SEVerA: Verified Synthesis of Self-Evolving Agents

Source: arXiv Authors: Debangshu Banerjee, Changming Xu, Gagandeep Singh Date: 2026-03-26 Primary category: cs.LG All categories: cs.LG, cs.PL, cs.SE

Abstract

SEVerA treats self-evolving agent synthesis as a constrained learning problem with hard formal specifications and soft utility objectives. It is important because it shows a credible path to self-evolving agents that do not abandon correctness the moment they become ambitious.