REFLEX: Reflective Evolution from LLM Experience

A new framework called REFLEX separates visual diagnosis from code generation in large language model-driven policy search, delivering auditable mutation traces and cross-run knowledge transfer, according to a preprint posted to arXiv on June 15, 2026 [1]. The REFLEX architecture, described as train-free, uses a vision-enabled Critic to produce structured behavioral diagnoses before a text-optimized Actor generates child policies. The Actor draws on a self-evolving Skill Memory of reusable code snippets, creating a transparent record of each mutation [1]. This design contrasts with prior approaches that entangled diagnosis and repair in a single model call, which the authors say obscured the reasoning behind mutations and blocked the reuse of algorithmic insights across independent runs [1]. In benchmark tests, REFLEX solved the Acrobot and Pendulum control tasks in under 10 LLM calls and achieved a best Normalized Weighted Score of 1.092 on Lunar Lander [1]. The framework was also evaluated on a 36-dimensional antenna array synthesis task, demonstrating what the preprint calls exceptional sample efficiency [1]. The paper does not include quotes from the authors. Large language models are machine learning systems with many parameters, trained on vast text corpora for natural language processing tasks such as language generation [8]. The preprint hosting platform, arXiv, is an open-access repository of electronic preprints that are moderated but not peer reviewed. It was founded in 1991 and surpassed two million articles by the end of 2021, with a submission rate of roughly 24,000 articles per month as of November 2024 [6]. The REFLEX paper appears alongside a suite of community-developed tools on the arXiv abstract page, including the Bibliographic Explorer for citation-tree navigation and the CORE Recommender for discovering related open-access papers [5]. These integrations are part of arXivLabs, a framework launched in 2020 that allows third-party collaborators to build experimental features on the site under guidelines emphasizing openness, community, and user data privacy [4]. arXivLabs is currently pausing new proposals while the development team focuses on migrating systems to the cloud, though existing projects remain unaffected [3].