AI tools company Adaption launched AutoScientist on May 13—an automated research loop system designed for model training and alignment. It gained widespread attention this week after being shared by renowned AI researcher Sara Hooker. By jointly optimizing data recipes and model training parameters, AutoScientist automatically iterates until the model’s behavior aligns with user-defined goals; developers no longer need to manually tune parameters or spend endless hours ‘searching for optimal hyperparameters.’ Even non-technical builders can now directly train and adjust models, rather than being limited to prompt engineering alone. Official test results show that across all runs involving 5,000 to 100,000 samples, multiple model architectures, and eight industry verticals, AutoScientist outperformed manual configurations created by Adaption’s internal AI researchers by an average of 35%. Its win rate rose from 48% under manual setups to 64% with AutoScientist, delivering consistent gains across all eight verticals without any signs of overfitting to specific domains. The system supports every fine-tunable model hosted on Together AI.
Adaption positions AutoScientist as the model layer complementing its Adaptive Data platform; together they form an end-to-end closed loop spanning data preparation to model adaptation. As mentioned in Sara Hooker’s post, Adaption is currently offering 30 days of free computing resources to users in East Africa to lower entry barriers—full application details are available on the official blog. What sets AutoScientist apart from mainstream AutoML tools is its self-improvement mechanism: the system continuously enhances its configuration capabilities based on historical operational data, rather than relying on static search strategies.