GigaEvo: An Open Source Optimization Framework Powered By LLMs And Evolution Algorithms

TL;DR

GigaEvo presents a modular, open-source framework that couples LLM-driven mutation with MAP-Elites-based quality-diversity search through an asynchronous DAG execution pipeline and a Redis-backed program archive. It emphasizes bidirectional lineage, context-rich mutation via lineage-informed insights, and heterogeneous LLM routing to explore problem spaces ranging from geometric optimization to NLP prompts. The authors reproduce and extend AlphaEvolve results on Heilbronn triangle configurations, circle packing, and kissing numbers, and demonstrate competitiveness on bin packing and prompt/agent evolution tasks. The work provides concrete implementation guidance, assesses reproducibility, and offers a scalable platform for future automated algorithm discovery and mathematical construction.

Abstract

Recent advances in LLM-guided evolutionary computation, particularly AlphaEvolve (Novikov et al., 2025; Georgiev et al., 2025), have demonstrated remarkable success in discovering novel mathematical constructions and solving challenging optimization problems. However, the high-level descriptions in published work leave many implementation details unspecified, hindering reproducibility and further research. In this report we present GigaEvo, an extensible open-source framework that enables researchers to study and experiment with hybrid LLM-evolution approaches inspired by AlphaEvolve. Our system provides modular implementations of key components: MAP-Elites quality-diversity algorithms, asynchronous DAG-based evaluation pipelines, LLM-driven mutation operators with insight generation and bidirectional lineage tracking, and flexible multi-island evolutionary strategies. In order to assess reproducibility and validate our implementation we evaluate GigaEvo on challenging problems from the AlphaEvolve paper: Heilbronn triangle placement, circle packing in squares, and high-dimensional kissing numbers. The framework emphasizes modularity, concurrency, and ease of experimentation, enabling rapid prototyping through declarative configuration. We provide detailed descriptions of system architecture, implementation decisions, and experimental methodology to support further research in LLM driven evolutionary methods. The GigaEvo framework and all experimental code are available at https://github.com/AIRI-Institute/gigaevo-core.

Figures (4)

• Figure 1: GigaEvo system architecture. The framework comprises four core components (highlighted in green): (i) Redis Database --- a Redis storage for evolutionary units, which contain UUID, source code, lifecycle state, metrics, and lineage information, supporting concurrent access via optimistic concurrency control; (ii) DAG Execution Engine --- an asyncio-based pipeline processing programs through configurable stages (execution, validation, complexity analysis, LLM inference); (iii) Evolution Engine --- an asynchronous loop implementing MAP-Elites quality-diversity algorithm that maps programs to behavior space cells (fitness × validity), performs fitness-proportional selection of elites, and supports single-island and multi-island configurations with periodic migration; and (iv) Mutation Operator --- a LangGraph-based agent that constructs prompts from task descriptions, parent code, metrics, insights, and lineage analyses, then invokes LLMs to generate offspring programs using rewrite-based or diff-based mutation modes with multi-model routing support. All components interact through the Redis storage layer, enabling modular, concurrent evolutionary search.
• Figure 2: Part of the DAG pipeline in the GUI used to process programs in our experiments. Programs are first validated (ValidateCodeStage), executed (CallProgramFunction), and evaluated via a problem-specific validator (CallValidatorFunction). Complexity metrics (ComputeComplexityStage) are merged with validation metrics (MergeMetricsStage) and ensured complete via EnsureMetricsStage. Individual program insights are generated (InsightsStage), while lineage analysis proceeds through multiple stages: LineageStage computes parent$\to$current transition analyses; AncestorProgramIds and DescendantProgramIds select relevant relatives; LineagesFromAncestors and LineagesToDescendants gather bidirectional lineage insights. Finally, MutationContextStage assembles all information into a structured context for the mutation operator. The framework supports arbitrary DAG topologies through declarative configuration.
• Figure 3: Heilbronn triangle problem results for $n=11$ points.Left: Point configuration discovered by GigaEvo, maximizing the minimum triangle area formed by any three points in a unit-area equilateral triangle. The evolved program places points both on the boundary and in the interior, exploiting geometric symmetries while avoiding near-collinear triplets. Right: Fitness evolution over 20 evolution generations, showing rapid initial progress followed by refinement. The dashed lines indicate benchmark results: AlphaEvolve achieved $0.0365$ (green), while the previous best friedman_main_website was approximately $0.036$ (red). GigaEvo achieved a fitness of $0.0364$, which is $10^{-4}$ below the reported AlphaEvolve result. By visual inspection (fig. 26 left panel in georgiev2025mathematical), the point configurations are nearly identical, suggesting GigaEvo successfully reproduced the AlphaEvolve solution. The small fitness difference could be eliminated with additional computational budget for fine-tuning the simulated annealing parameters in the final optimization phase. This demonstrates that our implementation captures the essential algorithmic principles underlying AlphaEvolve's success.
• Figure 4: Circle packing in unit square for $n=26$ circles. Optimal configuration discovered by GigaEvo, maximizing the sum of radii $\sum_{i=1}^{26} r_i$ of disjoint circles packed within the unit square. GigaEvo achieved a total radius sum of $2.636$, slightly surpassing the AlphaEvolve result of $2.635$ reported in novikov2025alphaevolve. Visual comparison with the AlphaEvolve solution (fig. 28 in georgiev2025mathematical) reveals nearly identical qualitative structure, with minor improvements likely attributable to refined local optimization of centers and radii. This demonstrates that our implementation successfully reproduces and marginally improves upon AlphaEvolve's results through effective exploration of the constrained geometric optimization landscape.