LoongFlow: Directed Evolutionary Search via a Cognitive Plan-Execute-Summarize Paradigm

TL;DR

LoongFlow tackles the cognitive bottlenecks of existing LLM-driven evolution by replacing blind mutation with a Plan-Execute-Summarize (PES) paradigm and a Hybrid Evolutionary Memory that preserves diversity via MAP-Elites and adaptive Boltzmann selection. The framework decouples reasoning from evolutionary dynamics, using lineage-aware planning, polymorphic execution with fast-fail verification, and abductive memory updates to enable meta-learning. Empirical results on the AlphaEvolve and Kaggle benchmarks show state-of-the-art performance with substantially lower compute, including rapid breakthroughs on high-difficulty tasks and strong pipeline optimization performance. This work advances autonomous scientific discovery by delivering a sample-efficient, reasoning-enabled evolutionary system with open-source agents for algorithmic discovery and ML engineering.

Abstract

The transition from static Large Language Models (LLMs) to self-improving agents is hindered by the lack of structured reasoning in traditional evolutionary approaches. Existing methods often struggle with premature convergence and inefficient exploration in high-dimensional code spaces. To address these challenges, we introduce LoongFlow, a self-evolving agent framework that achieves state-of-the-art solution quality with significantly reduced computational costs. Unlike "blind" mutation operators, LoongFlow integrates LLMs into a cognitive "Plan-Execute-Summarize" (PES) paradigm, effectively mapping the evolutionary search to a reasoning-heavy process. To sustain long-term architectural coherence, we incorporate a hybrid evolutionary memory system. By synergizing Multi-Island models with MAP-Elites and adaptive Boltzmann selection, this system theoretically balances the exploration-exploitation trade-off, maintaining diverse behavioral niches to prevent optimization stagnation. We instantiate LoongFlow with a General Agent for algorithmic discovery and an ML Agent for pipeline optimization. Extensive evaluations on the AlphaEvolve benchmark and Kaggle competitions demonstrate that LoongFlow outperforms leading baselines (e.g., OpenEvolve, ShinkaEvolve) by up to 60% in evolutionary efficiency while discovering superior solutions. LoongFlow marks a substantial step forward in autonomous scientific discovery, enabling the generation of expert-level solutions with reduced computational overhead.

Figures (4)

• Figure 1: Overview of LoongFlow.
• Figure 2: Expanded view of the LoongFlow evolutionary process. The framework iterates through a Planner-Executor-Summarizer loop. The Planner retrieves historical insights to prune the search space; the Executor generates and verifies code; the Summarizer extracts causal knowledge to update the Evolutionary Memory.
• Figure 3: Evolutionary Effect & Efficiency.
• Figure 4: Score Convergence over Time. Note: Faint lines represent individual runs ($N=3$), and bold lines represent the average trajectory.