Self-Evolving Agents

A novel framework that converts cross-episode signals into reusable procedural memory and distills it into the policy's weights during training. Memory functions as a training scaffold, enabling memory-free inference while achieving significant improvements over existing methods.

VPD reframes on-policy self-distillation as a Variational Expectation-Maximization problem. The teacher is actively refined on trajectory outcomes (E-step) and then distilled into the student's own rollouts (M-step), with a dynamic trust region anchored to the current policy — co-evolving teacher and student inside a single shared-weight network.