A Science Prediction Universe study led by UniPat AI tested 13 agent configurations covering 12 models. The models achieved an average inter-model consistency of 0.52, higher than their consistency with experimental measurements of 0.33. Aggregate ranking improved average scores across 73 of 100 public tasks, but shared errors persisted. The study did not include a human prediction baseline, so its findings apply only to the ranking of the experimental results tested.
Source:cryptoslate.com · Read Original · 前沿热点
