SEVerA (Self-Evolving Verified Agents) is a framework for building LLM agents with formal correctness guarantees. It introduces Formally Guarded Generative Models (FGGM), which enforce first-order logic contracts on every model call, ensuring outputs always satisfy specified constraints. SEVerA combines program synthesis, formal verification, and learning in a three-stage pipeline (Search, Verification, Learning) to produce agents that are both provably safe and high-performing across tasks like symbolic reasoning, program analysis, and tool use.
Code will be released soon