HyEvo: Self-Evolving Hybrid Agentic Workflows for Efficient Reasoning

Source: arXiv Authors: Beibei Xu, Yutong Ye, Chuyun Shen, Yingbo Zhou, Cheng Chen, Mingsong Chen Date: 2026-03-20 Primary category: cs.AI All categories: cs.AI

Abstract

HyEvo evolves hybrid workflows containing both probabilistic LLM nodes and deterministic code nodes. Its evolutionary loop searches not only workflow topology but also node logic, using execution feedback to reduce cost and latency while preserving reasoning quality. For harness design, this is a useful reminder that the most evolvable workflow substrate may be mixed-mode rather than LLM-only.