Menu

Using latent-agent swarms for math discovery and optimization [R]

Reading 0:00
15s threshold

Last year Google DeepMind's AlphaEvolve found a faster way to multiply 4x4 complex matrices, beating a record that had stood since 1969. The idea is to let an LLM write a program, score it, and iteratively improve it. So it's called " evolve ".

But AlphaEvolve runs on Gemini models. Similar open-source implementations are also usually slow and expensive. We tried to make it faster, cheaper, and better.

SwarmEvolve is our new paper, a scalable framework for math discovery and optimization.

We run swarms of hundreds of latent-reasoning agents on Qwen3 8B, 14B and 32B. In ~5 minutes, they reach scores that took OpenEvolve, ShinkaEvolve and CodeEvolve >1 hour.

Here's how:

Inspired by recent works in LatentMAS, our agents never write out their thoughts. They reason in latent space and pass KV caches to each other, so ideas from different agents mix right inside the model.…

Continue reading — create a free account

Join HashtagPLUS to read full articles, follow hashtags, vote, and join the conversation.