AutoTrainess teaches language models to autonomously improve other language models

AutoTrainess is a language model agent designed to automate the post-training process, addressing the human-intensive nature of training frontier models. It exposes planning, data preparation, training, evaluation, and logging as a repository of agent-computer interfaces.

• AutoTrainess externalizes prior human experience through explicit workflows, rules, and execution constraints to guide reliable training behavior.
• On PostTrainBench, it achieves an average score of 26.94 with GPT-5.4 (Codex), outperforming CLI-only baselines which scored 23.21.
• The system generalizes across models and harnesses, improving DeepSeek-V4-Flash (OpenCode) from 12.13 to 19.58.

By automating the complex cycle of iteration and evaluation, AutoTrainess enables language model agents to autonomously improve other language models without manual intervention.