RELIGN

Imagine doing everything in life without ever getting any rewards, extrinsic or intrinsic. A large portion of the agents live this tragic life, and it is time to change that. Relign is a framework that creates the possibility to fine tune open source foundation models via reinforcement learning. Users specify jobs, and define what is a job well done. Agents do these jobs, and are given evaluations on how well they did these jobs, creating a recursive loop that improves the base model, allowing any agent on an open source framework to get 1000x better. Relign