self-evolve

Definition

A model or agent that improves itself from its own outputs and execution results, without new human labels each round. The loop has the system generate tasks, attempt them, score the attempts by some verifiable signal, and train on what survived — an answer to human annotation being the bottleneck. It differs from distillation and RLHF in where supervision originates, which is also why such claims need care about reward hacking.

Explainers using this term

• Paper Walkthrough: Aspire — Can Models Self-Evolve from Vague Goals?Aspire: Can Models Self-Evolve from Vague Goals?