ATLAS: Embedding-Guided Quality-Diversity Search for Scaffold-Free Algorithm Synthesis

A recent paper published on arXiv (2608.15546) presents ATLAS, a quality-diversity framework guided by embeddings for scaffold-free full-algorithm synthesis in combinatorial optimization. Unlike conventional LLM-driven automated algorithm design approaches that focus on optimizing a specific component within a predefined scaffold, ATLAS empowers the LLM to select and reorganize components, interactions, and control flow, requiring only a minimal I/O interface to define instance and solution formats. The framework autonomously identifies execution, interface, and feasibility failures, recalculates objectives, and implements error-conditioned repairs. To prevent premature convergence, similarity-based archive management maintains algorithms across various embedding-space regions. A three-layer search process enhances the best design while exploring additional regions. The paper can be accessed at https://arxiv.org/abs/2608.15546.