SwarmEvolve: 450 latent-space agents boost program search 5-13x
aigclink · x · 2026-11-09
Brown University researchers present SwarmEvolve (arXiv 2611.08483), a multi-agent evolutionary search framework where agents reason in latent space without decoding text and share information via KV caches.
How it works
- 450 low-cost agents per round, split into swarms exploring different directions; each runs 40 steps of latent reasoning with no textual chain-of-thought
- A novelty filter picks the most distinct ideas by cosine distance between hidden-state vectors
- A summarizer agent merges swarm members' KV caches to generate new candidate programs, then automatic evaluation drives the next round
Results
- On 21 math discovery and scientific computing problems with Qwen3 8B/14B/32B-FP8, it beats OpenEvolve, ShinkaEvolve and CodeEvolve
- 5-13x more candidate programs evaluated per hour; median time to match baselines' 1-hour best: only 3-8 minutes
- On several problems, SwarmEvolve with Qwen3 8B improves on solutions AlphaEvolve found with frontier LLMs
More from coding & agent
- AI Agents Waste ~18% of Output Tokens, Paper Finds; Pruning 41-46% Barely Hurts Success — rohanpaul_ai · 2026-11-09
- Lovable Is Opus 5.5 at API Prices: Claude Plan Does Same Work ~45x Cheaper — PawelHuryn · 2026-11-09
- Systems Engineer: Write Code Yourself in the AI Era to Build Ownership and Taste — Abhishekcur · 2026-11-09
- Prefixwise names why your prompt cache misses — 19 causes, recovers 99.9% of savings — No-Individual159 · 2026-11-09
- Dev spends 12.5 months of nights building a local, Claude Code-style meeting scribe — Jazzlike-Swing-6198 · 2026-11-09
- After a 100% AI coding mandate, codebase triples and nobody understands it — TechPreacher · 2026-11-09