Description
Training stronger agents usually needs lots of human-labeled data, costly and hard to scale. The Agent0 series proposes agents that self-evolve through tool-integrated reasoning with zero external data.
One agent sets problems and another solves them, improving together; from UNC, Salesforce and Stanford, at ICML and COLM 2026.
Zero data:No labels.
Self-evolving:Curriculum and solver.
Tools:Integrated reasoning.
One agent sets problems and another solves them, improving together; from UNC, Salesforce and Stanford, at ICML and COLM 2026.
Features
Zero data:No labels.
Self-evolving:Curriculum and solver.
Tools:Integrated reasoning.
