Adaption announced the launch of AutoScientist, a new system designed to automate the research and development loop behind AI model training and alignment. The company said the platform enables users to co-optimize training data and model recipes automatically, with the goal of making advanced model adaptation accessible beyond a small group of frontier AI researchers.
According to Adaption, fewer than a thousand people worldwide currently possess the expertise needed to shape frontier AI models, with much of that knowledge concentrated inside a handful of labs and shared informally between researchers. The company argues that this dynamic has limited broader participation in model development and forced most users to rely primarily on prompt engineering instead of training or adapting models directly.
AutoScientist was created to address common challenges associated with fine-tuning and reinforcement learning, including catastrophic forgetting, overfitting on limited datasets, and conflicting training signals. Adaption said the system automates the entire research cycle, continuously improving both the training data and model recipes until the resulting model converges on a target behavior specified by the user.
The company said AutoScientist allows developers to move from concept to a customized model within hours rather than weeks. Adaption also positioned the platform as a tool for enterprises seeking to reduce reliance on proprietary external models, lower inference costs through smaller fine-tuned models, and unlock the value of proprietary datasets for AI systems.
Adaption reported that AutoScientist outperformed human-configured training setups by an average of 35% across internal benchmark runs. According to the company, win rates improved from 48% to 64% when AutoScientist recommendations were used instead of configurations selected by internal AI researchers. The benchmarks covered multiple dataset sizes, verticals, and model architectures available through Together AI.
The company also said AutoScientist demonstrated consistent gains across eight industry verticals without overfitting to any single domain. Adaption noted that many real-world fine-tuning efforts fail to deliver reliable improvements, while AutoScientist produced repeatable performance gains across different models and tasks.
Adaption described AutoScientist as one of its first major releases focused on automating AI research and development workflows. The company said future work will focus on real-time adaptation techniques that can modify model behavior without requiring additional training.
For the next 30 days, AutoScientist will be available free of charge.
