跳到正文
HuggingFace Daily Papers(社區熱門論文)·· 18 天前AI 評分37

SwarmEvolve:用於演化發現與最佳化的潛在智能體羣

SwarmEvolve: Latent Agent Swarms for Evolutionary Discovery and Optimization

AI 導讀

SwarmEvolve 是一個由 LLM 引導、用於數學發現與最佳化的演化搜尋框架。它以數百個小型智能體進行潛在推理,透過 KV cache 傳輸溝通,不會撰寫文字。使用開放權重 Qwen3 8B、14B 和 32B 時,它只需數分鐘便達到其他框架一小時才能取得的分數,並在少數問題上超越 AlphaEvolve 的結果。

正文

Author here!

SwarmEvolve is an LLM-guided evolutionary search framework for mathematical discovery and optimization.

Unlike many existing frameworks, SwarmEvolve runs swarms of hundreds of small agents that use latent reasoning and communicate through KV cache transfer, so they never write text. As a result, SwarmEvolve is much faster, more scalable, and more diverse in the ideas it explores.

With open-weight Qwen3 8B, 14B and 32B, SwarmEvolve takes only minutes to reach the scores that other frameworks reach in an hour, and it improves on AlphaEvolve's results on a few problems!

Like our agents, we think better as a swarm, so questions and ideas are very welcome!

來源:HuggingFace Daily Papers(社區熱門論文) · huggingface.co