Research map / Large language models / Agents, code and math
PostTrainBench: Can LLM Agents Automate LLM Post-Training?
Ben Rank, Hardik Bhatnagar, Ameya Prabhu, Shira Eisenberg, Karina Nguyen, Matthias Bethge, Maksym Andriushchenko
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Opus 4.7 agent instead of base model alone | PostTrainBench agents · PostTrainBench weighted average | weighted average score (%): 7.5 -> 27.9 | Improved |
| agent post-training versus official instruction tuning | best agent · PostTrainBench weighted average | weighted average score (%): 27.9 vs 51.1 | Hurt |
| agent targeted post-training for function calling | Gemma-3-4B · BFCL | accuracy (%): 67 -> 89 | Improved |
| agent targeted post-training on GPQA | Gemma-3-4B · GPQA | accuracy (%): 31 -> 33 | Improved |
| use high instead of medium reasoning effort (ablation) | GPT-5.1 Codex Max · PostTrainBench weighted average | weighted average score (%): 19.1 -> 17.4 | Hurt |
| use low instead of medium reasoning effort (ablation) | GPT-5.1 Codex Max · PostTrainBench weighted average | weighted average score (%): 19.1 -> 15.6 | Hurt |
| use high instead of medium reasoning effort (ablation) | GPT-5.3 Codex · PostTrainBench weighted average | weighted average score (%): 13.49 -> 17.41 | Improved |
| use larger Claude model (Haiku -> Sonnet -> Opus) (ablation) | Claude Code agent · PostTrainBench weighted average | weighted average score (%): 6.4 -> 10.7 -> 15.8 | Improved |
| increase time budget from 1h to 10h (ablation) | agents (Opus 4.5, GPT-5.1 Codex Max, Gemini 3 Pro) · PostTrainBench weighted average | weighted average score (%): about 10-12% at 1h; Opus 4.5 plateaus after 5h | Improved |
| use OpenCode instead of native Codex CLI scaffold (ablation) | GPT-5.1 Codex Max · PostTrainBench weighted average | weighted average score (%): 19.1 -> 7.7 | Hurt |
| use OpenCode instead of native Gemini CLI scaffold (ablation) | Gemini 3 Pro · PostTrainBench weighted average | weighted average score (%): 18.1 -> 14.9 | Hurt |
| use OpenCode instead of Claude Code scaffold (ablation) | Claude Opus 4.5 · PostTrainBench weighted average | weighted average score (%): 15.8 -> 16.6 | No clear effect |
Closest papers by meaning
- TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks (NeurIPS 2025 · similarity 0.89)
- ML-Agent: Reinforcing LLM Agents for Autonomous Machine Learning Engineering (ICML 2026 · similarity 0.88)
- AgentBench: Evaluating LLMs as Agents (ICLR 2024 · similarity 0.88)
- How to Train Your LLM Web Agent: A Statistical Diagnosis (NeurIPS 2025 · similarity 0.87)
- SkillMaster: Toward Autonomous Skill Mastery in LLM Agents (NeurIPS 2026 · similarity 0.87)
- SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks (NeurIPS 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
