TrainExplorer lets agents retrain small models on their own: literature-built strategy space lifts PostTrainBench weighted score to 39.02, past Codex
Synopsis
Agents that autonomously post-train large language models tend to stay inside their initial strategy, and a single experiment often changes data and training settings together, so the effect of each change is hard to attribute. TrainExplorer separates strategy planning from execution: it first builds a candidate strategy space from scientific literature, then splits data processing and model training into independently replaceable components, changing one per iteration and attributing the effect. On PostTrainBench's seven tasks and four base models of 1.7B to 4B parameters, it reaches a weighted aggregate score of 39.02, 7.32 points above Claude Code with the same agent LLM and above GPT-5.6 Sol-powered Codex at 36.23.
Interpretation
TrainExplorer lets a post-training agent switch training strategies mid-run instead of only making local adjustments inside the supervised fine-tuning scheme chosen at the start. Post-training agents previously relied mainly on the model's own knowledge to pick strategies, so iterations tended to stay near the initial strategy and other viable training methods went unexplored. Across the four PostTrainBench base models, TrainExplorer shows a strategy switch rate of 37.50% against Claude Code's 9.49%, a difference of 28.01 percentage points.
The framework splits an experiment into separately replaceable data-processing and model-training components, changing one component per round while inheriting the other implementations and configurations from the previous round. Previously a single experiment often changed the data pipeline and training configuration at once, so a performance change could not be assigned to a specific modification. On Qwen3-1.7B, removing execution decoupling lowers the weighted aggregate score by 3.55, and GSM8K still drops by 5.13 even when experiment history and feedback are retained.
Strategy planning first organizes scientific literature into a searchable training-experience library, then uses it to build a candidate strategy space for the current task. Candidate strategies no longer come only from methods familiar to the agent but cover types such as data selection, data synthesis, data augmentation, data cleaning, SFT and RL. On Qwen3-1.7B, removing strategy planning lowers the weighted aggregate score by 5.26, with marked declines on tasks such as AIME.
Perspective
Evaluation follows the PostTrainBench protocol, covering seven tasks and four base models (Qwen3-1.7B, Qwen3-4B, SmolLM3-3B and Gemma3-4B), with a 10-hour budget on a single 80 GB GPU per run and strategy planning performed once at the start of each run.
The overall weighted aggregate score of 39.02 remains below the official instruction-tuned models' 51.14, leaving a clear gap between autonomous post-training and instruction tuning. Strategy planning depends on whether the training-experience library holds relevant literature, so in emerging or specialized domains with limited reference material the library may give insufficient guidance for building candidate strategies.
