PACE: Primitive-Aware Code Evolution for Automated Algorithm Design

2026-08-07

AI Summary

This work addresses a key limitation in traditional large language model–driven automated algorithm design, which treats programs as monolithic entities and struggles to preserve and reuse valuable local logic. The authors propose PACE, a novel approach that explicitly models local algorithmic logic during evolution as persistent, executable algorithmic primitives (EAPs). By introducing primitive-aware evolutionary operators, PACE decouples local structures from global program architecture. Coupled with a Thompson sampling strategy guided by relative performance improvements over parent solutions, PACE enables efficient primitive selection and cross-program transfer. Experimental results across four benchmark tasks demonstrate that PACE not only discovers competitive algorithms but also effectively retains and reuses high-value components, thereby overcoming the constraints of holistic evolutionary paradigms.

Abstract

Large Language Model (LLM)-based automated algorithm design typically evolves algorithms as complete, indivisible programs. While this whole-program perspective simplifies the search space, it fundamentally couples the useful local logic to its host program. Consequently, valuable code snippets vanish when the overall program is discarded, making it highly difficult to assess the contribution of individual algorithmic components.To address this, we propose Primitive-Aware Code Evolution (PACE), which decouples local logic from complete programs by representing it as persistent units called Executable Algorithmic Primitives (EAPs). To enable code-level transfer, PACE maintains a dynamic set of EAPs. Algorithm evolution is driven by primitive-aware operators that structurally guarantee the retention and cross-program transfer of these components. To evaluate them effectively, PACE leverages Thompson sampling based on parent-relative performance improvements, guiding primitive selection from the set without requiring extra evaluation datasets. Experiments on four tasks demonstrate that PACE effectively discovers competitive algorithms while structurally preserving valuable algorithmic components.

Problem

• automated algorithm design
• code evolution
• algorithmic primitives
• program decomposition
• component reuse

Innovation

• Executable Algorithmic Primitives
• Code Evolution
• Primitive-Aware Operators
• Thompson Sampling
• Automated Algorithm Design