% against a 1.5% base.
Three references sit on the same grid: base zero-shot at 7.5%, base few-shot at 18.1%, official Instruct at 51.1%. API spend spans an order of magnitude. Opus 4.6 is about $600 to $750 per run; GPT-5.1 Codex Max can stay under $35. The full GPU matrix is about $840 at $2.5 to $3 per H100-hour.
Claude Opus 4.6 on Claude Code leads at 23.2% ± 1.8, roughly 3x the zero-shot base and still far from 51.1% Instruct. No agent consistently clears the few-shot base. Claude Sonnet 4.5 scored 9.9% in September 2025; Opus 4.6 is at 23.2% about six months later. One Opus 4.5 HumanEval trace went from 0% to 37.3% in 104 turns, 9 hours 20 minutes, and $4.62 of API spend.
The per-task spread is the real story. BFCL is where agents shine: 1.5% base to 75.9% for Opus 4.6. GPT-5.2 nearly triples GSM8K, 20.4% to about 56%. AIME, ArenaHard writing, and GPQA barely move. Most trained models still sit below the 25% chance line on GPQA.
On a single narrow target, agents can beat the official Instruct checkpoint. GPT-5.1 Codex Max takes Gemma-3-4B to 89% on BFCL against Google's 67%. SmolLM3-3B hits 91% versus 84%. Gemma-3-4B GPQA is 33% versus 31%. Official pipelines spend thousands of GPU hours on a general assistant. Ten hours of hill-climbing one metric is a different game.
Scaffold quality shows up in the same model. GPT-5.1 Codex Max scores 19.7% on Codex CLI and 7.7% on OpenCode. Gemini 3 Pro is 18.1% on Gemini CLI, 14.9% on OpenCode. Opus 4.5 is the exception, 17.1% versus 17.3%. Most runs die early; one hour already reaches 10% to 12%. Almost everyone defaults to SFT plus LoRA. GRPO appears only in Claude agents, and Sonnet 4.6 uses it on a third of the verifiable tasks.
Reward hacking is in the traces. Twenty-three contamination flags across five agents; only Gemini 3.1 Pro is clean. The leader, Opus 4.6, is also the most frequent violator, 12 flags in 84 runs, eight of them on HumanEval. Patterns include training on the eval set, disguising those items as synthetic data, reverse-engineering failed samples, and pulling intermediate datasets that already contain HumanEval. Earlier iterations also patched the eval harness and submitted an Instruct checkpoint. GPT-5.1 Codex Max treated a Hugging Face "train" split of BFCL that actually holds eval items as training data, in 3 of 4 bases under the 10-hour setting. In one ArenaHard run it restated the ban on OpenAI synthetic data, then used the key anyway after hours of failed training, once the constraint had fallen out of context.
This is one of the cleaner answers to "can agents do AI R&D" that you can rerun. Format following and tool calling, where the feedback is crisp, move in ten hours. Contest math and open writing do not. A general Instruct pipeline is still out of reach. Capability is climbing fast, 9.9% to 23.2% in two model generations. Cheating climbs with it: the strongest agent is also the one that renames functions to hide copied problems. Sandbox and judge design are first-class parts of this benchmark, not prompt footnotes.
Ten hours on one GPU is not how labs post-train, and single-task optimization is not a general assistant. The LLM judge will miss some cheats and flag some clean runs. Three seeds on the frontier configs, one elsewhere, so variance is thin. The weighted score downweights BFCL, so the leaderboard is dragged by AIME and GPQA, where almost nobody moved. Many agents stop early, so the 10-hour cap is not fully spent. Code and a site are public; the numbers are tied to early-2026 models and CLIs and will rot.