SwarmEvolve: latent-space agent swarms beat AlphaEvolve solutions with an 8B model
brownu · hf · 2026-11-07
SwarmEvolve introduces LLM-guided evolutionary search using swarms of agents that reason in latent space, targeting program optimization for math and scientific computing.
Key design:
- Hundreds of low-cost agents per round work in swarms, reasoning without decoding to text and sharing information by passing KV caches
- Each swarm studies the problem from a different angle; a novelty filter measures distances between hidden-state vectors to keep the most distinct agent ideas
- Crossover happens naturally: a summarizer agent receives the combined KV caches of all swarm members
Results: highest average score on 21 problems in mathematical discovery and scientific computing, reaching in a median of 3-8 minutes the best scores other state-of-the-art frameworks need 60 minutes for. On several problems, SwarmEvolve with Qwen3 8B improves on solutions AlphaEvolve first found with frontier LLMs.
More from coding & agent
- Scope benchmark: only 16.6% of 1,035 AI agent attempts can sign up for hosted databases — RaphaelDabadie · 2026-11-07
- Every cloned its Slack with 51 AI coworkers trained on chat history via DSPy GEPA — lateinteraction · 2026-11-07
- User taunts AI coding agent about an edge case — it had already handled it at line 803 — alexanderchen · 2026-11-07
- GitHub Copilot adds one-time sign-in across Figma, Sentry, Slack, synced to every surface — 0xkarasy · 2026-11-07
- 'No moat, it's all a .md file': recreating a lens perspective illusion with Claude Code — itsOmSarraf_ · 2026-11-07
- Open-source tool shows system prompt and tools can eat 98% of API tokens — behrouzbk · 2026-11-07