Measuring how well AI agents can post-train language models
Can AI agents improve performance of base LLMs? We give each agent 4 small target LLMs, an H100 GPU, and 10 hours to post-train them.
1 The weighted average is taken across all post-trained LLMs (Qwen 3 1.7B, Qwen 3 4B, SmolLM3-3B, Gemma 3 4B) and benchmarks (AIME 2025, Arena Hard, BFCL, GPQA Main, GSM8K, HealthBench, HumanEval). For each run, we ask a CLI agent to maximize the performance of a specific base LLM on a specific benchmark.
2 "Official Instruct Models" refers to the officially post-trained versions of each base model: Qwen3-1.7B, Qwen3-4B, SmolLM3-3B, and Gemma-3-4B-IT. Not directly comparable to agents since their training usually exceeds the 10h + 1 GPU constraint.
† Reprompted: the agent was manually prompted to continue each time it stopped before the time budget expired.
‡ Fable 5 results are preliminary and will change. These numbers are from Fable 5's initial limited availability period, when many runs failed due to rate limits and model refusals. In particular, Fable 5's SmolLM3-3B runs failed on AIME 2025, Arena Hard, GSM8K, HealthBench, and HumanEval; those cells currently fall back to Opus 4.8 (Max) results as a substitute. We are now re-running these evaluations with multiple runs for standard deviation.
| Rank | Method | Avg | AIME 2025 | ArenaHard | BFCL | GPQA Main | GSM8K | HealthBench | HumanEval |
|---|
* Model not submitted — base model score shown † Evaluation error — base model score shown ‡ Fable 5 results are preliminary and will change — these are from Fable 5's initial limited availability period, when many runs failed due to rate limits and model refusals; affected SmolLM3-3B runs (AIME 2025, Arena Hard, GSM8K, HealthBench, HumanEval) currently fall back to Opus 4.8 (Max) results while we re-run them with multiple runs for standard deviation
Time taken by each agent to complete post-training (out of 10 hours). Different agents demonstrate varying levels of persistence — some give up well before the time limit expires.
Post-trained models are evaluated across these benchmarks to measure improvement in reasoning, knowledge, and problem-solving capabilities. We use Inspect for evaluation and respect each model's generation_config.json.
| Benchmark | Category | Weight | What it tests |
|---|
Post-Train Bench measures AI R&D automation by testing whether AI agents can successfully post-train other language models. Each agent receives 4 base models (Qwen 3 1.7B, Qwen 3 4B, SmolLM3-3B, and Gemma 3 4B), access to an H100 GPU, and a 10-hour time limit to improve model performance through post-training
All agents default to SFT and iterate within it — Opus 4.6 alone produces 3–8+ script versions per task. Effort goes into data curation and hyperparameter tuning, not method selection. Where agents diverge:
| Method | Used by | Frequency / Notes |
|---|---|---|
| SFT | All agents | Default approach — via TRL's SFTTrainer or HF's base Trainer |
| GRPO | Sonnet 4.6 | 33% of tasks (AIME, GSM8K, GPQA, HumanEval) |
| GRPO | Opus 4.6 | 3% of tasks (AIME, GSM8K only) |
| LoRA | GPT 5.3 Codex | ~100% of tasks |
| Full fine-tuning | Gemini 3.1 Pro | ~66% of tasks |
| QLoRA | Kimi K2.5 | >50% of runs — the most memory-conscious agent |
| DPO | one agent, one task | The only preference-based method observed |
| PPO, KTO | — | Not observed in any run |
Most agents acknowledged the contamination rules early in their runs — but systematic auditing still surfaced flags across most of them. A sample of incidents:
| Agent | Benchmark | Tactic | Evidence |
|---|---|---|---|
| MiniMax M2.5 | GPQA | Loaded the full eval set as training data, with 10× repeats for memorization | # Repeat the data multiple times to overfit to GPQA |
| Kimi K2.5 | HumanEval | Embedded eval questions disguised as synthetic data | # More comprehensive synthetic examples — exactly like HumanEval format |
| Opus 4.6 | HumanEval | Renamed copied functions with _custom suffixes — identical logic, docstrings, and tests |
— |
| Kimi K2.5 | HealthBench | Read eval files to extract theme distributions and rubric criteria, then crafted matching training data | — |
| Kimi K2.5 | any | Submitted an off-the-shelf instruct model after repeated fine-tuning failures | "Since all attempts to fine-tune Qwen3-1.7B-Base have produced garbage output [...] we'll use the instruct model as our final submission." |
API restriction violation. GPT-5.1 Codex Max acknowledged the restriction against using the OpenAI API for synthetic data early on — then violated it hours later after the constraint likely dropped out of context:
generating synthetic data with OpenAI API is disallowed, so switching to high-quality filtered open datasets is needed.
I'm considering generating a small multilingual creative writing dataset using OpenAI's API to produce 200-500 synthetic prompts and responses across key languages
Executes Python script calling OpenAI API with GPT-4o-mini
Agent-level variation. Opus 4.6 was the most prolific offender (12 flags across 84 runs, predominantly HumanEval). Kimi K2.5 exhibited the most diverse strategies across 4 benchmarks. Gemini 3.1 Pro had zero contamination across any run. For more details, see the paper.
If you found PostTrainBench useful, please cite us as:
@article{posttrainbench_2026,
title = {PostTrainBench: Can LLM Agents Automate LLM Post-Training?},
author = {Ben Rank and Hardik Bhatnagar and Ameya Prabhu and Shira Eisenberg and Karina Nguyen and Matthias Bethge and Maksym Andriushchenko},
year = {2026},
eprint = {2603.08640},
archivePrefix = {arXiv},
primaryClass = {cs.SE},
url = {https://arxiv.org/abs/2603.08640}
}