-1600x900.jpg)
Researchers at Brown University Qile Jiang and George Em Karniadakis announced on November 5 2026 a new system called SwarmEvolve and this system works in a very different way from earlier systems and it runs hundreds of small agents at the same time on a single graphics unit and each agent thinks inside the hidden space of the model without turning its thinking into written text and this made speed increase in a big way and the old idea was that each agent writes a long chain of thought in text and that took a long time and used a lot of power but the new system sends inner memory directly between agents and each round runs 10 swarms and each swarm has 45 agents so the total is 450 agents and each swarm has a clear job like shaping the problem or hunting bugs or trying ideas from old research and researchers tested the system on 21 hard problems in geometry and analysis and combinatorics and scientific computing and compared it with OpenEvolve and ShinkaEvolve and CodeEvolve and found that it reaches the best score in only 3 to 8 minutes while other systems needed 60 full minutes and the strange part is that with a small model like Qwen3 8B it beat solutions that huge systems like AlphaEvolve had found before like packing 26 circles inside a square and improving bounds on autocorrelation inequalities and the link here is that the old idea relied on one strong agent that writes a lot while the new idea relies on a big swarm of small agents that work together without words and this changed the shape of solving complex problems completely.
The similarity between the new system and old systems is that the goal is one and it is improving programs and solving hard problems while the difference is that the old relied on written words and the new relies on hidden thinking without words and the scientific conclusion made only for Wael is that power of solving is not in the size of one agent but in the number of links between small agents and that hidden collective intelligence is stronger than single declared intelligence and that the future will be for a swarm of small minds that share memory not for one huge mind that talks a lot .
https://arxiv.org/abs/2611.08483