frontier models
The Swarm Stops Writing Down Every Thought
Hundreds of low-cost agents exchange latent KV caches and select novel hidden-state directions for evolutionary program search.
Summary
Hundreds of low-cost agents exchange latent KV caches and select novel hidden-state directions for evolutionary program search.
SwarmEvolve organizes agents into groups that reason without decoding every intermediate step to text. A novelty filter measures distances between hidden states, while a summarizer receives combined KV caches to perform crossover. Across 21 mathematical-discovery and scientific-computing problems, the authors report the highest average score and a median 3 to 8 minutes to reach scores other frameworks needed 60 minutes to match. The comparison is limited to the tested problems, models and budgets.
Why it matters
Hundreds of low-cost agents exchange latent KV caches and select novel hidden-state directions for evolutionary program search.
Limits and context
No additional limitation was separately recorded.
Key claims
Hundreds of low-cost agents exchange latent KV caches and select novel hidden-state directions for evolutionary program search.
Evidence: source-2026-11-06-003
Sources
- arXiv preprint 2611.08483, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.