Search
Mitigating Overthinking in Large Models: Excessive Cognition Can Hinder Performance
2025-03-26 /
Read about 0 minute
Author：小编
Researchers from esteemed institutions such as the University of California, Berkeley, and ETH Zurich have uncovered a fascinating insight: large models are prone to overthinking when executing agent-based tasks. To objectively measure this phenomenon, they devised a comprehensive evaluation framework and analyzed an extensive dataset comprising 4,018 trajectories. The findings reveal a compelling trade-off: while models configured with high reasoning capacity solved 29.1% of the problems, this came at a steep cost of $1,400. In stark contrast, models with lower reasoning configurations achieved a 21.0% success rate but at a significantly lower cost, approximately 3.5 times cheaper.
-
C114 Communication Network -
Communication Home
7 X 24 Track global technological trends
