Researchers have developed a new AI system called SwarmEvolve that accelerates the process of discovering and improving mathematical and scientific programs. Unlike previous approaches that rely on large language models (LLMs) generating lengthy text explanations at each step, SwarmEvolve uses hundreds of lightweight AI agents working together silently “behind the scenes” in a compressed, hidden form. This innovation allows the system to explore solutions much faster and more efficiently, potentially transforming how computers tackle complex scientific problems.
Key Takeaways
- SwarmEvolve uses hundreds of low-cost AI agents working in swarms to explore solutions simultaneously, rather than one agent generating long textual reasoning chains.
- The agents communicate and combine ideas in “latent space,” a hidden numerical representation, avoiding the slow process of converting thoughts into text.
- A novelty filter ensures diversity by selecting the most distinct ideas based on differences in hidden state vectors, encouraging fresh approaches.
- On 21 benchmark problems, SwarmEvolve reaches top results in just 3 to 8 minutes, compared to 60 minutes for other leading methods.
The core innovation behind SwarmEvolve lies in how it organizes and communicates the AI agents’ reasoning. Traditional LLM-guided evolutionary search involves generating detailed textual explanations (“chains of thought”) to iteratively improve solutions. While effective, this text generation is computationally expensive and slow, especially when maintaining a diverse population of candidate solutions.
SwarmEvolve sidesteps this bottleneck by having hundreds of agents operate in “latent space.” This term refers to a compressed, abstract representation of information inside the AI model—essentially a set of numerical vectors that capture the essence of ideas without needing to be expressed in words. Instead of writing out thoughts, agents pass around these vectors (called KV caches) that encode their current state and insights.
These latent representations enable agents to “think” and share information much faster and at a lower computational cost. Additionally, SwarmEvolve organizes agents into swarms, each focusing on different perspectives of the problem. A novelty filter compares agents based on the distances between their latent vectors, promoting diversity by favoring unique ideas. When it’s time to combine the best insights from the swarm, a summarizer agent merges the KV caches, effectively performing a “crossover” of ideas in latent space rather than mixing textual outputs.
In tests on a suite of 21 mathematical and scientific programming challenges, SwarmEvolve consistently outperformed existing state-of-the-art frameworks. It achieved the highest average scores and did so significantly faster—often in under 10 minutes where others took an hour. Notably, when paired with the Qwen3 8B model, SwarmEvolve improved solutions originally found by previous top systems like AlphaEvolve, demonstrating its potential to push the boundaries of automated scientific discovery.
Looking ahead, this research suggests that enabling AI agents to collaborate silently in latent space can dramatically speed up evolutionary search methods. This could make automated program optimization more practical for real-world scientific and engineering problems, where time and computational resources are limited. Future work may explore scaling this approach further, applying it to more complex domains, or combining it with other AI techniques to unlock new discoveries more efficiently.
Based on research published on arXiv by Qile Jiang, George Em Karniadakis.