Researchers led by Časlav Nedeljković replicated MIT's SwarmWorld study using a 27B parameter model on consumer hardware. The experiment involved 50 software agents that completed all 16 research milestones, outperforming solo agents that reached only six. This single run suggests that collective advantages persist even with smaller models, though the results are limited by the lack of repeated trials. The team also found that adjusting the KV cache precision affected outcomes more than the difference between model sizes.
Read the full story at science.report.
Photo: rosh8111 / Pixabay