SwarmEvolve, an AI framework for improving computer programs, reports faster search and higher average scores across 21 mathematical and scientific-computing problems than three competing frameworks. The work was posted to arXiv on November 5, 2026. Peer review is not verified, and the findings come from the framework’s own authors.
Evolutionary program search repeatedly generates candidate programs, evaluates them and uses successful candidates to guide subsequent changes. SwarmEvolve aims to reduce the time spent generating written reasoning by having agents work with continuous internal model representations, known as latent states.
Each round launches 10 swarms of 45 agents, or 450 agents altogether. Different swarms refine strong programs, combine approaches, explore alternatives or seek a fresh direction. Each agent takes 40 latent reasoning steps. A filter compares hidden-state vectors and retains 30 agents per swarm to preserve diversity. A summarizer then receives their KV caches—the stored attention keys and values used by a transformer—and generates a complete program as text. Programs are evaluated and archived, while a separate agent proposes numerical parameter changes.
The evaluation compared SwarmEvolve with OpenEvolve, ShinkaEvolve and CodeEvolve using Qwen3 8B, 14B and 32B-FP8. Each framework ran three random seeds per problem and model, with a 60-minute budget, one GPU and eight CPU cores. GPU hardware differed across model sizes, so comparisons between sizes also change hardware.
Sixteen tasks came from published AlphaEvolve problems in packing, geometry, analysis and combinatorics. Five author-designed tasks covered numerical linear algebra and differential equations. Mathematical scores were normalized against AlphaEvolve’s published constructions; scientific-computing scores reflected correct digits against high-fidelity references. The combined average is therefore a benchmark measure, not a percentage accuracy.
With Qwen3 8B, SwarmEvolve’s average best-within-budget score was 0.804, compared with 0.693 for OpenEvolve, 0.642 for ShinkaEvolve and 0.700 for CodeEvolve. It also had the highest average with the two larger models, although it did not lead on every task.
The authors report 5–13 times as many evaluated programs per hour. Median time to reach competitors’ 60-minute scores was 4–8 minutes with 8B and 3–6 minutes with larger models. Some target scores were not reached within the hour.
With 8B, SwarmEvolve reportedly exceeded AlphaEvolve’s published constructions on three tasks, versus one for each comparison framework. Its results included matching the cited best-known sum of radii for 26 circles in a square, 2.635983, and improving two autocorrelation bounds. These comparisons concern published constructions, rather than a matched new AlphaEvolve run.
The evidence supports the complete framework under these test conditions. However, SwarmEvolve also changes search organization, diversity filtering and parameter tuning; the main comparison cannot attribute all gains to latent communication alone.
Only three seeds were tested per condition, and the main benchmark comparison provides no confidence intervals or significance tests. Baselines used default settings, while generalization to unseen problems and possible model exposure to existing solutions remain unresolved. Independent replication and direct monetary or energy-cost measurements are not supplied. All agents must share one model, the implementation requires standard attention, and KV-cache memory grows with agent count, reasoning length and model depth.
Cover illustration generated with AI; conceptual, not research evidence.