A computational parallel to collective intelligence in human groups

Junsol Kim, Shiyang Lai, Nino Scherrer, Blaise Agüera y Arcas, and James Evans authored a paper discussing the computational parallel to collective intelligence in human groups. Their research employs quantitative analysis and mechanistic interpretability methods on reasoning traces, revealing that reasoning models such as DeepSeek-R1 and QwQ-32B demonstrate significantly greater perspective diversity compared to instruction-tuned models, highlighting broader conflicts related to heterogeneous personality and expertise features.