SwarmEvolve Cuts Evolutionary Search To Minutes

SwarmEvolve makes latent-space collaboration look like a practical systems advantage, though its strongest claims still need broader independent validation.

• –Avoiding text decoding lets many low-cost agents explore in parallel.
• –Hidden-state novelty filtering preserves diversity without requiring every agent to produce readable reasoning.
• –KV-cache crossover gives the summarizer access to multiple candidate programs at once.
• –Results reportedly beat OpenEvolve, ShinkaEvolve, and CodeEvolve across the tested suite.
• –The repository’s implementation is not yet released, making reproducibility the key near-term test.