Recursive AI develops automated methods that improve neural network performance. Our tools find improvements that human engineers miss — verified across multiple random seeds.
Get in touchWe built a pipeline that automatically discovers modifications which improve neural network performance. The system is model-agnostic, requires no human guidance, and verifies every result across multiple random seeds before reporting it.
ATLAS evaluates candidate modifications in parallel on cloud GPUs, then verifies promising results across multiple random seeds. In our strongest verified result, ATLAS discovered a modification that improved a pretrained transformer's task performance by up to 7.8 percentage points, with no seed showing degraded performance.
Recursive AI LLC develops automated tools for improving neural network performance. Our research has been validated across multiple architectures and is targeting publication at top-tier AI conferences.
William brings over 30 years of academic experience in AI and machine learning. He has held faculty positions at Ritsumeikan Asia Pacific University (Professor Emeritus), Sophia University, Waseda University, and Temple University Japan. He currently serves as Associate Teaching Professor at Northeastern University and directs the Agentic AI Research Lab, a cross-campus research group.
We're exploring partnerships and early applications of our technology.
wclaster@recursiveai.online