Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation

Researchers introduced Explorative Modeling, a generative training method that compares multiple candidate outputs against data and learns from the best match.

The approach adds exploration as a third scaling axis alongside model size and data, improving performance across images, video, language, and control tasks.

For reconstructive generation, it can reduce inference steps while improving efficiency, boosting both sample efficiency and FLOP efficiency.

The work points to a new pretraining path for more direct multimodal learning and could make large-scale generative AI more accurate and practical.